But thanks to the relentless AI hype machine, many others have not only failed to improve, but have actually gone backward. They’ve lost focus, wasted time, and shipped slop that undermines the trust of the people they work with.

The difference between these two groups is having a thoughtful answer to three questions:

1. What problems should I solve with AI?

2. How should I interact with AI?

3. Which tools should I use?

So, let’s try to answer them.

Subscribe now

1. What problems should I solve with AI?

The world is currently full of “productivity theater”.

People design intricate agentic workflows that solve problems that didn’t need solving. People spend the weekend vibe-coding an app that they could have paid $2.99 for.

This is fine. Great, even! Many people have fun building this kind of thing, and by doing it, they’re helping us learn about all the new stuff.

But it is not actually making them more effective. If effectiveness is your goal, the first test of which problems are a good fit for AI is simply “is it big enough?” Does it actually take a lot of your time? Is it something you must do repeatedly? A surprising number of people are failing this test today.

There is a second, more subtle test that even more people are failing. That is understanding whether the work is about execution or about exploration.

One way to understand this point is to watch the debate currently happening on X about whether AI will kill spreadsheets. Many people in tech are convinced it will, but many in finance think they’ll never die. What is happening here?

Spreadsheets are used for two very different things:

1. “Mini software”: dashboards, inventory trackers, marketing attribution, and thousands of other hacked-together apps

2. A decision-making tool: financial models, ROI models, scenario models

The first case is about execution. All you care about is accurately charting a metric or keeping track of inventory, and spreadsheets are a quick way to do that. AI will indeed replace a lot of this because the spreadsheet is just janky software, and AI will make it easy to build non-janky software.

But in the second use case, the goal is actually exploration. When you set out to build a financial model, the whole point is that you don’t actually know yet whether you should invest in the company yet. You have to go through the process of collecting the inputs, building the model, tweaking it, adding to it, playing with sensitivities. AI can’t automate this because you don’t even know what the right answer is until you get into it.

A lot of work is exploratory. Take writing, for example: sometimes you just need to fire off an email or fill out a form. Hopefully, AI automates all of that. But if you’re writing a strategy doc or an essay or your senior thesis, the point is that you have to do the writing yourself to figure out what you’re trying to say. Writing is how you learn.

So there is a basic decision tree for how you should apply AI to different kinds of problems:

If the problem is not big enough to be worth your time, you can use it to learn about new tools, but that’s about it. For most people, building an agent to give them a “morning briefing” that summarizes their calendar and the news and the weather is not worth doing, unless they want the practice.

If the problem is big enough, and if the goal is simply execution, that’s a good candidate for automation. This would include an engineer writing the actual code, an analyst writing SQL, or a marketer generating tons of ad copy to test. Pretty soon, people will be spending very little time on these kinds of problems.

But it turns out that a lot of the most important work — the stuff that actually separates great people from mediocre ones — is in a third category. These are big, exploratory problems. This could be writing a strategy doc, designing an org chart, or deciding which features to launch. It is not possible to automate this work. Instead, we can best think of AI as a thought partner.

Which brings us to our second question:

2. How should I interact with AI?

Your mind will go through three phases as you navigate any sufficiently large, exploratory project:

1. “Wow, this is simple. It will be easy!”

2. “This might be harder than I thought. I’m getting conflicting information and running into real tradeoffs.”

3. “I’m starting to make sense of things. It seems simple again, but this time I actually understand it.”

Each project is essentially a microcosm of the Dunning-Kruger effect, in which those with low competence have high false confidence, this evaporates as they start to learn more, and then they slowly rebuild true confidence.

The temptation is to use AI to skip over the “valley of despair” entirely. You could come up with a reasonable scope for the problem, ask Claude some reasonable questions, and then just accept its reasonable-sounding answers.

The problem is that the initial answer is likely wrong in meaningful ways. And you didn’t learn anything through the process that would let you reason about where it is wrong.

What you actually need to do is use AI to navigate through the complexity with you. It should look something like this:

For example, I used Opus 4.6 to research this post about which SaaS companies will be defensible against AI. Here is what that entailed:

• I started by writing a draft of the essay cold, with no AI assistance

• I gave Claude this draft, along with a mountain of other context, including the ~20 articles linked at the end of the piece. I asked it to synthesize what everyone else was saying and use it to critique the framework in my essay

• I asked it to create a spreadsheet with the ~70 SaaS companies with >$1B in annual revenue, and iterated around the fields to fill out for each (category, business model, pricing, financial performance, market cap)

• I asked Claude to score all of the companies against the framework to see if the more defensible ones had fared better in terms of financial performance and market cap in recent months (they had, but not as strongly as I hoped)

• I then used that as a jumping-off point to iterate on the defensibility framework and re-score the companies until it was predicting their performance quite well

• I started having Claude pull out specific company examples to cite in the post, asking it to pull in public filings, analyst reports, and other context to refine them

• And so on, resulting in a thread with 300-400 back and forth questions which had to compact context 5-6 times.

When we were done, my interaction with Claude looked like the bottom of this graphic from Liz Fosslien, just with many more “human” lines than could possibly fit on it:

My original essay and Claude’s initial critique of it were almost entirely unrecognizable from what I used in the final essay. I needed to wade through the complexity with it. Using it probably sped me up a bit, but the more important effect was that it deepened my understanding of the problem. That’s how to make AI an actual thought partner.

3. Which tools should I use?

Every day, you can go on X or LinkedIn and find five new tools that you “must” use if you don’t want to become part of the permanent underclass.

And yet despite all this hype, the vast majority of people are getting the vast majority of value out of just a few things today:

1. Core chatbots: Claude, ChatGPT, Gemini

2. AI on top of general productivity tools: Either directly embedded (e.g. Notion AI) or, increasingly, via connectors from the core chatbots

3. Domain-specific tools: e.g., Harvey for legal or Clay for sales

4. And for anyone more technical, Agentic coding tools (Cursor, Claude Code, Codex).

Because of the speed of progress, much of what you see beyond the most common tools are immature, unvetted, and highly likely to waste your time. As new products mature, the core use cases and workflows get worked out, those workflows get easier to use and less brittle, and a small number of products emerge as the category leaders.

If you’re trying products before any of that happens, you’re basically doing free user testing in exchange for potential alpha for being early. For most people, that is not a good trade.

One way to reason about when you should actually adopt a new tool is the diffusion of innovations curve, originally popularized by Everett Rogers. Where you belong on this curve depends on what you’re trying to accomplish.

Innovators are typically those actively trying to apply a new technology. It may include founders and certain members of technical teams. Importantly, most of what they are doing is just trying to learn about the new technology, not trying to use it to make themselves more effective.

These tools are often hard to use, unformed, and present security or other risks. Despite this, they frequently go viral, causing many more people to try them than otherwise would. This was OpenClaw in November of last year. Today, OpenClaw (and alternative locally run agents being built by Anthropic and others) have moved into Early Adopter territory.

Early Adopters include teams building adjacent products, those leading AI adoption efforts inside companies, and creators and investors exploring an industry. A big part of the goal here is still learning, but the tools are starting to get more useful.

They often now look like “real products” but are still rough around the edges. Claude Cowork is in this territory now. Agentic coding tools like Claude Code or Codex were here in the second half of 2025, but for technical users are now in the Early Majority.

The Early Majority includes most people working in tech (certainly everyone in the product organization) and anyone else who is actively trying to keep their edge. This is where the goal crosses firmly from learning to being about improving performance.

Tools here are gaining real adoption with consumers or businesses, are easier to use, have established pricing models, and it’s often clear who the handful of winners in a category will be. They include things like AI meeting assistants (Granola, Otter), and research tools like NotebookLM or deep research modes in Gemini or Claude. The paid offerings of core chatbots were in the Early Majority a year or two ago, but are now in the Late Majority.

The same person can take on multiple of these personas, depending on their role and how they are trying to use the technology. Casey Winters has a good piece on how to think about this, which is worth reading here.

The most important thing is simply being intentional about where you sit and being willing to ignore the noise if it’s too early.

Into the future, serenely

Anxiety levels in tech are through the roof right now. It feels like things are moving at 1,000 miles per hour, and if you don’t keep up, you might do irreparable damage to your career.

One of the reasons it feels this way is that there are many people incentivized to push the fear narrative: people building or selling new products, investors in these products, and a legion of influencers trying to scare you into clicking.

It is true that AI will drive unprecedented gains for every knowledge worker in the world. But it is also being shoved on us with unprecedented speed and hype. New tech always rolls out more slowly and less completely than it seems like it will at first.

For most people, if they’re using a small set of the most common tools, engaging with them thoughtfully, and applying them to their most important problems, they’re going to be just fine.

Credits

Thank you to Casey and Gabriela for their feedback on this essay.

The thinking is that because it is now much easier to create software, new companies will be built to serve every niche use case and low-TAM market. Clones of major tech companies will emerge to carve out market share.

This is similar to the intuition people had about the internet, and they were wrong then too.

AI will drive greater concentration of value into fewer companies and create the biggest businesses of all time.

This essay is a brief exploration into why and how this will happen.

Subscribe now

The historical trend

The economy is on a long arc toward concentration. We started with small shops in every town and ended up with Amazon. We started with millions of local musicians and ended up with Taylor Swift.

For at least the last 100 years, this has been a trend toward concentration of value (revenue, profit, capital) into the top companies, but not a concentration of labor.

A recent study demonstrated that the share of value accruing to the top 1% of companies has steadily increased since the early 1900s, and is now at about 80% across most major market economies.

Over the same period, however, the share of employment in those top firms has been essentially flat.

This divergence between value and labor helps to explain why many people got the internet wrong. It did, in fact, enable the creation of many more small businesses in industries like consumer products, retail, media, and services. This was decentralizing in many visible ways: more entrepreneurs, more startups, and greater diversity in the products and media we consume.

But over the same period, concentration of value just continued to march upward. One reason for this is that while the internet enabled a wave of small businesses, it also enabled a new kind of mega-company, which Ben Thompson called the aggregator. Companies like Google, Meta, and Apple control the distribution that all of these new businesses need in order to grow, and as a result, command much of the newly generated economic value.

Why AI will accelerate this trend

There are two mechanisms by which new technology has consistently driven greater concentration, and AI will reinforce both.

The first driver of concentration is that technology increases the power of the leading companies relative to their competitors.

Power (sometimes called defensibility) is the ability for a company to durably earn more than its competitors. You can represent it with an s-curve, with an increasing input (scaling operations, growing customer network) on the x-axis and an increasing benefit (lower costs, more valuable product) on the y-axis.

Technology tends to push this curve outward: the top companies now have farther to move along the curve before the benefit asymptotes, making it harder for anyone to catch up. Technologies like manufacturing and the semiconductor pushed the curve out for scale economies. The telephone and the internet pushed it out for network effects.

Certainly for chip manufacturers, and probably for the frontier model companies, the scale economies curve has again shifted out. At this point, it’s difficult to imagine anyone but an incumbent with existing scale entering the competition.

For the application layer of AI, the change will come in the form of strengthening data effects, where the larger and more proprietary the dataset, the more value a product can deliver.

Prior to AI, data effects came in a milder form: Amazon’s product ranking improving with more conversion data, Stripe’s fraud model getting better with more transactions. These are powerful but have capped upside on the y-axis: purchase conversion or fraud prevention can only get so good.

The data effects of AI applications will be different for two reasons. First, they don’t just optimize the product; they determine what it is capable of doing in the first place. A moderate amount of tax return data might allow a company to build a pretty good CPA assistant. But a lot of tax return data may mean they can just close the books and file returns for any small business in the US. Second, the data produces a generalizable capability: that same tax return data may mean the company can also underwrite any SMB in the country for a new credit line.

The second driver of concentration is that technology lowers transaction costs within companies.

In The Nature of the Firm, Ronald Coase answered the question: why do we need companies at all? If the free market is so good at allocating resources via prices, why don’t individuals just contract with each other for every task?

His answer: coordinating each transaction has a cost (setting a price, negotiating, enforcing contracts), and companies can lower these costs by coordinating them internally. In most cases, for example, it is more efficient to hire a marketer than to contract out each marketing task to a 3rd party.

Why, then, isn’t the whole economy just one big company? Because as you organize more activity within a company, it starts to get too complicated, and resources start getting allocated inefficiently. From Coase:

“Other things being equal, therefore, a firm will tend to be larger:

a. the less the costs of organizing and the slower these costs rise with an increase in the transactions organized.

b. the less likely the entrepreneur is to make mistakes and the smaller the increase in mistakes with an increase in the transactions organized....

This isn’t just theory - if you’ve spent any time around big companies, you’ve felt it viscerally. As companies grow, they start to trip over themselves. There are too many people and too much process. This is when companies start failing to make the next leap into a new product extension or market expansion. This is when they start hiring vendors and consultants instead of doing things themselves, because it’s easier to just over-pay instead of wading through the internal molasses.

Technology often helps with this problem. Software like Excel, Salesforce, Notion, or Slack makes it easier for companies to manage themselves, and in effect raises the size of the team or level of complexity at which the company becomes unmanageable.

AI will do this much more powerfully than legacy software, because it won’t just give humans tools to manage more efficiently, it will do a lot of the work itself. AI means many fewer people will be required to produce a unit of output. This will lower headcount, lower complexity, and cause internal coordination costs to plummet.

In the process, it will make companies much more valuable per employee, which will allow them to retain their best people by both paying them more and creating an environment in which they can actually get things done. It will take longer before a company’s best people start jumping ship for somewhere with less bureaucracy and more upside.1

The combination of these two drivers means that both companies have more to gain from getting larger, and they will be able to get much larger before they become slow, dumb companies and start making mistakes.

These dynamics will concentrate more value into fewer super-companies, a dynamic that will play out from two directions: (1) consolidation of existing software companies and (2) turning non-software industries into software and consolidating them in the process.

Software consolidation

The average large company in the US runs about 350 different SaaS apps.2 That number is about to come down a lot.

It would be advantageous for a software company to use a shared infrastructure and data model to address a wide range of use cases. But so far, they have been held back from becoming massively multi-vertical by three things:

1. The sheer volume of software to build and maintain

2. The difficulty in understanding the needs of many different kinds of customers and use cases

3. The inability to expand without creating a product that is much too complicated for the typical user to navigate

AI weakens all of these constraints. The first is the most obvious: coding agents dramatically reduce the cost of building software. And they won’t just ship the initial product; they will also help to respond to and adapt it to bugs, edge cases, and feature requests.

Second, it will now be easier for one company to maintain more high-quality product teams. Previously, creating a new product pod might have meant hiring 10-12 people. Soon it might mean hiring 3-4. And as we discussed above, it will be more feasible for companies to retain their very best PMs, designers, analysts, and engineers who are doing the hard work of customer needs finding.

Finally, there will be agents running on top of all of the software that is built. Instead of users drowning in a growing sea of menus, settings, and workflows, they will have a simple entry point. Agents will abstract the complexity away when they don’t want it, and direct them to the correct workflow when they do.

As a result, we’ll start seeing companies jump across functional and industry lines that used to look static. Do you really need Workday as the system of record, Greenhouse for recruiting, Lattice for performance management, Culture Amp for employee engagement, Pave for compensation, and Checkr for background checks? Why isn’t that all just one thing?

It’s hard to tell who will do this consolidation — probably some mix of software companies expanding horizontally into new use cases, frontier models expanding downward into apps, and PE rolling up struggling software businesses. But no matter who does it, consolidation among existing companies is coming.

You might be thinking that this makes sense for existing players, but what about the long tail of new software companies that just couldn’t be built before AI and now will? Won’t those drive fragmentation?

I’m not convinced that small-TAM software will be built by individual companies. It might be that multi-vertical software businesses expand into ever-narrower use cases. But it’s certainly possible that there is just such a long tail of use cases that big companies can’t serve them well, and a wave of new entrepreneurs and companies do emerge to serve them.

If this happened, however, it would follow a pattern similar to what we saw in media, consumer products, and online retail during the rise of the internet: an explosion of many new companies driving the number of new firms up, but the bulk of value accruing to the aggregators who own distribution.

That is because AI will make it easier to build, but it will not make it easier to get distribution, and that will become the bottleneck, allowing aggregators to extract most of the value.

In this world, we may come to think of small-TAM software producers more as “creators” than as companies. A single person can do it, and many of them will. It could be fun and meaningful work, but it won’t be a great business model. It will be hyper-competitive and have a high failure rate, with a strong power-law distribution, meaning only a small number break through and generate real profits.

The need for distribution will ensure that even if there is fragmentation at the producer level, there will be ever steeper concentration at the platform level.

Turning non-software into software

AI will also drive the concentration of currently non-software industries. Perhaps the most dramatic example here will be in the services, which are highly fragmented today. In the US, there are 1.2M consulting firms, 400K law firms, and 85K accounting firms.

They are fragmented because (1) there are limited returns from scale and (2) the transaction costs to hire them on the open market are low: their services are discrete, well understood, and easy to price.

But if you’re not just renting out time but instead building agents that can do the work autonomously, that “returns from scale” dial gets turned way up. It’s a race to aggregate as much data and as many completed jobs as possible to enable a product that can do more work at higher accuracy and lower cost.

In our accounting example above, an automated services firm could reduce the cost of serving the marginal customer to close to zero. And as the data model grows, it would get faster and less error-prone across any type of customer and edge case.

This is the kind of company that Julien Bek wrote about in his excellent piece Services: The New Software:

“A company might spend $10K a year for QuickBooks and $120K on an accountant to close the books. The next legendary company will just close the books.”

Admittedly, tax prep is a bullseye use case because it has an agreed-upon set of rules, and the outputs are verifiably accurate - that is why many people are using it. But when you start to dig into the services, you find many other examples that look similar: claims adjusting, underwriting, paralegal, payroll processing, and regulatory and compliance, to name a few.

Automated services companies will get massive, and in doing so, begin to consolidate the fragmented services landscape we see today.

We’ll see AI’s centralizing effects across a much wider range of industries than we’ve discussed here. We haven’t even covered anything that touches the physical world, for example: industrials, transportation, manufacturing. AI can also be a centralizing force in these industries, particularly when paired with robotics.

The lesson of history is that new technology leads to more concentration, and AI has the potential to be the most powerful example of this yet.

This will not be an incremental change. As every major new technology before it has done, AI will produce companies that are so large they make the last generation look quaint.

End note

This essay is meant to be descriptive, not normative. I’m not saying it’s a good thing that we’ll see greater concentration, just that it is likely to happen. There are many benefits from centralization - the last wave of super companies produced unparalleled consumer surplus. But it also risks disrupting small business owners, contributing to economic inequality, and driving political unrest and a policy response that could be destructive.

It is worth noting, however, that the study we cited earlier also found that increasing business concentration has not necessarily led to greater inequality, in part because the ownership of companies has also become more diffuse over the same period. This is one reason to hope that companies like OpenAI and Anthropic get public quickly.

From the study:

Interestingly, across the countries and time periods in our sample, rising concentration of production activities is not necessarily accompanied by rising inequality among households, which suggests that their determinants can be different.

Although production has become more concentrated, ownership appears to become more diffuse over time: large companies used to be owned primarily by individuals and families; now, a broader group of households own their equity through investment funds and retirement savings.

Is it oversold? Unclear. Multiples are the lowest they’ve been in over 10 years, but so are growth rates.

The more important point is that the selling was indiscriminate - everything with perceived exposure got crushed.

We’re now entering the shakeout, where markets try to sort out the true winners and losers. Below is a framework for determining which is which.

Subscribe now

What’s the problem?

The threat is not that companies are going to vibe code their own software. Software companies are not their code but the business wrapped around it, the hard parts of which are understanding customer needs, taking the product to market, and bearing the heavy weight of maintenance and customer support.

What will actually happen (as has happened in each past wave) is that existing forms of software will remain, and new and even more valuable forms will get built on top. This is the point Steven Sinofsky made in his excellent essay:

“The lesson here is that whatever the world thought would end just ended up being vastly larger than anyone thought. And the thing that people thought would forever be replaced was not simply legacy but ended up being a key enabler.”

Software as a category will be fine. But it’s unclear if these specific companies will be fine, because there is an explosion of competition coming from every direction.

New startups can now build viable alternatives at much lower cost. The major tech platforms and frontier model providers will try to verticalize into every attractive space. Other incumbent software companies will take advantage of cheap code to move horizontally into new use cases. And customers themselves might now be more inclined to spend their budget on tokens instead of more software.

Whether a company survives will come down to two questions:

1. How hard is it for customers to switch to a competitor? This determines how long a company has before it must compete with all of these new competitors on merit.

2. Does the value of the product compound with scale? This is the merit itself. Is the product actually more valuable or lower cost than what the new competition can offer?

Taken together, these dimensions produce three distinct segments. Companies with compounding advantages will prove much more durable. Companies without them will need to adapt their products and monetization models, and switching costs will determine how much time they have to do so.

Let’s break down the drivers of switching costs and compounding value in more detail.

Switching costs

We’ve heard a lot about “systems of record” recently. But this is really just a proxy for switching costs, and increasingly not a very good one as AI brings down the cost of migration. Switching costs are indeed coming down, but there are three things that still matter a lot:

Workflow tax

If a lot of your company has spent a lot of time learning to use a particular UI, it is costly to switch. Some are claiming that this moat is gone because a text box is going to eat that UI. That might be true for some use cases, but not most. For workflows that are sufficiently deep, frequent, or repetitive (which is a lot of them), chat becomes very frustrating very quickly.

Workday and Atlassian have to worry about better products getting built, but they have some time because of how painful it is to get a whole team to switch to something new.

Business disruption risk

If a product manages payments, infrastructure, security, or any other business-critical function, it is much riskier to rip out.

Retailers will think twice before replacing Square, and restaurants will think twice before replacing Toast, because getting it wrong is very costly. This is why historically, new POS companies have carved out market share more by becoming the way new businesses start and less by getting existing businesses to switch. The “theoretical hull speed” of growth is how fast businesses are turning over in an industry, and AI is not changing that very much.

Legal and regulatory risk

You want to rip out ADP for a brand new competitor? Your funeral. Here is what Aaron Levie had to say about it at YC:

“Three years from now there’s going to be a bug… and that bug is going to pay people the wrong amount of money. I don’t want to have to go and call my IT team in the middle of the night to be like, ‘Sh*t, you have to go fix this bug that paid everybody the wrong amount of money.’ I want to be able to go to a company that I know that I can sue if they f*ck up … Because I can’t sue my internal IT team and I certainly can’t sue Anthropic.”

It’s the same dynamic for companies operating in regulated industries. For better or for worse, AI is not going to crack Epic’s hold on the EHR market any time soon.

Compounding value

Some software products get more valuable the more they are used or the greater scale they achieve. This is more important than switching costs, because it determines whether all of this new competition has a chance at competing in the first place.

These forms of defensibility are almost fully intact and, in some cases, strengthened by AI. There are three main ones:

Data effects

In some cases, data is valuable because it is proprietary, typically generated over many years of operating. CoStar has 40 years of commercial real estate data, and unlike the MLS for consumer real estate, there is no public equivalent. ADP has payroll data for 20% of the US workforce. Epic’s Cosmos network has longitudinal medical records for ~80% of Americans.

The other way data effects manifest is when the scale of data directly improves the product experience. Many security businesses are highly defensible for this reason. CrowdStrike processes 100 billion security events a day across all of its customers, making it faster and more effective at identifying threats for any one of them.

AI does very little to erode the advantages of proprietary or high-scale data and will likely help these companies apply it even more effectively.

Scale economies

This is not about the ability to amortize high fixed costs over a growing base of low marginal cost customers - virtually all software companies have that. And you could argue that AI will both lower fixed costs and increase marginal costs, upending the way the economics of software businesses have worked historically.

But some software businesses actually improve their unit economics with scale. This is Stripe having lower costs because its fraud model improves with scale, or AppLovin getting better at pricing impressions as its auction liquidity improves. The cold start problem will be difficult for new entrants in these cases.

Network effects

The most common software network effect has been ecosystem: networks of 3rd party apps and integrations built around companies like ServiceNow and Salesforce. This is weakening in a world where it’s easier for anyone to build apps or integrations around any new product that takes off. So there is some value here, but it’s worth discounting.

But there are other forms that are not impacted, including true cross-side networks (e.g. Shopify Shop Pay in which the product gets more valuable to customers as more merchants use it and vice versa), and cross-company workflows within an industry (e.g. Veeva being used by pharma companies, CROs, and their partners to collectively manage drug development).

How companies stack up

Based on how well they score on switching costs and compounding value, let’s make some (rough) predictions about where different companies belong.

Segment 1: Durable

It matters much less to these companies if there is an explosion of competition, because new competitors start without the network, data, or scale advantages that they have. In many cases, AI will actually be a tailwind as it will allow them to pull forward their roadmaps and further compound their advantage.

The primary thing these companies need to watch out for is adjacent companies that already have access to the data or customer base to bootstrap the compounding advantage, and may now be able to do so more quickly with AI.

Example companies in this segment include CrowdStrike, CoStar, Shopify, Zscaler, Stripe, Datadog, Veeva, Epic, and ADP.

Segment 2: Eroding slowly

These companies have limited compounding advantages. There will be an explosion of alternative products that are comparable, or in many cases better, because it will be easier to build products tailored to narrower segments and use cases.

But they have some time, because it is hard for customers to switch due to workflow complexity, business disruption risk, or legal/regulatory risk. They should be using this time to sprint to improve their offerings and add AI agent functionality on top of the core product.

Example companies in this segment include Atlassian, Salesforce, SAP, and Workday.

Segment 3: Eroding quickly

These companies have neither compounding advantages nor high switching costs, which will lead to a fast bleed of market share as new alternatives show up. They’ll likely need to shift their offerings and monetization model entirely if they want to survive the transition, which is why Gokul Rajaram recently predicted that some companies may need to go private in the coming years if they aren’t already.

Example companies in this segment include Monday, DocuSign, UiPath, and Zendesk.

Where do we go from here?

The market seems to be internalizing these differences in durability to an extent: some of the largest YTD losers have been in more exposed buckets (Atlassian, Monday), and some of the more durable companies have fared better (CrowdStrike, Datadog).

But I expect there is more separation coming. We’ll look at the past 10-15 years as a momentary blip in which the artificial barrier of code scarcity made it possible for too many of these companies to get extraordinary returns. Take away that barrier, and we’ll see who sinks and who swims.

As he often does, Ben Thompson put this perfectly:

The real risk I see for software companies is the fact that while they can write infinite software thanks to AI, so can every other software company. I suspect this will completely upend the relatively neat and infinitely siloed SaaS ecosystem that has been Silicon Valley’s bread-and-butter for the last decade: identify a business function, leverage open source to write a SaaS app that addresses that function, hire a sales team, do some cohort analysis, IPO, and tell yourself that you were changing the world.

Historically, most of the value accrued to the companies that did the very hardest things: scaled networks, improved the structural economics of an industry with data, took on outsized risk, and navigated legal and regulatory complexity. That’s the world we’re headed back to.

Appendix: What is not in the framework?

A number of other dimensions have been proposed for distinguishing between software winners and losers. I think it is worth addressing two of them:

Per-seat vs. outcome pricing: The argument is that AI will put pressure on per-seat models by both reducing the number of seats and creating alternatives that are outcome-based. This is valid, but I would argue that it is mostly downstream of the dimensions outlined above. Most of the examples of durable companies price based on outcomes, and all of the eroding companies price based on seat. Pricing by seat seems to indicate a business model in which value does not compound with scale and in which ROI is harder to measure, so seat-based monetization is the only option.

Hardware: Companies with hardware will be more defensible both because AI helps less with the creation of new devices and because they contribute to customer stickiness. This is valid, and I only excluded it from the framework because (1) it’s not all that common (roughly 10% of the 68 companies in the EMCLOUD index) and (2) most of the companies with hardware are already predicted by the framework to be durable.

Credits

Thank you to Zach Grannis for his feedback on this essay.

Sources

Essays

• 10 Years Building Vertical Software: My Perspective on the Selloff - Nicolas Bustamante

  • Rebuttal to Nicolas - @atelicinvest

• AI Agents to Boost Productivity and Size of Software Market - Goldman Sachs

• Build vs. Buy - Jamin Ball

• Death of Software. Nah. - Steven Sinofsky

• In Defense of SaaS - Finbarr Taylor

• Is SaaS Dead? No. But One Thing Is Clear: It’s Unstable. - Jason Lemkin

• Microsoft and Software Survival - Ben Thompson

• SaaS Isn’t Dead. It’s Worse Than That. - Michael Bloch

• Shopify Earnings, Shopify’s AI Advantages - Ben Thompson

• Time’s Up for SaaS (Grow Faster or Vanish) - Alex Clayton

• The Agent Will Eat Your System of Record - Zain Hoda

• The Crumbling Workflow Moat: Aggregation Theory’s Final Chapter - Nicolas Bustamante

• Why that $2 trillion software stock wipeout didn’t derail the AI bull market - Jim Edwards

Podcasts/video

• Ben Thompson from Stratechery on AI Ads, the end of SaaS, and the future of media

• Gokul Rajaram: Lessons from Investing in 700 Companies

• Aaron Levie: Why Startups Win In The AI Era

• Why AI won’t destroy existing software

• All Things AI (BG2)

Share

“The intangibles seem to be becoming more important, not less important over time. It seems to be becoming more like an art and less like a science as we go.”

This was Marc Andreessen describing how venture capital is changing1, but he could have been describing any industry.

Science is everything you can measure, test, and replicate. In one word, science is verifiable. That is also the word that Andrej Karpathy used to describe where AI will win.2 LLMs will suck in all of the writing, all of the images, video, code, and data. And then they will practice until they simply don’t make mistakes.

What is left for humans is art. Intuition, feel, anything you can’t put into words or systematize. The “special sauce” that anyone who is at the top of their field has.

There are basically three forms of art in the business world. These skills will not only be defensible but become wildly more valuable as AI progresses. This essay breaks down what they are and how to develop them.

Subscribe now

Taste

Taste is the ability to evaluate quality. That might be knowing which design customers will love or which words will resonate most strongly. It is sometimes called discernment, craft, or product sense.

When you study people with the very best taste, you hear one theme repeatedly: taste is not about adding things; it is about removing them. It’s an act of curation or editing.

Dieter Rams might be the greatest industrial designer of the last century. Jony Ive copied was influence by him.3

Rams lived by the axiom “weniger, aber besser” (less, but better). LLMs generate creative work at unprecedented speed, but they don’t know which of it is good. “Slop” translates roughly to “more but worse.”

As the volume of creative work explodes, we need people with great taste more than ever. The rise of vibe coding has led many companies to expect their designers to ship code. That’s great if it helps them prototype and edit faster, essentially acting as an accelerant for applying taste. But we should not turn designers into mini engineers. Because increasingly, code is cheap and taste is expensive.

In marketing, the ability to know what will resonate with an audience will resist AI. Generating many variations of creative, analyzing funnels, and running A/B tests will not. The best brand and product marketers will be standing long after the last performance marketers are gone.

Judgment

Judgment is the ability to make good decisions under uncertainty. It is sometimes called strategic thinking, systems thinking, or prioritization.

The one thing that’s going to be truly future proof is judgment … In an era when you can do everything, the question is which of these things matter. - Gokul Rajaram

Information is no longer the bottleneck: it’s easier than ever to know everything about how a business is performing, synthesize all of the feedback from customers, and see everything your competitors are doing. Judgment is the ability to parse all of that and decide what to do.

Investing is close to pure judgment because it presents the same information to everyone and requires them to make the right decision. Charlie Munger called it the act of “inversion” - conceptualizing the outcome you want and working back to today. That requires weighing all of the messy game theory, incentives, human irrationalities, and real-world constraints that will lead to the end result.

New versions of LLMs are getting better at solving harder rules-based problems and making fewer errors. But they’re making very little progress in coming up with “the answer” to strategic questions. You still have to provide LLMs with your hypotheses, and then you have to cut through the noise they generate to decide on a direction. You need to inject your own judgment.

There are many jobs today for which judgment is just one part of the scope. They are part art, part science:

Analysts are expected to make good business assessments, but also query, parse, and visualize data.

PMs are expected to prioritize the right product work, but also to document and coordinate.

Software engineers are expected to architect systems that are robust and scalable, but also to write the code itself.

Over time, these jobs will become pure judgment. Those with great judgment will do the work that ten people used to do. Those without it will need to find new jobs.

Influence

Influence is the ability to shape human activity.

Judgment tells you what kind of business to build, taste tells you what kind of product to build. Influence gets people to actually help you build it.

Influence is sometimes called leadership, persuasion, or emotional intelligence. To be great at it, you must understand human motivation and desire. And you must be able to communicate with clarity and resonance, creating shared meaning.

But most importantly, you must be able to build trust with other humans.

Tobi Lütke has a metaphor called the “trust battery”. A relationship starts at some baseline charge - say 30%. Each interaction either charges or drains the battery. Trust grows as you repeatedly make choices that demonstrate reliability and good intent.

AI can’t build trust because it can’t make choices. Being trustworthy is predicated on having the option to do the wrong thing but deciding to do the right one.

Sheryl Sandberg is archetypal of someone who excels at influence. At Facebook, she took the vision and did the messy work of getting a big, complicated org to row in the same direction to achieve it. It’s not surprising that she talks a lot about trust:

Your strength will not come from your place on some org chart, your strength will come from building trust and earning respect. You’re going to need talent, skill, and imagination and vision, but more than anything else, you’re going to need the ability to communicate authentically, to speak so that you inspire the people around you and to listen so that you continue to learn each and every day on the job.

Any job with influence at its core will be highly resistant to AI. Some people have observed AI’s success in automating customer support and predicted that it will do the same for sales, but they’re wrong. Customer support is about solving a specific problem, and humans are OK letting a machine do that for them. But sales is about influence, trust, and credibility. For that, you need a human. Great salespeople will be standing longer than almost any other profession.

How do you develop these skills?

It is by definition difficult to say anything precise about how to learn taste, judgment, or influence - they are art, after all. But it does seem that as they become more important, certain approaches to learning are becoming more important as well.

Information diet

If you ask an artist or a designer how they developed their “eye,” often you’ll hear something about the other authors or designers they love. They’ve made a conscious attempt to expose themselves to the best in their field in order to understand what great looks like.

This is not just about taste. Learning judgment requires studying the best businesses, being part of highly functional ones, and watching great leaders make decisions. Learning influence requires cultivating relationships with many kinds of people and exposure to those who are charismatic and genuine.

Slop is career poison. The algorithm feeds you things that play to your base emotions, which is close to the inverse of the best quality. Do that long enough and your ability to recognize greatness will atrophy. More than ever, we need to intentionally curate what we’re exposed to.

Feedback loops

The unsatisfying answer to “how do you get good at something like making strategic decisions?” is often that you need a lot of reps. There is no rulebook, so you just have to practice.

And you don’t just need attempts, you also need to see if they work. You need a closed feedback loop.

This is one of the dangerous things about fields like consulting. They are excellent at giving you exposure to many topics, but bad at allowing you to see the consequences of your actions.

For people early in their careers, the best environment is often something like a mid to late-stage startup. They are operating at sufficient scale and rigor that the work has real consequences, but they are still nimble enough that you can take many swings.

Physical presence

It’s very dangerous for junior people that AI and remote work are rising at the same time.

The kinds of things you can uniquely get in person are often exactly those that are hard for AI to replicate, because it’s the only kind of data AI can’t get its hands on. The subtle cues you get about how someone is feeling when you speak in person. Watching someone use the product over their shoulder instead of over Zoom. The tone of the “meeting after the meeting” where product decisions actually get made.

Being in person, surrounded by talented people, is an accelerant for developing taste, judgment, and influence. Anyone (especially anyone early in their careers) with the ability to work in the office at least a few days a week probably should.

Mentorship

Perhaps the biggest risk is that AI automates entry-level work, but entry-level work is how you get the reps to build taste, judgment, and influence. If law firms don’t need associates, how do they train partners?

This problem will be solved, because it has to be solved if great companies and firms want to survive. But it could be a very rocky ~decade.

One way to protect yourself is to have great mentors. That could mean mentors in the traditional sense: people who have done the things you want to do and can give you advice along the way. But it could also mean making sure you have an excellent manager, who will invest in you even if the incentives to do so are lower than they were before.

Specialty

There’s a debate happening now over whether the future belongs to generalists or specialists. Both sides are right.

The most successful people will be generalists when it comes to the way they solve problems. AI tools will abstract so much of the job that roles will start blending. In AI-native startups, we’re already seeing product, engineering, and design blend into one “builder” role. Analytics, finance, and research could similarly blend into an “insights” role.

But they will be specialists when it comes to the problems themselves. Going deep into a particular industry, business model, or customer is how you can get deep enough to develop taste and judgment and to build up the credibility necessary to be influential.

Paul Graham has a recipe for doing great work. It requires getting to the “edge” of a field by going deep in it:

Four steps: choose a field, learn enough to get to the frontier, notice gaps, explore promising ones.

So if you’re trying to become an artist, this seems like useful advice:

1. Consume information intentionally

2. Get many feedback loops

3. Be there in person if you can

4. Seek out great mentors

5. Generalize on skills, specialize on problems

This advice is worth following even if AI progress plateaus. Taste, judgment, and influence have always been the highest forms of professional work. Becoming great at them has always been the surest path to impact, earning, and satisfaction.

It may just turn out that before too long, it is the only path.

Credits

Thank you to Tim (who has great taste) for his thoughts on this essay.

References

• AI Will Change What It Is to Be Human. Are We Ready?

• Gokul Rajaram: Lessons from Investing in 700 Companies

• How to Do Great Work

• Inversion and The Power of Avoiding Stupidity

• Is There Such a Thing as Good Taste?

• Sheryl Sandberg’s HBS Speech

• Taste Is Eating Silicon Valley

• Tobi Lütke: The Trust Battery

LLMs are on a collision course with marketplaces. The Verge calls it The DoorDash Problem:

So what, exactly, is the DoorDash problem? It’s what happens when an AI interface gets between a service provider, like DoorDash, and you, who might send an AI to go order a sandwich from the internet instead of using apps and websites yourself.

Marketplaces have always had to pay a lot for new customer acquisition, but they then pay it back through an annuity of repeat transactions that they get for free.

If ChatGPT (which I’ll use as a stand-in for whichever LLM wins) becomes the place customers start every search, that’s a big problem. Marketplaces would have to spend much more to acquire each transaction individually, and they would also lose out on all of the other things they do with customer engagement, like ads and upsells.

It’s true that this will happen to some marketplaces. But many will be just fine (and ironically DoorDash is one of them).

This essay unpacks how defensible each type of marketplace is, and what they can do to improve their odds.

Three factors will matter the most in determining whether customers decide to start their search on an LLM or directly on a marketplace:

1. Difficulty of supply aggregation

2. Degree of marketplace management

3. Nature of customer engagement

1. Difficulty of supply aggregation

A marketplace’s first and most important job is to aggregate supply. OpenAI hasn’t tried to do this yet, but they will soon. They’ll allow suppliers to create accounts, similar to how every local business ultimately created a business profile on Google.

But in many industries, it will be very difficult to match the supply of the leading marketplaces.

The hardest supply to aggregate is fragmented (many small suppliers), heterogeneous (each offering unique things), and illegible1 (difficult to understand service offerings, availability, quality, price).

Matching Expedia’s supply of hotels is not very hard. Hotels are large businesses with standardized offerings and teams of people whose job it is to explore new growth channels like LLMs.

Matching Airbnb is much harder. It’s taken them many years and billions of dollars to aggregate millions of individual homes and rooms. They’ve had to invest heavily to make amenities, photos and quality legible to customers.

Comprehensive supply is one key to keeping customers coming back. Show someone you don’t have what they’re looking for too many times, and they’ll stop trying.

2. Degree of management

The second thing marketplaces do is a series of jobs to help customers find and purchase from suppliers. There are four types of jobs:

1. Searching for suppliers

2. Facilitating the transaction

3. Managing risk

4. Managing service delivery itself

The number of these jobs a marketplace does determines how heavily managed it is (more on this here). The further to the right, the more defensible it will be.

Google has long been unsatisfied with just customer acquisition dollars, and has been trying to eat the entire marketplace profit pool.

Google Flights, Google Shopping, and Google Local were an effort to do the “searching for suppliers” part of the marketplace job directly in the SERP.

And to an extent, it worked. Google already takes most of the profit away from pure lead gen marketplaces, making them pretty bad businesses.

ChatGPT is just another aggregation layer like Google, but with two big differences:

1. It makes search much better by understanding intent, enabling multi-turn (back and forth) querying, and parsing and aggregating information across many sources

2. It can actually transact on your behalf, which we’re starting to see take shape with initiatives like Instant Checkout

Superior search plus the ability to transact will allow ChatGPT to push further into the marketplace stack than Google, and do everything a transactional marketplace can.

But is ChatGPT going to try to do the final two jobs of managing risk or managing service delivery itself? Are they going to start accepting returns? Offering financing terms to buyers? Are they going to manage their own drivers or build their own logistics network? That seems very unlikely.

The more value a marketplace’s customers get from those final two jobs, the more ChatGPT will rely on them to deliver the kind of experience customers expect. For DoorDash, even assuming the LLM could aggregate all of the restaurants they have (very hard), customers want to know that their food will get to them on time, still warm, not tampered with, and that they get their money back if any of those aren’t true.

Together, difficulty of supply aggregation and degree of management predict whether LLMs need a marketplace at all.

The most dangerous quadrant below is the bottom left. Marketplaces that (1) are not managed and (2) have supply that is easy to aggregate could be replaced entirely by an LLM.

The upper left is also dangerous, because the LLM can provide the same service with at least a subset of the marketplace’s supply, starting to unravel its network effect.

3. Nature of customer engagement

Managed marketplaces (the two right quadrants above) aren’t at risk of being replaced outright. But they’re still at risk of having to pay for more transactions than they do today if customers begin to start more of their searches in an LLM. How often that happens will be driven by the nature of customer engagement, on two dimensions:

Transaction frequency

The more frequently the customer uses a marketplace, the more likely they are to go directly to that marketplace instead of a general-purpose tool like Google or an LLM.

When you use a product a lot (e.g. take rides with Uber) you’re more likely to remember its name, more likely to have the app on your phone, more likely to be accustomed to their UI, more likely to join their rewards program. High frequency marketplaces just have deeper brand hooks into the customer and usually have much better retention rates as a result.

Purchase consideration

The second dimension is how heavily considered the purchase is.

Planning a vacation or buying new skis usually requires a bunch of research on which options are best and time spent comparing brands, options, and prices. This is an ideal LLM use case. Even if someone is a loyal customer of a marketplace, they’re likely to start their search on an LLM for high consideration purchases.

The marketplaces least likely to lose transactions are high frequency, low consideration use cases like ride sharing. And the most exposed are low frequency, high consideration like travel.

How different industries will play out

The result is a spectrum of outcomes, all the way from extinction on the left to business as usual on the right.

Hotel marketplaces (e.g. Expedia, TripAdvisor) will fare very poorly. The current winners rely heavily on paid and organic search traffic, and as that shifts to LLMs, the LLMs are likely to initially take much more of their economics, and ultimately replace them entirely.

Local services marketplaces (e.g. Thumbtack, Angi) already have relatively low repeat rates and are trying to address this by going deeper into home maintenance and planning tools. However this is a core LLM use case and customers are likely to start many of their searches there, shifting even more of the economics away from the marketplaces. The way out is to fully rebuild the experience around AI and become managed marketplaces.

Home rental marketplaces (e.g. Airbnb, Booking) will do better than hotels, because supply is significantly harder to aggregate and make legible. But they are still at risk of losing economics on a lot of transactions because they are often tied to high consideration travel planning use cases.

E-commerce marketplaces (e.g. Amazon, Walmart) will overall do well because they are heavily managed (creating significant speed and convenience benefits for customers) and have long tail, fragmented supply. But they will start to lose economics on more heavily considered purchases, unless they can successfully go AI-first in their experience and are willing to give up ad revenue, which we discuss more later.

Food delivery marketplaces (e.g. DoorDash, Eats) will be mostly unaffected. They may lose some transactions tied to heavier planning occasions like business and events, but these aren’t very common. The “DoorDash problem” is not really a problem for DoorDash.

Ride sharing marketplaces (e.g. Uber, Lyft) will be almost entirely unaffected by LLMs because they are heavily managed and have very frequent, low consideration purchases. This is good because they need to spend their time focusing on the other big AI disruption coming in the form of driverless cars.

What marketplaces should do now

It will take multiple years for LLMs to march through each industry, particularly now that ChatGPT and Gemini are going to have to spend a lot of their energy competing with each other. Here is what marketplaces can do in the meantime to strengthen their position:

1. Do things that LLMs won’t

This includes underwriting the transaction with returns, guarantees, and financial terms, and it includes participating in physical service delivery, like building your own fleet or logistics network. All of these reduce gross margins, add operational complexity, and take a long time to get right. But that is precisely why LLMs won’t follow.

Similarly, supply that is difficult for LLMs to aggregate is a defensive wedge. Amazon went direct to factories to create a whole new category of supply. DoorDash rolled out the red carpet with custom deals and heavy support to get the best restaurants on their network. This is the kind of thing marketplaces should be pursuing.

2. Build AI search natively

Particularly for high consideration use cases, marketplaces must stay as close to the LLMs on search experience as they can, so that other advantages can continue to tip the experience in their favor.

Marketplaces likely do not need to be first, but they do need to be building talent and capability, and watching closely to be ready to fast follow as great new experiences are created in their industry.

3. If you have high market share, play hardball

Michael Morton (analyst at MoffettNathanson) was interviewed by Ben Thompson of Stratechery in November:

I think Amazon should take their time with ChatGPT, because it’s important to underline that ChatGPT needs Amazon more than Amazon needs ChatGPT.

If you’re ChatGPT and you’re trying to present an e-commerce offering to the consumer, you can’t do it without 50% of the market. Amazon is half the e-commerce market, they have the world’s inventory and the world’s distribution network.

In any industry with a dominant leader, OpenAI faces the same conundrum. Smaller players like Walmart have an incentive to partner to try to take share, and they will be able to shave off incremental transactions. But if the largest players hold out, LLMs just can’t create a consumer offering that is very compelling.

Marketplaces with high market share have quite a lot of leverage. They should hold out entirely, or wait for great economics.2 And if they do partner, they should retain optionality by holding back their most valuable data, including supplier and product information and conversion and repeat rates.

4. Don’t get greedy

Amazon currently has a lot of margin that is looking like ChatGPT’s opportunity. Ads generate $60B in revenue and contribute ~all of the profit for the e-commerce business.

But it has also made the consumer experience much worse. More than half of product impressions are promoted, and it takes a long time to find what you’re looking for. They could get away with it prior to LLMs, but now it’s a huge vulnerability because the experience of getting the best product recommendation right away in ChatGPT is just so much better.

Amazon is one of the most egregious examples, but many mature marketplaces have been willing to degrade the customer experience in favor of monetization. If ads or upsells are making search worse, or if commission is too high and increasing prices for customers, marketplaces need to disrupt themselves before LLMs do it for them.

It’s going to be a tumultuous few years for marketplaces as the guard changes between Google and the LLMs. Customer acquisition dollars will shift fast, and then LLMs will come knocking for the whole profit pool.

But many marketplaces, particularly if they run the playbook above, are going to come out of it just fine.

Credits

Thank you to Casey Winters for his help on this essay. He is one of the top marketplace thinkers in the world and you should read his Substack.

Have you ever had to make the case for a new product or feature? Decide whether to launch a market or channel? Debated a major change to customer policy?

There is a single tool that can make you much more effective any time you need to align the team on a course of action.

It is sometimes called a 6-pager, a brief, a strategy doc, or a decision doc. Amazon famously uses one as a pre-read for every major meeting.

I’ve written hundreds of them and read thousands. And I have a template that works in almost every situation: CRISP.

• Context

• Requirements

• Insights

• Solutions

• Proof

This essay includes an overview of how to write one well, templates you can copy, and a prompt to instruct an LLM to write one for you.

Subscribe now

What’s in a CRISP doc?

For each section I’ll outline its contents as well as examples of what you might include for an illustrative project - a startup trying to decide whether or not it should launch in a new market.

Context

The goal of this section is to quickly get everyone on the same page. Why are we working on this? What prior work or assumptions is this building on?

In our market expansion example, you might include which past markets you’ve expanded into, how those launches went, and a reminder of a project the team did last year in which they decided to wait on launching more markets.

Requirements

This section aligns your audience on what success looks like. What are we trying to accomplish? What strategic principles do we want to align to?

This section might remind your readers that in order to deem a market launch a success, it must contribute 5% of revenue and be profitable within 24 months.

Insights

This section lays out what you recommend doing and why. It should lead with the recommendation, followed by a series of supporting arguments and evidence.

Here you might recommend that we launch the Australian market. To support this you could include an evaluation of multiple markets on a framework that includes market size, addressability, competitive overview, and level of effort. You could also include the projected financial results over the first 3 years of the market’s life.

Solutions

This section lays out how to deliver against the recommendation. What new GTM initiatives must be launched, what new products or features do we need to build? For each initiative, include the owner, timing, and cost.

This might include hiring a local sales team in Australia, enabling new payment methods in the app, and increasing the size of the customer support team to handle new customers.

(Additional) Proof

This section is a way to provide additional support for your argument and anticipate questions that may come up in discussion, but it is not required to understand the overall findings and implications of the work. It might include additional analysis, customer and market research, and other resources.

This section could go into more detail about how you sized the markets, what you learned about each competitor currently operating in the market, and the assumptions behind your financial projections.

Length

Amazon settled at 6 pages as the optimal length for a doc. That’s about 3,000 words, 500 words per page. But that should be considered a maximum — the length required for the most complex projects which are setting the strategic foundation for a major new initiative or part of the business. Most should be shorter.

The context and requirements sections should collectively take up about half a page. Insights is typically 1-2 pages. Solutions is another page. Everything else is additional Proof.

Some might be asking: where is the executive summary? If you write a doc like this well, you don’t need one. Anyone can read the first four sections (a maximum of 3 pages) in about five minutes and understand the entirety of the findings and recommendations.

It’s of course very hard to write a concise doc, but doing so will both make your writing better and significantly increase the chances that it will be read and understood.

Answer first

Let’s zoom in on the most important section of your doc: Insights. This is where you turn the corner from what’s been done in the past (Context and Requirements) to what you recommend doing in the future (Solutions). It’s also the most contentious, because it’s where you’re trying to change people’s understanding of the business.

To do this effectively, it should be laid out deductively - meaning start with the answer, followed by supporting evidence. Barbara Minto outlined how to do this in her Pyramid Principle:

Here’s what this might look like for our market launch example (the final version would be in paragraphs, not bullets):

We recommend launching in Australia

1. It meets our investment criteria

   1. It will contribute 6% of revenue after 24 months

   2. We expect the market to generate contribution profit within 18 months

   3. …

2. There is limited competition today, but new startups are emerging quickly

   1. 95% of market share today goes to older, non-technology companies

   2. However, there are 3 emerging tech-first competitors we’ve identified, and we know from past market launches that being early is important to success

   3. …

3. Customers in the market are similar to the US, reducing the effort to launch

   1. Aside from a few new payment methods, the Australian market will require limited localization of our app

   2. The types of products Australian customers buy and the way that they shop is similar to US customers

   3. …

This allows your readers to very rapidly understand the points you are trying to make, at different levels of fidelity:

• They can read just the first sentence to understand that the recommendation is to launch in Australia

• They can read the first line of the three supporting arguments to understand why you’re recommending this

• They can read the bullets underneath the supporting arguments to understand the rationale for each argument

Clear writing = clear thinking

Why is all of this done in long form prose, instead of in powerpoint or bullets?

The answer is that writing in long form benefits the writer as much as it does the reader. Writing in prose has a magical quality of forcing you to figure things out, because there is nowhere to hide.

That means that the CRISP template is not something you bolt on at the end when you’re basically “done”. It’s a way to help you think.

You should start writing it right away, beginning with what you know about the context and your hypotheses as to the insights you will uncover. And then as you get into the actual exploration — talking to customers, working with data, building an execution plan — you should keep updating the doc as you go to find weak spots. (For advice on how to do this, start here: How to be Strategic.)

The other implication is that over-reliance on LLMs will unavoidably impair your understanding of the problem. They can be a useful tool at the start (to generate lots of ideas) and at the end (to help you clean up your doc and put it in the right format), but you need to fully own and grapple with the messy part in the middle.

Templates

Here are templates you can copy in Google Docs and Notion.

Additionally, here is a prompt you can use to instruct an LLM to translate a rough doc into a finished product using the CRISP format.

Thank you to Alexa, Ali, Jolie, and Sam for their feedback on a draft of this essay.

Vibe analysis could be an even bigger deal: there are about 2 million software engineers in the US, but at least 5 million people who use data to answer questions every day.1

That means that in the US alone, we’re spending 10 billion hours a year reporting on business performance, assessing new products and features, and deciding which experiments to ship and which growth opportunities to pursue.

I worked with some of the team at Faire who are at the edge of applying AI to analytical work — Alexa, Ali, Blake, EB, Jolie, Max, Sam, Tim, and Zach — to shed light on the change that is coming.

We’ll look at both the bull case (how AI could massively increase the efficiency and quality of data analysis) and the bear case (why it will be harder than many people think).

What becomes clear is that no matter how conservative your assumptions, within a few years the way analysis is done and who does it will be unrecognizable from today.

Subscribe now

The bull case

Data analysis is full of the kinds of things that humans are bad at and machines are great at. There are basically four components.

The hardest part is often simply knowing what data to use - understanding the schema, how different tables and fields interact, and what is up to date. If you hook up ChatGPT to a data warehouse today, you get a tool that is pretty dumb out of the gate but gets smart quickly as it develops a semantic model of the dataset2. Instead of asking each team member to learn this for themselves, you can ask a model to learn it once.

Another component of analysis is writing SQL queries themselves. This is such an obvious use case that off-the-shelf LLMs are already very helpful at quickly cleaning up and generating queries. Cursor is pretty good at it too. Products that are purpose-built for data analysis will be excellent at it.

The third component is manipulating data into a useful format. Some of this can be done through the query, but many forms of analysis require a secondary tool like spreadsheets. New solutions for this are exploding, such as the viral launch of Shortcut (a “superhuman Excel agent”) just a few weeks ago.

Finally, there is visualizing and dashboarding data. This is probably where the tools are weakest today, but there are sparks of genius. All of the charts below were one-shotted by Claude based on some data and a quick description of the format:

As incumbents and startups race to build solutions to these problems, two distinct UIs are emerging:

1. “Cursor for analytics” where the core workflow is autocomplete, editing, or refactoring of existing code. Startups like NAO and Galaxy are building for this use case and incumbents like Mode are incorporating it into their products.

2. “Data chatbots” where the core workflow is natural language conversations that output basic datasets and charts. Many incumbents are building this, including Snowflake and Looker.

Today the former is accessible only to more sophisticated users, and the latter just have very limited capabilities. There are strong incentives for a tool that does both, because this would allow the work of power users to tune the semantic model for the benefit of everyone else.

This tool will also need built-in visualizations in order to avoid analysts constantly jumping between workflows, and to enable static dashboards and charts for sharing.

The space is begging for a full stack solution which:

1. Is native to your data warehouse and holds as much of it in context as possible

2. Constantly updates its semantic model of the data schema as it is used

3. Provides data and visualizations by default, but exposes SQL to those who want it

4. Has a built-in visualization tool that allows rapid iteration on charts

A product like this would supercharge analysis, collapsing the time it takes by 5x or more. There would no longer be distinct steps required to figure out what tables to use, write a query, slice the data, and visualize. You could just… talk to your data. You could vibe with it in the same way Karpathy imagined vibing with code

It wouldn’t stop there. Give a tool like this access to email, Slack, and Notion, and now it can learn not only what analysis gets done but what action it leads to. It would be a context sponge for how the business is performing and what is working and not working.

It would put us on a path that looks something like this:

1. Augmented Analyst - This is where we are today. AI is speeding up certain tasks, providing power users a productivity improvement on the order of 20%.

2. Accelerated Analyst - End-to-end tools like the one we describe above arrive. Analysis speeds up by 5x or more and is accessible to many more people.

3. Agentic Analyst - You can delegate a basic end-to-end analysis to an agent entirely - no human in the loop.

4. Autonomous Analyst - Instead of assigning analysis to an agent, it constantly assesses the business and prioritizes analysis to run on its own.

All of the pieces needed to achieve the Accelerated Analyst phase are in place, we just need to build the products and test them inside companies. A 1-year timeline to get to a reasonable product in the wild is feasible.

From there, I will borrow the timelines proposed by Dwarkesh Patel, which are significantly more conservative than something like AI 2027. He outlines two milestones:

• 2028: ability to complete an end-to-end knowledge work tasks of moderate complexity. This maps to the Agentic Analyst phase.

• 2032: “continuous learning” ability (the product can accumulate tacit knowledge, refine approach, and internalize feedback). This is the point at which the AI is acting autonomously.

The Vibe Analysis timeline:

The bear case

There is just one problem: real analysis is a lot messier than we’ve been describing.

The example that Dwarkesh gives of an end-to-end knowledge work task that AI could do by 2028 is filing taxes for a small business. That activity is guided by a set of rules and has an output that is verifiably accurate.

Most data analysis is neither of those things — it’s an exploration of an undefined problem space with an undefined range of outcomes. This means that (1) the actual efficiency gains will be lower in the Accelerated Analyst phase and (2) it will take longer to get to the Agentic Analyst phase.

So far we’ve only been talking about the part where you’re pulling or manipulating data — the “analysis” itself. But to actually go all of the way from question to insight requires a bundle of four things, and machines are less good at the other parts.

Hypothesis is the phase at the beginning where you use accumulated context and intuition to answer the question “what do we expect the answer or range of possible answers to be”? It’s maybe 10% of the total amount of time spent, but a huge amount of the leverage.

Diagnosis is where you apply everything you’ve learned from the data to form “the answer”. It is probably 20% of the time investment today.

Both of these will be highly resistant to change. They require the generation of novel ideas, which is a particular weakness of the current generation of AI products. Certainly tools today can help with this, for example by generating an exhaustive list of possible answers or pressure testing the best ones. And if the core constraint here is simply “context”, won’t models get better as they learn more and more about your company? Probably. But let’s be hyper conservative and say that these two steps don’t get more efficient at all.

Synthesis is the phase at the end where you put everything into a format that is clear, structured, and can be shared, like a doc or a presentation. This is the final 20% of time spent today.

Part of creating a final document is actually clarifying your understanding of the problem. Writing is thinking, and that piece won’t get much faster with AI. But there is also a big part of it that is simply a “clean-up” effort to structure and trim the doc, make it easy to read, and ladder supporting insights under each claim. If AI is great at one thing, it is great at summarization. It will become trivially easy for it to generate a doc in whatever format and level of depth you want.

At Faire we built a custom GPT with templates, examples, and writing principles that can translate a rough “answer” into a full output doc, and help you improve that doc over a few iterations. This will get much better as more specialized tools are built for this purpose. It’s relatively conservative to say that synthesis will get about twice as fast.

If you’re keeping score, we’ve said that:

• The analysis phase gets 5x faster (not automated)

• Synthesis gets twice as fast

• Hypothesis and diagnosis don’t change at all

Even with these relatively conservative assumptions, we’re talking about speeding up the entire process by 2x and significantly shifting the composition of work away from technical data skills and toward judgement, intuition, and communication.

Some kinds of analysis will collapse even more, depending on how much judgement they require. On one end of the spectrum you have more rote analysis like analyzing the results of an experiment or generating a weekly performance report. On the other end you have the assessment of a greenfield expansion opportunity or figuring out which teams to resource.

While the latter requires a lot of judgement, it’s possible that AI could compress the former by much more and get to the agentic state faster.

Predictions

No matter how conservative you try to be, it’s clear that analysis is going to change a lot, and that this will probably happen inside of five years. What does that mean?

Analysis will decentralize

Tools like Looker, Sigma, and Mixpanel promised to democratize data by abstracting SQL and letting anyone play with drag and drop UIs. But they all tended to be limited to certain datasets, applicable mainly to standardizable parts of a product like growth funnels, and quite brittle and time-intensive to maintain.

AI tools will deliver on this vision for real. Anyone will be able to just talk to the database and get an answer. Many more people will do analysis, much more often. The CEO will be able to understand how a new product is performing in just a few minutes. Any salesperson will be able to generate a report to close a new prospect on the fly. Engineers might (gasp) not need to talk to a product manager to figure out what to build next.

The role of analyst will transform

When hiring analysts today, companies are looking for both technical skills and strategic judgement. It’s possible but very difficult to find people that are excellent at both. A lot of companies give up and hire analysts who spike on technical skills but lack judgement, and rely on product managers or others to “be strategic”.

As the more technical parts of the job are abstracted, the purely technical analyst will become significantly less valuable. Analysts will be hired primarily based on their ability to apply judgement and answer hard questions.

In the process, analysts will become massively more impactful. We will start to see “super ICs” that can do the work that a manager and team of 6-7 people used to do.

I suspect this will not mean fewer analysts, just that their collective output is much greater. Jevons Paradox (which says that increased efficiency of resource use often leads to more consumption of that resource, not less) is very much at play. I’ve never seen a company run out of questions to ask.

Companies will grow faster

All companies are effectively trying to turn this flywheel as fast as possible:

They are learning about their customers, the market, and how the product is performing. Then they are using these insights to build and sell new things. Repeat.

As AI makes engineers more productive and code less scarce, the constraint on the right side of this flywheel is lifting. The bottleneck is now on the left side.

In a world where you can generate insights twice as fast and many more people can do it, this whole flywheel can again get unstuck. Vibe analysis will accelerate the rate at which companies learn, and thus how fast they can grow.

2

Here’s ChatGPT explaining how its integration with Faire’s data works:

The CEO asked me a question which has rattled around in my head ever since: do you want to build, or do you want to sell?

My experience since then has added some nuance to that framework. Some jobs are a mix of more than one - for example the role I landed on after that conversation was in growth, which is a mix of building and selling.

And today I find myself in a job which is neither of those two: I lead the Strategy & Analytics teams at Faire, whose primary job is to understand.

Add this up, and you get a simple framework that can be clarifying for anyone trying to figure out what path to go down. There are only three fundamental jobs at any company:

1. Building - creating the product or the company itself

2. Selling - getting more customers to use the product

3. Understanding - learning more about what to build or sell

They form three interlocking circles, and every function can be mapped somewhere within them:

Some functions are closely aligned to one job. Most of engineering and operations is about building the product or the company. Most of marketing and sales is about driving growth. And most of analytics and research is about better understanding what to do.

Some functions are hybrids. Growth emerged as a way to combine the jobs of building and selling. Designers are tasked with not just designing the product but deeply understanding the customer. PMMs play a similar role between selling and understanding. The job of founders and general managers is often to do all three.

If you’re currently thinking about what path to take, I’d offer three meditations on this framework:

Ask yourself where you derive joy (and what you can put up with).

I began my startup journey in growth, but found myself constantly pulled toward roles in analysis, strategy, and investing. I realized it is because I love the “understand” part of the job - the process of going from hypothesis to conclusion. I find joy in ideas, learning, the hunt for the right answer. And I can put up with the most common criticism of this kind of role, which is that is one step removed from the execution. You often have to be OK with someone else acting on your ideas.

Some people like building beautiful products or designing efficient operations. Builders find joy in the act of creation - bringing something new or better into the world. But these roles usually come with a heavy logistics burden. There is just a lot of project management and complicated coordination between many stakeholders. When people first become PMs they often love the strategic side of the role but are surprised by how much blocking and tackling there is. To be great, you have to at a minimum put up with this, and better yet enjoy it.

Other people like figuring out how to craft the perfect message or close the customer. Sellers find joy in the deal, in winning hearts and minds, in making the numbers go up. But it often means you play a role that is downstream from the people building the product. You have to be comfortable evangelizing someone else’s vision.

There are other pros and cons to these kinds of roles, of course. The point is to reflect on what those are, and make sure you love the upsides and can tolerate the downside.

The job of selling comes for most of us eventually.

Recruiting, managing, and working cross-functionally are all mostly about selling. You’re trying to convince people to take a job, stay in a job, or do something they might not otherwise.

That means that if you stay in almost any function long enough, you’re probably going to end up doing a lot of selling.

In fact, one of the most common sources of mid-career frustration is when builders or understanders realize that they are being asked to get into sales. “Why am I in so many meetings? Why do I spend so much time interviewing?”

As a result, everyone needs to either (1) embrace the job of selling and get good at it or (2) intentionally steer themselves toward a role that doesn’t require much of it.

The main way to do #2 is to pursue a senior IC track, such as becoming a staff engineer or designer. Many functions don’t have an equivalent role today, but as AI tools magnify the output of the best ICs, we’ll see “staff” functions emerge in many other fields like marketing and operations as well.

Consider where the puck is going.

At most tech companies today, the split of building, selling, and understanding is probably about 50%, 30%, 20%. The engineering team is usually the largest, and the sales and marketing team is the second largest.

But AI is improving the efficiency of building and selling at a much faster rate than it is impacting the “understand” part of the job. We’re already seeing PM to engineering ratios fall. AI tools are beginning to do a lot of front-line CX and marketing work, and it will move up the stack from there.

What it hasn’t done a good job replacing yet is the generation of novel insights or answers to hard problems. The shift over time will be toward greater and greater share of time spent on the understand task. That means that whatever role you are in, it’s likely to become a bigger part of your job, and it is worth investing in.

But being pretty good at many things is not a stable equilibrium, because most customers only care about one or two of them.

“In business we often find that the winning system goes almost ridiculously far in maximizing and or minimizing one or a few variables – like the discount warehouses of Costco.” - Charlie Munger

Costco (and Walmart and Temu and Shein) won on price. Amazon won on convenience. Nordstrom and D2C brands won on quality. Independent local retailers won on experience. They collectively stole market share until there was none left, and department stores died.

This is not just about retail. General-purpose products are dying everywhere. In this essay we will cover:

1. Why every industry is following the same pattern

2. Two more examples of this trend from media and education

3. Why AI will be the technology that finally reverses the trend

Why is this happening?

Many people describe what is happening as the “barbell” effect. The theory is that in every industry, both very large companies and very small companies win, and everything in the middle dies.

This theory is incorrect, or at least incomplete.

Michael Porter has a series of tests of good strategy which do a better job of articulating what is really happening. He said that in order to win, a company must have:

1. A distinctive value prop: a unique benefit to customers

2. A tailored value chain: a company designed to deliver those benefits

3. The ability to make trade-offs: choosing what to do and what not to do

The first two criteria are what he called “the core” of strategy. Pick a consumer value prop, like convenience, then build a business fully optimized to deliver it, like Amazon.

The third test explains why Amazon can’t take over the entire retail market. It’s not just difficult to provide everything customers want simultaneously; it’s impossible. To deliver the highest convenience, Amazon had to build a supply chain optimized for delivering small quantities of products very fast, which is more expensive than what Costco or Temu built. It has to stock as many products as it can, which means it can’t vet quality or protect the customer experience as well as Nordstrom or D2C.

The result is a predictable pattern as industries mature. They start with general purpose products in the middle. Then a new technology allows an incumbent or a startup to deliver better on a specific benefit. When others see it is working, a new cluster forms around it. Each company gets better and better at optimizing its chosen strategy. The clusters become increasingly tightly packed and farther away from each other. Anyone who tries to stay in the middle dies.

Two more examples:

Media

One thing people want from media is for it to cater to exactly their specific interests and tastes. That’s YouTube and TikTok.

Another thing they want is shared cultural experience. Something to laugh and cry and bond over with friends and family. This segment is thriving: Hollywood movies are increasingly high budget and focused on stories and characters everyone already knows. Paris hosted the highest rated Olympics ever. The NFL had 70 of the top 100 broadcasts last year.

Some people want to feel smart, intellectually stimulated, high status. Prestige TV (Succession, The White Lotus) pushes the boundaries of production quality and subject matter to meet that need.

Sitcoms and variety shows were once the general-purpose products of this industry. Now they’ve been relegated to the much smaller use case of mindless or “passive” entertainment.

Education

Universities are a general-purpose product. They started losing share in 2016, a trend that will accelerate because most of them do a mediocre job at many things.

One thing universities offer is prestige: the ability to be surrounded by elites and become one by association. This will continue to be solved by universities, but probably only the top 10 or 25 of them in the US that actually do it well.

Another is credentialing: proving you can do a job to potential employers. Online programs and employers themselves will be able to do this much more effectively.

And of course there is actually learning things. Cohort-based programs, Youtube, and Substack are slowly winning at this.

What comes next?

In every story so far, the internet removed a bottleneck in distribution, allowing new types of products to work and further decentralizing the industry around distinct value props:

• The internet removed the physical constraints on how much inventory you could put in a store, enabling Amazon to build the “endless aisle”

• It removed the constraint of 3-4 big TV networks with a limited amount of programming and allowed new providers to “go over the top” with many more kinds of content

• It removed the constraint of the physical college campus, allowing students to learn in many new ways

AI will not just impact the way products are distributed, but the nature of products themselves.

AI’s core strength is the intelligence and flexibility to tailor outputs to each customer’s individual needs. This will enable products that can be great at multiple value props simultaneously, causing a re-centralization of industries around fewer clusters led by new, massive general-purpose products.

In retail, the clusters we see today around quality and convenience will collapse back into one. The ability to offer a massive selection without undermining quality is a ranking and personalization problem. If Amazon could deliver exactly the right product for each customer and use case, they could protect the high-end experience for some customers without eliminating the massive selection for others.

To achieve this, e-commerce products may start to look more conversational, gathering more input on exactly what each customer is looking for and using this to deliver more tailored results.

In media, platforms will emerge that can take any user prompt and create bespoke video, audio, and text catered entirely to that person’s tastes and interests.

That will clearly do the job that YouTube or TikTok are doing, but it could also swallow the passive and prestige clusters as well, because it will be able to achieve similar or better production quality at virtually no marginal cost.

The only cluster that these products will find it hard to crack is shared cultural experience, because the whole point is that everyone else is watching it too. However you could imagine this need ultimately being served by pieces of content that were originally generated for individuals going hyper viral (a dynamic we’re already starting to see with TikTok).

What happens in education might be the most extreme of all. Superintelligent AI tutors will provide 1:1 learning catered entirely to each person’s goals, interests, and learning styles. They will generate video, audio, and text on demand that is massively better than what universities, online courses, YouTube, or Substack can provide today. And they will know exactly how far each learner has progressed, so they will become much better at credentialing as well.

Prestige is hard to address, because the whole point is that most people can’t have it. But it’s not hard to imagine a world where there is a small set of elite universities for a very narrow audience, and AI super-tutors for everyone else.

In each industry, Porter’s 3rd test of strategy starts to collapse. There will no longer be trade-offs between things like convenience and quality or between personalization and production value. A new kind of product will emerge that can be great at many things simultaneously. Welcome back to the era of general-purpose products.

Thank you to Brian Balfour and Zach Grannis for their feedback on this essay.

Create functions like product and marketing. Hire functional leaders, and delegate decision-making to them. As the team grows, hire managers within each function and further delegate decision-making. As the business develops distinct products, break them out under GMs and replicate the whole structure of functions and managers.

This structure allows a company to make more decisions. The team can scale quickly and focus on many things at once. As accountability is pushed further into the company, it becomes less reliant on the founder until it is a machine with interchangeable parts.

But it comes at the expense of the quality and speed of decisions. Good founders have good judgement, and this reduces their ability to exercise it. If they need to change the direction of the company, it takes much longer and may fail if they have to cascade the message through many layers of management.

This is often not a good trade. It’s certainly not a good trade when the market or technology is changing rapidly, which puts a premium on being right and being fast.

In Founder Mode, Paul Graham wrote about a different way of running a company and how little we know about it:

"There are as far as I know no books specifically about founder mode. Business schools don't know it exists. All we have so far are the experiments of individual founders who've been figuring it out for themselves. But now that we know what we're looking for, we can search for it. I hope in a few years founder mode will be as well understood as manager mode.”

One way to understand founder mode is to understand the kind of organization you need to design to achieve it.

Conventional wisdom is to diffuse decision-making as much as possible throughout the org. In contrast, founder mode organizations maximize the number of decisions made by the founder.

This famous parody of org charts from 2011 was meant to expose the particular dysfunction of each of the tech giants at the time:

Apple’s “dysfunction” was that red dot: Steve Jobs. He was intentionally positioned as an organizational bottleneck. Despite having many product lines, there were no business units - every function rolled up directly to him. He was directly involved in the details of every launch.

It’s very hard to argue that this was actually dysfunctional. In 2011, Apple had been on a ten-year tear, launching the iPod, iPhone, and iPad and growing their market cap by nearly 100x to the 2nd largest in the world.

Examples of orgs designed this way are rare. But if you look at just the companies that are leading in the highest growth categories - the best of the best - they are less rare than many people think. Three current examples:

Airbnb

Airbnb was on its way to becoming a “traditional” big company. But when it lost 80% of volume overnight during the pandemic, it gave Brian Chesky an opening to run the company the way he wanted to.

He eliminated the business unit structure, and all functions reported to him. He took out layers of management.

Everyone put what they were working on into a single Google Sheet, and then they cut 80% of it. All of the remaining projects got put on a cadence - Brian would meet with them every 1, 2, 4, 8, or 12 weeks.

This meant that every week, Brian would see what he called a “semi assembly of the entire new product”. He could identify problems and directly force resolution.

In his own words: “I stopped pushing decision-making down, I pulled it in.”

Shopify

Tobi Lütke is famously executing against a 100 year vision. But the plan to achieve that vision evolves much faster.

He reviews the entire roadmap with each team every 6 weeks. Powering these meetings is a custom-built tool that collects information on every project happening across the company. It is focused on what is being built and how the engineers are building it, not on metrics. Teams can’t bring anything except what is in the tool to the meeting - no framing, no presentations.

He then decides what to keep and what to kill, and allocates a certain number of engineers to each team. The rest of headcount planning is done in a GitHub repo based on ratios with engineers.

Nvidia

Jensen Huang has designed the flattest organization possible. He has more than 40 direct reports.

Every meeting is open to whoever wants to join - he shares his perspective as widely as possible so that it doesn’t have to be cascaded through many layers of the org.

In turn, he gets his information from everywhere in the org, not just from leaders. Anyone in the company can email him the "top five things" on their minds. He estimates that he reads 100 of these emails every morning.

There is no formal planning cycle; he evaluates changing market conditions continuously and makes changes whenever needed.

These examples are highly idiosyncratic and built around the founder and the way that they think. But commonalities across them can help us deduce a few general principles for founder mode org design:

1. They give the founder as much surface area as possible to engage directly with the team (flat orgs, many direct reports, open meetings, often no business units)

2. They give the team as much surface area as possible to feed ground-truth information back to the founder (weekly standups, open communication, custom-built docs)

3. They allow the founder to overwrite the plan whenever needed

Clearly, this approach has limitations. The biggest is that founders must apply constant force to sustain it. No matter how strong their judgement is to start, they only maintain a decision-making edge by continually deepening their knowledge of the customer, the business, and the org. Their strength comes from being the only one who can truly “hold the whole thing in their head”.

This model also works better for decisions about innovation: what new things to build and how to build them. It’s not always necessary or productive for decisions related to doubling down on a known path.

Amazon had a lot of the latter type, and is known for a more traditional structure. Bezos has been openly critical of the “genius with a 1,000 helpers” model of organization design.

But you better believe that Bezos overrode the team on the decisions that really mattered, especially when there was a lot of uncertainty. Here he is at NYT Dealbook:

“I’m a very easy person to influence. I change my mind a lot. But a couple of percent of the time, no force in the world can move me. I can’t tell you how many people tried to talk me out of Fulfillment by Amazon. But I said: you guys will never talk me out of this. We are going to do this. I will do this through sheer force of will.”

A more nuanced version of the principle: founder mode organizations maximize the number of decisions made by the founder, up to the limits of their ability to make better or faster decisions.

Provided they have the humility to recognize those limitations, more founders should design their orgs this way. In fact, if a company is still founder-led and they’re not breaking the so-called “best practices” of org design, it’s a very bad sign.

I work at a 1,000-person wholesale marketplace called Faire. When people join, they’re often surprised at how many everyday decisions our CEO Max makes. At any given time, he’s embedded in at least 10 projects, meeting with the team weekly or even daily and getting into exactly how the product is going to be designed and exactly which words we’re going to use to talk about it with customers.

The teams on these projects have to work harder, and their decisions are more likely to be overridden.

Which brings us to the best counter-argument for founder mode org design: won’t it disempower the team? Don’t great people want autonomy?

Perhaps counter-intuitively, the teams on these projects tend to gain energy, not lose it. The explanation seems to be that in exchange for some autonomy, they get a great deal of clarity. This is a point Brian Chesky made on Lenny’s Podcast:

Way too many founders apologize for the way they want to run the company. They find some midpoint between how they want to run the company and how the people they lead want to run the company. That’s a good way to make everyone miserable. Because what everyone really wants is clarity.

Clarity is the only way to move fast enough to win. Moving fast is fun. And winning is what people want most of all.

Subscribe now

Thank you to Archie, Lenny and Max for their thoughts on a draft of this essay.

The business school version of why DoorDash won is that they had the right strategy. They launched in the right markets, acquired the right restaurants, and designed the marketplace the right way.

The Silicon Valley hustle culture version is that they out-executed everyone else. They just shipped faster until they had better selection, a better product, and more reliable delivery.

The financial markets version is that they got lucky. Grubhub and Uber were both public and playing with a hand tied behind their back during the most pivotal moment in the fight.

The reality is that you can’t understand what happened without all three perspectives. Increasingly, success in every competitive market will require the right strategy, rapid execution, and good luck.

Subscribe now

Strategy

Strategy turns on a few big decisions. For DoorDash, three of them really mattered.

The first was recognizing that owning delivery was the key to unlocking supply. Most restaurants can’t support the economics of running their own delivery fleet, so the Grubhub model of just routing orders to restaurants was bound to hit a wall.

DoorDash was the only company that launched with the business model that everyone now uses. The initial idea for Uber Eats was to load cars with fresh meals made at scale so they could be delivered as fast as possible. Postmates started with packages, not food.

DoorDash recognized the importance of logistics from the start. This was CEO Tony Xu’s final slide at YC demo day in 2013:

The second big decision was seemingly at odds with the first. Rule #1 of building a logistics business is to increase network density and thus driver utilization. This led everyone else to major city centers.

But DoorDash realized the suburbs were a better place to start. There, the alternatives to delivery were much worse, customers were more affluent, and average order values were higher. Customers immediately understood the value prop and spent enough to make delivery economics work.

Most importantly, no one else was in the suburbs yet. This is a key principle in Sarah Tavel’s Hierarchy of Marketplaces:

To win in a market, you can’t just be the market leader. You have to be the market leader by a lot. This creates a flywheel in which you have much more demand, which allows you to bring on more supply, which in turn compounds your demand advantage. Every market in food delivery was ultimately hard fought, but DoorDash’s initial wedge in the suburbs gave them an advantage no one else had.

Finally, DoorDash recognized that the most important thing was a wide selection of restaurants, even if that meant sacrificing other things that customers cared about, like delivery speed and price.

They realized that as long as they could deliver in about 40 minutes, there were limited gains from being faster. Uber tried to optimize for faster delivery, which led to the wrong decisions, including launching with the wrong business model in the first place.

DoorDash was also willing to initially make the service more expensive for consumers, which allowed them to give commission breaks to important restaurants to convince them to join. Uber instead wanted a flat customer fee, which required them to play hardball with restaurants. Later, DoorDash did begin to win on price (in particular through DashPass and lower markups on food items), but only once they had great selection.

Execution

DoorDash had the right strategy, and still almost failed.

They exploded out of the gate, raising a Series A from Sequoia in 2014. Legendary investor John Doerr effectively came out of retirement to lead their Series B in 2015 at a $600M valuation.

But then the music stopped. They were burning cash fast, and Tony couldn’t find a lead for their next round for six months. In 2016, Sequoia ultimately had to step in and lead their Series C at a $700M post-money valuation—a down round. By the end of 2017, they were almost out of money again and had to do a $60M bridge round just to keep the company alive.

Uber and Grubhub had much deeper pockets during this time. DoorDash only survived through incredible speed of execution.

The culture starts with Tony and the leadership team he built. Execs like Prabir Adarkar (President and COO), Keith Yandell (CBO), and Jessica Lachs (CAO) are cut from the same cloth - relentless, data-driven, and capable of operating at the lowest level of detail.

They learned to do many things at once. This is Keith Yandell on the Crucible Moments podcast:

People would come in and say, “Well, do you want to focus on growth or do you want to focus on profitability?” And we quickly realized that if we were going to survive this time, we had to do both at the same time. Eventually, we just made it a core value. It’s “and,” not “or.” You can’t pick or choose. We must grow, and we must become more profitable.

This speed translated to better metrics, but also to a better customer experience. Most customers in the early days of food delivery would switch between apps, but eventually became loyal to DoorDash.

When you ask them why, you hear things like “It was just more reliable”. The ops team focused on making deliveries a little faster and reducing defect rates every week. They scrutinized the quality of every one of their restaurants and dashers.

Or you hear “The app was just better”. The product team made sure it was easier to use, less cluttered, and the checkout was faster. The photos were bigger, it was easier to select items, and easier to re-order what you had last time. They were first to let you see where your dasher was so you knew when you were going to get your food.

They launched DashPass before Uber launched Uber Rewards. They launched the Chase Sapphire partnership before Uber launched the Amex partnership.

All of this added up to better and better customer retention, and markets began tipping in DoorDash’s favor.

Luck

By 2018, the business was on track to be a great outcome, but a little luck ensured it became a monster one.

The period from 2018-2019 was the most crucial, because it determined the footing each company was on going into the pandemic, when food delivery demand soared.

What was happening during this time period? Grubhub was already public, and CEO Matt Maloney had sold the markets on an efficient, asset-light model, a position he would continue to double down on as late as 2021:

“[Food delivery] is and always will be a crummy business”. Everyone else in the industry is doubling down on their logistics plays and talking about how smart they are and what a great technology company they are. I think the right choice is to be a better restaurant company.”

It would have cratered their market cap to take on the burn required to pivot to DoorDash’s model. Instead, they stayed asset-light and posted EBITDA margins of 14% ($186M on 1.3B in revenue).

Uber was recovering from #DeleteUber and a CEO change. They went public in May 2019, which forced them to cut back on their famously high burn. The combined business had EBITDA margins of -19%.

Meanwhile, DoorDash let it rip. They raised $535M from Softbank in 2018, another $250M later that year, and $1B across two rounds in 2019. In 2019 they lost $475M, a whopping -54% EBITDA margin.

They leaned heavily into market launch (including 5x-ing their market footprint in 2018 alone) and into customer acquisition, including running major brand campaigns. From 2019-2021, DoorDash spent more than $3 billion on marketing and sales.

This catapulted them from a distant third in 2017 to the market leadership position in March 2019. As the world shut down and food delivery became critical infrastructure in early 2020, DoorDash capitalized and compounded their gains.

Things might still have gone differently had Uber been able to acquire Grubhub in 2021 and built a combined business to rival DoorDash. Instead, Uber was outbid by Just Eat, which had no presence in the US and sold the business in 2024 at a loss.

Doordash had the right strategy, shipped fast, and was able to take full advantage of a once-in-a-generation tailwind.

We could formulate a crude success equation as:

The first term must be strategy because without it, you’re running in the wrong direction.

But this is raised to the power of execution velocity. Shipping things is a way to pressure test and improve your strategy. Ship fast, and you’ll quickly find new things that work and compound your gains. Ship slow, and you’ll get passed by competitors, even the ones that started with the wrong idea.

Luck is a force multiplier - a good break will magnify the outcome, and a bad one can turn a company into a zero.

Twenty years ago, it might have been sufficient to have just two of the three. But today, the tools available to startups are better, information travels faster, and venture markets are more liquid. Every big market is going to be full of well-funded, aggressive competitors.

To win, you need to be in the right place at the right time, have the right idea, and still run very fast.

Sources

1. The Changing Market for Food Delivery - McKinsey, 2016

2. The Complete History and Strategy of DoorDash - Acquired, 2020

3. The Hierarchy of Marketplaces - Sarah Tavel, 2020

4. Food Delivery Wars: 3 Takeaways From The UberEats, Postmates, Grubhub, DoorDash Ecosystem - Sarah Tavel, 2020

5. DoorDash from application to IPO - Paul Buchheit, 2020

6. Thread on why Doordash won - Michael Bloch, 2020

7. DoorDash and Uber Eats Are Hot. They’re Still Not Making Money - WSJ, 2021

8. Ordering in: The rapid evolution of food delivery - McKinsey, 2021

9. DoorDash ft. Tony Xu – The “Wrong” Moves That Built a Giant - Crucible Moments from Sequoia, 2024

10. Which company is winning the restaurant food delivery war? - Bloomberg Second Measure, 2024

11. DASH, UBER, GRUB public filings

Share

At their best, network effects are a source of deep defensibility that make markets winner-take-all. At their worst, they’re trivially easy for anyone else to replicate, leading to zero-sum competition for market share. What’s the difference?

Any defensive moat can be overcome with enough capital. If someone was willing to give you hundreds of billions of dollars without expecting a return on their investment for twenty years, you could build a viable Amazon competitor. But no one will do that, because it would be too expensive and risky. So the question is not can a moat be overcome, but at what price.

In the case of a marketplace network’s effect, that price is closely approximated by how much it would cost to acquire and retain all of the supply needed to make customers happy.

There are three dimensions which determine that price: geographic reach of a market, heterogeneity of supply, and multi-tenanting. We’ll unpack them using the analogy of a bucket. The more water (supply) it is holding, the more it would cost someone else to fill their own, and the less likely they are to try.

1. Breadth (geographic reach)

The more narrowly a market is defined, the easier it is to aggregate enough supply to deliver a good experience to customers, and the less defensible it will be long term.

At the smallest end of the scale, you have local marketplaces like food delivery (Doordash) and local services (Thumbtack). Their markets are defined at the city or even neighborhood level because customers only care about the supply in their area. The best way to think about these businesses is not a single marketplace, but instead a collection of smaller, independent marketplaces running on shared infrastructure.

At the other end of the spectrum, you have national or global marketplaces, where a customer is selecting from supply across a much wider area. Most of e-commerce is at least national, and always expanding as cross-border freight gets easier and cheaper. For travel marketplaces like Airbnb, choosing between options in any country is a core part of the value.

The other important factor is whether customers move between markets or not. In the case of Thumbtack, they usually don’t - most customers probably only have a home in one area, so all they care about is whether or not coverage is good there. But for Lyft, it’s valuable for the product to work anywhere you travel, which makes it a competitive advantage to have many individual markets aggregated in one product. These markets are local, with a cross-market benefits.

2. Depth (heterogeneity of supply)

Once we’ve defined the size of the market, we have to determine how much supply it takes in that market to make customers happy. Again, the less it takes, the less defensible a network effect is.

On one end of this spectrum, you have a marketplace like Lyft. Supply is homogenous, meaning they all essentially provide the same thing, and customers don’t care much about who the supplier is. They care about wait times and price, and it doesn’t take much supply before wait times are less than five minutes and pricing is pretty efficient.1

On the other end of the spectrum is a marketplace like Amazon where supply is heterogeneous. Customers shop across thousands of categories, millions of brands, and hundreds of millions of products, and everyone wants them at different price points, quality levels, and aesthetics. The core benefit is selection and convenience - and that continues to improve even as you add large amounts of supply.

In the middle of the spectrum are markets that are heterogenous, but shallow. The customer cares who the supplier is, but is usually satisfied with <10 good options for a given search - think grocery stores on Instacart2 or home painters on Thumbtack.

Taken together, breadth and depth determine how much supply you need at any given time to make customers happy. The further up and to the right, the more defensible a network effect is.

3. Holes (multi-tenanting)

The final dimension is how feasible it is for suppliers to multi-tenant, i.e. use multiple marketplaces at one time.

If multi-tenanting is easy, a new marketplace can get traction even if someone else has already aggregated a lot of the supply in a market. Everyone will have to spend a lot to re-acquire or re-engage supply over time.3

It’s actually even worse than that sounds. As the first mover in an industry, a marketplace has to do a lot of work to vet supply, convince them of the value of the marketplace, and onboard them onto the product. If suppliers can easily switch, the next marketplace can draft off of all of that work and just acquire the same suppliers who are high quality and already bought in. This is part of what fueled the capital bonfire in the food delivery space.

There are natural dynamics in markets that make it harder for suppliers to multi-tenant. For example, if a supplier has a larger or more complex operation (truckers, manufacturers) it’s less likely they will use many platforms than simpler ones (drivers, landscapers). But unlike the other two dimensions, most markets aren’t inherently very different on this dimension.

Instead, the best marketplaces intentionally make it harder to multi-tenant over time. To do this, they find features that simultaneously add a lot of value for suppliers and increase switching costs.

For example, they could build integrations into the supplier’s business, such as with their accounting software. This saves them time, but also makes it less likely they’ll also integrate with someone else. Or they could go even further and take on a part of the value chain altogether. This is what Amazon did by spending billions of dollars to build the logistics network that powers Fulfilled by Amazon. That makes it much easier for suppliers to sell on Amazon, and also a lot less likely that they’ll use other marketplaces because a bunch of their inventory is actually sitting in Amazon warehouses.

The defensibility of network effects is the difference between incredible and brutal businesses

Pulling all of these dimensions together, we can see that network effects range from the deepest moats in the world to virtually no moat at all.

Amazon has aggregated an exceptionally heterogeneous long tail of supply across the world. And it has built programs like Fulfilled by Amazon and Prime that make it very hard for suppliers to engage in other marketplaces in the same way. The result is both an incredible customer experience and a very deep moat.

Contrast that with “Uber for X”, which failed almost everywhere it was tried. That is because those businesses all had a few things in common: they were hyper-local, focused on one or a handful of homogenous service categories, and didn’t provide an opportunity to build enough value for suppliers to keep them from switching.

If you’re considering founding, joining, or investing in a marketplace, first assess how much water their bucket can hold. It will tell you a lot about how defensible the business will turn out to be.

Thank you to Casey Winters and Lenny Rachitsky for their feedback on this essay.

2

Instacart moved up on this dimension. They initially built a network effect between grocery stores and consumers, but it was not highly defensible because the markets is not deep: grocery stores are relatively consolidated and thus do not require much aggregation. But Instacart’s advertising product has CPG brands on the supply side instead of stores. This market is much deeper and harder for someone else to aggregate, which makes their ad product more defensible and gives them dollars to spend that competitors don’t have.

Game One is optimizing for company outcome. Focus all of your energy on finding the highest leverage problems in your company and fixing them.

Game Two is optimizing for personal outcome. Seek external signals like title, scope, and team size. Make sure you’re on the right team and working on the right things.

If you’re a founder or working at a very small startup, it’s usually obvious that you should play Game One. Your outcome and the company’s outcome are close to synonymous, so if you want to succeed, you need to make the company succeed. And everyone knows who is pulling their weight, so the right people get credit.

If you work at a giant company, it’s usually in your best interest to play Game Two. It’s very hard to personally impact the outcome of the company enough that it will make a difference to you. Even if you do have an impact, it’s often not visible to the right people, so you might not get credit. As a result, everyone is looking out for themselves.

But the size of company at which it makes sense to play Game One is much larger than most people think. At almost all startups - even those with hundreds or thousands of people - it’s the right choice. It will result in a better outcome for the company, but also a better outcome for your career.

One reason is that startups are an iterated game, and the startup ecosystem is a small world. If you get a reputation for playing Game One, it will start to open doors. Conversely if you play Game Two, there will be growing dissonance between your external image and what references actually say about you when they’re asked. It can become career poison.

If you’re at a giant company now but ultimately plan to work for or found a startup, you should play Game One. It’s very hard to suddenly switch the entire way you operate, so many people who are used to Game Two have a rough time when they make the transition. They quickly get a reputation for having sharp elbows and lose trust.

But there is an even more fundamental reason to play Game One. You may think you know which title you want or which team you should lead or which projects to take, but you probably don’t.

Success is mostly about finding the intersection of what you’re great at and what the market needs. Figuring out what you’re great at takes a lot of iteration, and what the market needs is always changing. So if you sit down and try to chart out a 20, 10, or even 5-year career plan, it’s going to be wrong.

If you instead play Game One and let yourself get pulled to where you will have the greatest impact, it has a way of naturally putting your career on a good track. And as the skills and reputation you build start to compound, it will accelerate your career in ways that are very hard to predict.

The best counter-argument to Game One is simply that if you try to do it inside a not-great company or with not-great people, it often backfires. Despite all of your efforts the company might fail, or your peers or manager might take advantage of you.

There isn’t a good response to this argument. You just have to try very hard to work at great companies with great people. You should spend a lot of time thinking about which company to join and when to join it.

But once you’re there, play Game One. I think you’ll be surprised by how well it works.

The problem is not that there isn’t much strategic work to do, it’s that they are focused on the wrong kind.

What they have in mind is making big, consequential decisions - what Jeff Bezos called “one-way doors” because they’re hard to reverse. Launching Amazon Prime was like this. It required an attempt to see multiple steps into the future to develop conviction that it would work. These kinds of decisions happen rarely, and most people don’t spend much time on them, even if they’re on the senior leadership team.

However there is a different kind of strategic work in extreme abundance, and that is what comes after launching Prime. Thousands of smaller decisions that will determine whether or not it actually works. How do we convince people to sign up? Which suppliers should we onboard? How should we set up our fulfillment operation?

Being good at these kinds of decisions is a different thing entirely.

You’re not trying to see many steps in the future. You’re just trying to get clarity on what is actually happening right now, so you can decide what to do next.

And speed matters just as much as accuracy. You want to try to get these decisions right if you can, but you also want to give yourself as much time as possible to keep trying if you’re wrong.

It’s not a linear process, it’s a flywheel: 

This means that the metric you are optimizing for needs to have both a direction and a velocity. It’s something like “good decisions per hour”.

Making good decisions rapidly requires four inputs, and we’ll explore how to get good at each of them.

1. Intuition: short circuiting as much of a problem as possible and driving quickly to a hypothesis

2. Insight generation: a toolkit to rapidly validate or invalidate that hypothesis

3. Solutioning: understanding enough about execution that you can recommend the right things to try 

4. Synthesis: pulling everything you’ve learned into a story that is cohesive, true, and will lodge in people’s brains in just a few minutes

Intuition

Intuition is a machine that takes in questions and outputs good hypotheses as to their answer.

Ben Thompson of Stratechery was asked how he can possibly have an insightful take on what is happening in tech every day, often just a few hours after it happens. His response:

I have a framework - an overall view of the world and how it works. It’s like a machine. When a piece of news happens, I feed it into the machine and out pops a conclusion. If A happened, then B then C then D then E.

How do you develop this?

Intuition is one of those mythical things that it seems some people have and others don’t. But you can absolutely get better at it if you try, especially if you’re intentional about understanding the problem space at three levels of fidelity:

1. The customer. Most fundamental is understanding who the customer is and what they actually care about. You need a mental model of how they decide between alternatives, and what single or small set of variables matter most.

For example, Amazon figured out relatively early that convenience was the thing that mattered most. Specifically ubiquitous, free, fast shipping. Eugene Wei wrote about how clarifying this was: 

“You can't imagine what a relief it is to have a single overarching obstacle to focus on as a product person. It's the same for anyone trying to solve a problem”

2. The industry. The next level of fidelity is the surrounding industry. What is its cost structure? Who are the key competitors and on what variables are they competing? Which of those variables are fluid and which are hard to change?

One big thing changing in online retail is the Temu and Shein business model. They partner deeply with overseas factories and use a tax loophole that allows them to bypass import duties. This makes products so much cheaper that many customers are willing to accept slower shipping. Amazon has been forced to re-think its customer value calculus, and they are launching a Temu competitor.

3. The Idea Maze. Finally, you have to understand what the cutting edge is inside your own company. Your colleagues are all navigating the idea maze with you, and once a company is sufficiently large you can get pretty far just by talking to them, reading their docs, and understanding which goals they are hitting and missing. 

If you were working on the team at Amazon launching the Temu competitor, it would be quite important to understand things like how much other teams have been able to negotiate prices with factories and how much price and shipping times impact conversion rates.

Insight generation

Once you have hypotheses, you have to go actually validate or invalidate them.

An important part of being good at this is simply being clear that this is what you’re doing. Start by stating what the hypotheses are and what conditions must be true to believe them. 

It’s then about having a toolkit that allows you to rapidly assess those conditions. There are many different types of businesses, roles, and problems, but if you’re at a tech company, 90% of the way you generate raw insights comes down to just two things: learn SQL, and learn to talk to customers.

The way to get to the data you want inside most tech companies is SQL. Sometimes people who have only spent time in fields like consulting think they can get by with spreadsheets because someone else will give them a nice neat dataset to work with, but they are wrong. Just invest in getting great at this right away so you can unblock yourself on working with data.

Talking to customers is an important counterbalance, because it can give you something data can’t: anecdotes. Until you’ve heard at least one customer talk about the dynamic you’re seeing in the data, you should be highly suspicious that that dynamic actually exists. Bezos has a line about this:

The thing I have noticed is when the anecdotes and the data disagree, the anecdotes are usually right. There's something wrong with the way you are measuring it.

To get value from customer conversations you must avoid leading the witness. You have a hypothesis, and if you’re not careful you can probably get the customer to agree with you even if they don’t. To counteract this, start your interviews with general topics and let the customer take you to more specific ones.

You’ll often have to vacillate between data and customer feedback as you zero in on an answer. At BCG I was taught a principle for how to do this: “qual → quant → qual.”

In other words, start with a few customer calls while you’re still in hypothesis generation mode to make sure you’re smoking out all of the possible issues. Then use data to size and validate the things you’re hearing. Then return to customers at the end to make sure what you found in the data is real. 

Solutioning

Now you must drive to recommendations. Coming up with good ideas is usually not the hard part. It’s knowing which of them are worth trying and in what order.

That doesn’t mean trying to predict exactly what will work, which is usually too difficult. It’s the more fundamental task of knowing which ones are feasible and how much effort they will be to implement, so you can assess whether it’s worth taking the shot at all.

This requires knowing how long things take, what is already on roadmaps, where there are dependencies on other things that must be built first, and a bunch of other things you can’t really know until you understand the teams that are doing the building, like engineering, operations and marketing. 

That is why strategic work can’t be decoupled from execution. The people working on strategy should either be the same ones that execute it, or embedded with them and working closely from the start of the process.

Synthesis

The most important skill of all is writing. That is because it is the best proxy for the thing that really matters: synthesis. Synthesis is taking all of the data and anecdotes and recommendations and telling a story that is cohesive and true.

Slides or bullet points are not as good at forcing synthesis, because you can list things that sound smart without actually figuring out if they’re important. Writing in prose forces structure and exposes gaps. Cultures that write (like Amazon) are not just making a stylistic choice; they actually learn and execute faster as a result.

Getting good at writing seems to be mostly about doing it a lot. There are certainly things that help: for example Barbara Minto created an excellent structure for good writing, and Paul Graham has published a lot of useful things on writing, including the appropriately named How to Write Usefully. But mostly you just have to write (and read good writing) a lot.

We’ve listed synthesis last, but it shouldn’t actually happen at the end. Start writing a skeleton of what you think the ultimate output will be at the very beginning. This has a way of exposing where there are holes in your logic.

The ultimate output should be something that lodges an understanding of the situation and the path forward in your reader’s brains in a few minutes. Not in thirty minutes, not after a follow up discussion. If you can make it understandable with just a few minutes of reading, your work can spread through your company and have much greater impact.

Are small and large decisions so different?

The sheer magnitude of smaller strategic decisions that require this process - going from hypothesis to insights to recommendations to synthesis - means that you can get a lot of reps at it. As you do, you’ll probably notice that you’re developing stronger and stronger opinions about the bigger strategic decisions as well.

That is because almost all decision making in business follows a pattern: one big decision (launching Prime) followed by an endless series of smaller decisions that will determine whether or not it actually works.

The only way to develop the instincts to be good at those big decisions is to understand deeply what it would take to make them work. So start there and the rest will follow.

Credits

Thank you to Max and Lenny for their feedback on a draft of this essay.

Share

There are five types of risk:

1. Technology risk - can we get the product to work?

2. Market risk - can we build something people love?

3. Scaling risk - can we acquire lots of customers?

4. Business model risk - can we serve customers profitably?

5. Defensibility risk - can we maintain market share?

Remove all of them, and a business will become one of the most successful in the world, worth billions of dollars. 

YC’s motto is “make something people want”, which is a great objective for early stage startups. The longer term goal is more of a mouthful:

VCs think about risk all the time. When a form of risk is removed, the expected value of a business goes up massively, so it’s lucrative to be the first money in after it’s clear that will happen.

But not enough potential startup employees think about risk, even though it is the #1 thing that determines their economic outcome.

Let’s look at each form of risk and the questions you can ask to understand if they’ve been solved. Then we’ll explore how this typically plays out over a startup’s lifecycle, and what the good entry points are for those considering joining.

Technology risk

Solving technology risk means answering the question “Does this product actually do the thing we want it to?”

Many people who have been in tech for less than ~20 years have never really had to think about this, because most companies were applying existing tech to new problems and customers. Basically every SaaS, marketplace, and e-commerce company jumps right to market risk.

But new waves of technology bring back tech risk. With the rise of AI, many people are joining startups where technology risk is not actually solved, perhaps without even realizing it. Despite the hype, it remains to be seen to what extent and in what situations a model can actually replace a customer support agent, a developer, or an attorney.

Market risk

Solving market risk is often called finding product-market fit, and it is the canonical thing that early stage founders are searching for.

Understanding if a business has hit product market fit is classically fuzzy and sometimes gets summarized as “you’ll know it when you see it”.

However there are two things that are very strong indicators, especially when taken together:

1. Do customers obviously love the product? Do they talk about how much they love it with their friends? Would they be very upset if it was taken away?

2. Does the product have strong retention? Specifically, if you draw a survival curve which charts the percentage of customers in each cohort that remain active over time, does it flatten out at some point? The fact that it flattens is more important than where it flattens - some subset of customers must be finding value and sticking for a long time.

Scaling risk

Product market fit often kicks off a wave of organic growth, but it won’t be sufficient to achieve real scale. The next phase is figuring out how to bring a product to many more people.

The thing to look for to assess whether scaling risk has been removed is, well, scale. But more precisely, two things:

1. Is there at least one channel of growth that is large and can grow reliably in the future? There aren’t many options - in this piece, Lenny and I wrote about how for consumer companies, there are basically only three (virality, paid marketing, and SEO). For B2B businesses you can add sales. One of these channels needs to be absolutely cranking and have a path to continued growth.

2. Is there a large enough market to continue to grow? This can be hard to assess because the best startups expand their markets and grow into new ones. But even once the business has produced meaningful scale (20, 50, 100 million in revenue) they should probably be less than 5% penetrated in their core market.

Business Model risk

Solving all of the previous forms of risk is considerably easier if you’re selling dollars for 90 cents. The next big risk is figuring out how to actually make money selling something people love.

The best two questions to assess the potential for good economics are:

1. Does the business have sufficiently low payback periods on customer acquisition spend? Good paybacks are usually about 6 months for B2C companies and 12-18 months for B2B.

2. Does the business have a healthy contribution margin? Contribution margin is the amount left after you take out all of costs that can be attributed to a transaction, like COGS and marketing. These vary widely but for technology businesses should generally be north of 50% of revenue.

Defensibility risk

Companies that solve for the other forms of risk will inevitably attract a lot of competition. Some businesses will hold up well to this, and others will get locked into a zero sum battle that destroys their margins. 

The difference is defensibility, and the best framework I’ve found for understanding this is Hamilton Helmer’s 7 Powers. It outlines the way that businesses can continue to earn differential returns, i.e. protect their position.

It’s worth just reading his book. However one shortcut: most defensibility for new technology companies derives from one of three things:

1. Network effects: does the product get more valuable as more customers use it?

2. Scale economies: does the cost to serve customers get significantly lower as the business gets larger?

3. Switching costs: once they have started using the product, is it hard for customers to switch to a similar one?

The journey

Solving for these risks is of course less linear than I’ve presented here. It’s not always clear if a form of risk has actually been removed, and companies can bounce between stages.

In particular, there is often a circular relationship between scaling and business model risk. Getting bigger requires pushing into new channels and new customer segments, both of which put more pressure on economics, so companies must vacillate between expanding, figuring out economics, and expanding again. 

A typical journey looks something like this:

Funding rounds map to these stages of risk only very roughly, and that is kind of the point: it’s more important to focus on risk than stage. Companies often get funded before risk has actually been removed. This is especially true in hype cycles like the one we’re going through with AI right now, which can create particularly bad outcomes for employees.

When should you join?

The most basic principle is to try to do the same thing VCs do: invest (join) as soon as you can tell a major risk will been removed, before many other people have figured it out. This is of course not particularly easy to do, but even considering risk gives job seekers a big advantage.

It’s common for candidates to flock to a startup that recently got a big new funding round and big new valuation. This is exactly the kind of thing to avoid, because the upside was just built into the value of your equity grant.

Beyond this basic principle, there are two entry points that are often good, and one that is often dangerous.

One very good time to join a startup is immediately after they have solved market risk (i.e. found product market fit). Most companies that achieve product market fit will not ultimately build a sustainable business. However most of them will go on to grow dramatically, which is great for your career and could produce a meaningful economic outcome. And you will almost certainly learn a lot in the process.

The other great time to join is during the messy dance between scaling and business model risk. These businesses are actively breaking out, and a higher percentage of them are on their way to an IPO or other meaningful exit.

It is an especially good time to join if the company has recently made a lot of progress on one of the forms of risk, such as tapping into a new growth channel or making improvements to their unit economics. One or both of these should be leading to rapidly expanding contribution profit.

A typically bad point to join a startup is before market risk is solved, if you’re not a founder. You’re simply taking on almost as much risk as the founders, with much less equity in return.

As the table below shows, employees #1-5 usually get around 50 bps of equity, which is at least an order of magnitude less than founders. Are the chances that the company gets to a meaningful outcome also an order of magnitude better? Occasionally, for example in the case of repeat founders who know their market exceptionally well. But not usually. 

It might still be a good trade for some people - working on early stage businesses is very fun if you’re wired for it, and it may be a mission you just have to be a part of. But in purely risk vs. reward terms, it’s usually better to wait until the company has found product market fit.

Learning to assess risk

The most important lesson of all: be intentional about learning to assess risk.

Early in a career, that often means working at at least one company that is sufficiently de-risked and sufficiently well run so that you can see what “good” looks like. This will help you get comfortable accepting more risk later on.

Later in a career, it means building enough expertise that you know how to spot breakthroughs in risk before the rest of the market can. That might mean specializing in a stage (such as helping companies scale after they achieve product market fit) or specializing in an industry (such as spending so much time with marketplaces that you know what it looks like as they transition through stages).

Ultimately, the assessment of risk is just one part of a broader framework for how to choose a startup. But if you start thinking a little more about risk and asking some of the questions above as you’re searching, you’re likely to make better decisions.

Credits

Thank you to Casey Winters, Chris Erickson, and Zach Grannis for their feedback on a draft of this essay.

When a business equation is translated into its analytical form (usually a spreadsheet), it is often called a “growth model”. They can be a powerful tool for understanding how a business works and where to allocate resources to inflect growth.

I’ve built growth models for many businesses and observed it done for many more. I’ve also noticed that it has become popular to post, blog, and podcast about it.

But here is the dirty secret: growth models usually don’t get used. It is often a useful exercise for the person building it, but they are rarely adopted broadly within a company.

In this essay, I outline the five most common pitfalls that cause growth modeling efforts fail and how to avoid them. I also include an example model to help bring these lessons to life.

Pitfall #1: Using a growth model to forecast your business

Here is a very common failure mode:

1. Build a fancy growth model

2. Try to use it to forecast revenue or user growth

3. The results prove to be inaccurate

4. Others at the company lose confidence in the model

Growth models don’t produce reliable forecasts for the simple reason that they rely on many assumptions that are each hard to predict. When you stack these assumptions on top of each other, it compounds this unpredictability and produces unreliable results as soon as you get 6-12 months into the future. 

Growth models are primarily useful for making relative tradeoff decisions. For example, if a growth team could focus on lifting global activation rates by 10%, or improving conversion to signup from paid marketing by 25%, or increasing referrals per customer by 5%, which of those would drive more revenue growth?

This raises a key distinction with a different kind of tool: an operating model. Operating models are typically owned by the finance team and produce a forecast that the company is goaled against and is reported to the board. To create the reliability needed for this purpose, they are based on fewer and better understood inputs. For example, in an operating model you typically would not break out the different steps in a conversion funnel, but in a growth model you usually would.

The people maintaining these models should absolutely learn from each other. But keeping the two models separate is an important element of success.

Pitfall #2: Not putting your growth model into a spreadsheet

Initially, simply laying out your equation in a doc or presentation can help the team start to understand the key metrics in the business and how they interact.

However, if you stop there, you will miss most of the value, for two reasons. First, before it is in a spreadsheet you can’t actually play with inputs to see what happens to outputs. How much is that 10% improvement in activation rate actually worth? The results are often nonintuitive.

But more importantly, in a spreadsheet there is nowhere to hide. You must actually figure out how all of the metrics link up to produce sensible outputs. You have to go track down all of the baseline rates for each assumption. In going through this process, you’ll inevitably realize you didn’t totally understand the equation or your business. Simply getting the model to work is a meaningful part of the value of the whole exercise.

Pitfall #3: Building “one model to rule them all” 

For relatively simple businesses, such as D2C or some SaaS models, you might be able to get away with one model that has all of the inputs you need. But many businesses are simply too complex to express in a single model, and you’ll end up with junk in, junk out.

Let’s take marketplaces, which is the business model I’m mostly deeply familiar with. For Airbnb, you have the demand side of the business, which may be simple enough to model: you acquire travelers, they retain at some rate, retained travelers make a certain number of bookings per year, and each of those bookings has an average price and margin structure. Great!

But what about the supply side? You’re also adding properties over time, and as you do it will increase traveler conversion rate and bookings per year. And as demand goes up, that will make it easier to acquire even more properties. I’ve seen many models that try to capture this flywheel dynamic between demand and supply, and they are almost always junk.

If you’re working on a business with this level of complexity, break your business into its component parts and model them separately. In the Airbnb example, at a minimum you will want distinct models for the supply and demand sides of the business. 

Pitfall #4: Choosing the wrong person to own the model

One of the hardest things about building a growth model is that it requires two skills that are rarely found in the same person. 

They must have excellent business judgment and an intuitive sense for the company’s strategy. But they also have to understand the business on a molecular level, and be able to get the right data on each individual input. If no one fits this profile, you can pair a few people to bridge the gap. But in my experience the closer you are to a single person owning the process end to end, the better.

To add a further constraint, it is usually better if this person is an “unbiased observer”. In other words, they don’t have an ingoing preference for what the model predicts you should work on. You will be shocked to find out that when someone who owns a particular growth channel or part of the product builds a growth model, they often find that their part of the business is very important!

Given these requirements, the person with the highest chance of success is usually one of the very strongest people on a team that is one step removed from operating, like analytics or strategic finance. 

Pitfall #5: Pitching the model for the model’s sake

Despite what LinkedIn and Twitter posts may indicate, no one cares that you built a fancy spreadsheet. The only thing they care about is growing the business.

So instead of telling people at your company about why growth models are great, or all of their different use cases, or how you’re going to build and maintain it, just pitch them on one thing they should do differently as a result of your interpretation of the model. That’s what is going to get attention. 

Growth model template

What does this look like in practice? Here is a very simple growth model for a transactional business (e.g. an e-commerce company or the demand side of a marketplace). There are three things to call out which reinforce the lessons above:

1. The first thing you see is a long list of assumptions

This is a good place to start, because the whole point of the model is to make relative trade-off decisions, and these inputs each represent something unique that you could invest in. Starting by laying them out will help you make sure you’re covering your option set. 

2. The core output is a comparison of scenarios

This is the basis for making those relative trade-off decisions. A good place to start is playing with one input (like conversion rate) or a small set of closely related inputs (like paid marketing spend growth and cost per click) to get a sense for how sensitive the model is to these changes. As you can see in this case, we’ve tweaked one assumption around referral rate, and it produces quite different outcomes, which become more and more different over time.

Just remember: the key output is the difference between the scenarios, not the absolute values they produce.

3. The model has very few “looping” assumptions

The model doesn’t include compounding assumptions which are typically very hard to capture accurately, e.g.:

• How changes in supply might impact demand (as in the Airbnb example above)

• How a growing customer base could increase awareness and accelerate the growth rate in Direct signups 

There are a few places where assumptions do explicitly compound, which are limited to those that are more reliable to model, such as how a larger customer base creates more and more viral referrals to the product:

Conclusion

There is a through-line in all of the advice above: don’t overplay your hand.

Growth models are a helpful tool alongside the other ways that you understand your business, like talking to customers, understanding the market, and bottoms-up goal setting. They aren’t a magic solution that invalidates your need for those other things.

All models are wrong, some are useful. Focus on making yours useful.

Credits

Thank you to David Weinstein, Lenny Rachitsky, and Zach Grannis for their contributions to this essay.

The first is probably what comes to mind when most people think of a “bad manager”: micromanagers who get into every detail and undermine the autonomy of their team.

However the inverse problem is at least as common, especially in startups: “macromanagers” who give too little direction to their team and undermine their performance. 

A large part of the journey to becoming a better manager is about correcting (and often over-correcting) until you find the optimal balance. The two paths look something like this:

Too often, managers wait until a crisis to course correct. Micromanagers are usually good at getting results in the short term, at the expense of eroding trust and losing engagement from the team over time. Crises often show up as people problems, like someone quitting without providing enough warning to do anything about it.

Macromanagers initially create happier teams, but they are likely to produce worse results. Crises often first manifest as output problems: the team fails at a project, or gets a reputation for not being able to ship. Over time, teams with macromanagers realize they are having less impact and not learning as much, which results in people problems as well.

Tuning in to leading indicators can help new managers diagnose the path they are on before it creates a crisis. Here are a few dimensions to consider, and what may manifest on either end:

Course correcting

Once you know the path you are on, here are some steps to accelerate course correction:

1. Name your fears

2. Know your team

3. Apply an intervention framework

4. Become a context sponge

The first step is naming your fears, because you can analyze the situation all you want, but until you get to the root of what is driving it, it’s too easy to fall back on old habits.

Micromanagers are often afraid they won’t be able to get their team to produce. This is especially true for previously standout ICs whose career reputation is staked on their ability to get things done. They’re nervous those talents won’t translate into management and just try to do things themselves.

Macromanagers are often afraid their team won’t like them or will rebel if they push too hard. Often the highest potential people are promoted early in their careers and as a result their team have similar levels of experience and may have previously been peers, leading to feelings of imposter syndrome.

Simply recognizing what kind of fear is most powerfully motivating your behavior helps a lot. Better yet, explore it with a manager, a mentor, or your team.

Another piece of foundational work is to deeply understand your team. That means knowing them well enough to understand their competence on any given type of project.

In probably the best management book of all time (High Output Management), Andy Grove coined the term task-relevant maturity:

"How often you monitor should not be based on what you believe your subordinate can do in general, but on his experience with a specific task and his prior performance with it - his task-relevant maturity.”

Something I’ve personally observed when doing this exercise: not only is a team member capable at a particular task, but they are often much more capable than I am. There is no clearer sign to back off.

Once you know yourself and your team, consider a framework for when to intervene in a team member’s decision.

The best one I’ve found comes from Keith Rabois, which Delian Asparouhov wrote a great piece on.

How high are the consequences of getting a decision wrong? How much conviction do you have in the right answer? The combination of these two dimensions is a useful recipe for what action to take:

The team needs room to fail, ideally on low stakes decisions. If you don’t have conviction, delegate fully. If you do, let the team roam and consider nudging them along the way.

High stakes decisions are harder. If you do have conviction and the team is headed in the wrong direction, override them. But if you don’t have conviction, you need to work with the team to gather more information until you do.

It is strictly better to be higher on the y-axis more often. Which leads us to our final point on how to find the optimal path: become a context sponge.

The most fundamental job of a manager is to have good judgement, and perhaps the most important ingredient of good judgement is having as much context as possible.

Part of this is about developing expertise in your functional area, which is why the best managers were often elite ICs in a given domain. 

But the context you need goes well beyond that. What actually makes this business tick? What are we really trying to accomplish? How are other teams trying to accomplish it? And most importantly, what is actually happening at the company, on the ground, right now?

Jensen Huang (CEO of Nvidia) is perhaps the greatest context sponge of all time. He has an approach he calls “stochastically sampling the system”. Instead of asking for status updates which would be highly filtered by the time they get to him, anyone in the company can email him a list of the “top five things" on their mind. He estimates that he reads 100 of these every day. Clearly this approach wouldn’t work for most people; the point is just how much he invests in building context.

It’s a long path

We are only scratching the surface of one of the hardest learning curves any manager faces. I know that I personally have a very long way to go.

However if you can rapidly identifying your natural tendency and what is driving it, and then begin to take steps to course correct before it creates problems, you’ll be well ahead of most people, perhaps on a new path entirely.

Credits

Thank you to Brian for feedback on this essay.

Step 1: Get your dream job. Doing product or growth or whatever it is that makes your heart flutter. Great comp, lots of equity.

Step 2: Waste years of your life because the company goes nowhere.

One reason this is so common is that people over-apply lessons learned from big companies where it’s safer to optimize for role and comp because they are more stable. But at a startup, what you do and how much you make can change rapidly and are almost entirely downstream of how well the company does.

I propose a “hierarchy of needs” when picking a startup to join, outlined in the diagram below. Start at the bottom, and only move up when a condition is met. Let’s take them each in turn.

Stage

When someone is considering joining both a startup and Google, it usually means they don’t really know what they want, and that probably what they want is not a startup.

The stage of business determines what game you are playing in the first place: what the job will be like, who will be successful in it, and the mix of risk and reward you are accepting. So start there.

The best framework I’ve found on this topic is from Nikhyl Singhal in Stage of company, not name of company.

He outlines four stages. I split his 3rd stage into two, because I think they can produce quite different experiences. You should be targeting at most two of these stages:

1. Pre-product fit (~Seed - Series A) These companies are figuring out what to build and if people want it at all. They need people who thrive in chaos, can apply a wide range of skills, and need very little hands-on management.

2. Immediately post product fit (~Series B) These companies are turning an initial spark into a real business. They often need people with some functional expertise (product, marketing, ops) to instill best practices for the first time.

3a. Early growth (Series C-D): These companies are experiencing explosive growth. Everything breaks every 6-12 months as a new stage of scale is achieved, so they need people who eat complexity and build process.

3b. Late growth (pre-IPO to immediately post-IPO): These companies are growing fast but starting to show signs of slowing. They need people who can instill discipline as they prepare to go public, but they also need people to help seed new growth bets like geographic expansion or moving upmarket.

4. Scale (FAANG, Microsoft, Salesforce, etc.): These companies are stable and more focused on defense than offense. They need people who thrive in structure and know how to get things done in a more bureaucratic environment.

There is no right or wrong answer. It’s mostly about the kind of environment you will thrive in (more or less guidance, higher or lower rate of change) and the level of risk you’re willing to take.

However a word of caution: if you find you are inclined toward the earlier part of this spectrum, watch out for the danger zone between the founding of the company and product market fit.

If you are not a founder but join a company without product market fit, you’re often taking on a similar level of risk as the founders, but with much less of the upside. As employee #1-5, you’ll likely get in the range of 50 bps of equity, at least an order of magnitude less than founders. That might be an OK trade for some people. But in my view it is often better to either be a founder or wait until the company has found product market fit.

Team

Aside from stage, everything is downstream of the quality of team: how successful the company will be, what your experience will be like working there, and the relationships you build that will serve you later in your career.

Here are three lenses through which to assess the team:

The founders 

In this essay, Paul Graham identifies five traits that YC looks for in founders: determination, flexibility, imagination, naughtiness and friendship. It’s about as good as you can do on this topic and it’s short, so I’d suggest just reading it.

Perhaps it goes without saying, but founders are usually pretty intense. It can be harder to relate to them in the same way that you relate to the rest of the team. Before taking that as a negative signal, consider that it might be a good thing: doing extreme things requires extreme traits. Elon Musk in his opening monologue on SNL:

“To anyone I've offended, I just want to say: I reinvented electric cars and I'm sending people to Mars in a rocket ship. Did you think I was also going to be a chill, normal dude?”

The team

Here I’d start with Warren Buffet’s quote on the traits he looks for when hiring: integrity, intelligence, and energy. They are all important, but as Buffet notes, the first is critical: 

“Somebody once said that in looking for people to hire, you look for three qualities: integrity, intelligence, and energy. And if you don't have the first, the other two will kill you. You think about it; it's true. If you hire somebody without integrity, you really want them to be dumb and lazy.”

It can’t just be a few people that stand out on these traits, it has to be most of the team. Talent density matters a lot because great people want to work with other great people, and if there aren’t enough of them, they’re ultimately going to leave.

The culture

The only advantage that startups have over incumbents is that they can move faster. As a result, the most important aspect of startup culture is that it enables velocity. This means a culture that is good at making decisions rapidly, and good at course-correcting when those decisions are wrong. It means a culture that has a low degree of politics and bureaucracy for its scale. And it means a culture that operates transparently, allowing everyone in the org to act autonomously.

A few questions to ask to assess culture:

• When was the last time two leaders disagreed? What did you do and why?

• Tell me about a bad decision you made. Why did it happen? How did you course correct?

• What decisions will you expect me to make? Who needs to review my decisions?

Metrics

If you’re fortunate enough to join a company that takes off, your equity will become valuable, so comp is mostly downstream of company performance. And the team will scale rapidly giving you bigger jobs and more optionality, so role is mostly downstream of performance as well.

While you’ll want to consider more than these, here are four metrics that collectively do a very good job of assessing the quality of a business:

1. Retention

Retention is by far the most important metric because it is both the best indicator of product market fit and the driver of all other metrics. Look at both user retention (how many customers are still active after x months) and revenue retention (how much revenue a cohort produces relative to its first year). You can find good benchmarks on both from Lenny Rachitsky here.

2. Growth rate (especially “natural” growth)

Here are more good benchmarks from Lenny on growth rate by stage:

While harder to benchmark, more important than aggregate growth rate is the “natural” growth rate, net of paid marketing and SEO.

Paid marketing and SEO can be important ways to grow, but they can also be influenced by marketing budget and content, and they are not entirely in a company’s control. A product manager at Google can wake up one morning and decide to run an experiment that tanks your traffic. What is left without these two channels is organic and viral growth, which tells a story of natural pull from the market: how fast are users finding and telling others about the product? 

3. Unit economics

This is particularly relevant for transactional businesses like marketplaces, but don't just assume SaaS businesses have great economics. In particular, focus on contribution margin, which answers the question “how much is left after you take out all of the direct costs of a transaction?” More on how to calculate contribution margin from CJ Gustafson here.

4. Payback period

Payback period (the amount of time it takes for an acquired customer to produce more contribution margin than the cost to acquire them) is the best measure of customer acquisition efficiency, because it measures how quickly marketing spend can be re-invested in more growth. Benchmarks from Lenny here. 

Role

Startup roles are much more fluid than big company roles. It’s unlikely you’ll do the same thing for very long, especially if the company is highly successful. 

So instead of focusing on a specific function or team, optimize for a role that will allow you to maximize your impact, and put up wins quickly to open up your options within the company. This comes down to basically three things:

What you are naturally good at. The main failure mode here is getting distracted by what you think you “should” be doing. Many people genuinely love and are great at product management, but many others flock to it because it’s the “cool” job to have. The more you can ignore this kind of thinking, the better off you will be.

Where you have expertise. Over time, the way to get real career leverage and produce outsized impact is to specialize (2). You could specialize by function (product vs. marketing vs. ops), or stage (product fit vs. growth), or type of business (marketplace vs. saas vs. hardware). Often the way to be exceptionally valuable is to combine multiple of these that are rarely combined.

What the company needs. Finally, map what you’re good at to what makes the company tick. This is mainly a product of:

• Company stage: refer back to that step.

• How the company grows: I’ve seen many traditional marketers get frustrated at companies for which growth is almost entirely product-led and the marketing team is starved of attention and resources. Or product managers who are frustrated at sales-led companies.

• How the company makes decisions: If you are a data person, join a company that is data-driven. Don’t join a company that doesn’t value data and try to convince them otherwise, because you will likely fail. Similarly, consider the difference between writing cultures and presentation cultures and where you’re more likely to be successful.

A final note: I’m talking only about role, not title. I would explicitly not optimize for title, and probably optimize AWAY from it early in your career. This is why:

Comp

Comp is important and anyone who tells you it isn’t is either kidding themselves or trying to sell you something. It’s important because money is useful and everyone wants more of it. It’s also important because it tells you a lot about the company you are considering joining. 

The mix of comp and equity tell you about the stage of the business and how to think about risk vs. reward. And the total size of the package tells you about how much they value you and often something about the quality of the business.

When I was considering my first startup role, I got a few offers from startups that were in the same ballpark, and an offer from a D2C business that was close to half of the others. The fact that this outlier offer was so low was a good sign that we were either misaligned on what I could contribute, or it simply wasn’t as good of a business as the true tech companies I was talking to (and probably both).

So of course it is worth paying attention to comp, and to maximize it when all else is equal. But if you stick with startups long enough, the majority of your comp will come from equity, and that is mostly determined by how well the company does and how well you do within it. That’s why comp is last on the list.

If you are highly intentional about choosing startups using a framework like this one, and give yourself time to take multiple swings, I believe your likelihood of picking one that goes on to grow dramatically is pretty high. Doing that will be an incredible accelerant for your career.

Credits

Thank you to Carla Pellicano, Jolie Zwick, and Lenny Rachitsky for their thoughtful feedback on drafts of this essay.

Notes

(1) Two additional thoughts on metrics:

1. The metrics discussed here mostly apply at the product market fit stage and beyond. Before that, you’re looking for the potential to achieve this kind of performance, which is harder.

2. Once you get an offer, a startup should be willing to share all of their numbers, so you should have pretty good data on which to evaluate these metrics. If a startup isn’t willing to go open book on their numbers, that is a red flag.

(2) My main advice on expertise is to not fall into the trap of remaining a generalist for too long. Here is something I wrote on this trap: Generalist Disease

Subscribe now

I once fell asleep in the lobby of Disney’s Animal Kingdom Lodge at 3:00am while editing a 200-page deck about how to get people to visit more theme parks.

But contrary to the bad rap that big consulting firms get, BCG was a great place to start my career. I met a lot of ambitious people and got a bootcamp in the basics of how to write, speak, and use data in a business context.

The problem is what often takes root in people (including myself) in these “elite” professions after a few years. Many develop an incorrect mental model for how to build a successful career. I call it Generalist Disease.

The hallmark symptom of people with Generalist Disease is optimizing for optionality. In order to open as many doors as possible in the future, they chase breadth (exposure to as many industries and functions as possible) and chase prestige (logos and titles that look good on resumes). 

The classic places to find breadth and prestige are fields like consulting, banking, and private equity. As tech has grown in prominence, the list now includes fields like product management and VC. Top MBA programs can be a good way to make sure Generalist Disease metastasizes. 

These jobs are very good at making you feel like you’re on the right path. Your parents are impressed, and you start to make more money. You may not be enjoying the work, but you can tell yourself a story about how you’re paying your dues to generate incredible options.

But after the 3rd or 4th impressive line on the resume, where are those really good options? They often fail to materialize, and people with Generalist Disease start to feel stuck. 

Goal seeking

I had a professor who said:

"The first 10 years of your career are for learning what you want to do. The next 10 are for getting great at it. The next 10 are for making an impact and making money.”

Of course it is never that linear, and some people do it faster. But there is a kernel of truth in it: generalism is not the goal. It is just a means of goal-seeking to find what it is you really want to do. Ultimately you must specialize to build real leverage, which requires the exact opposite of increasing optionality. It means being intentional about taking options off the table.

Unless you are one of the rare few who knows what they want to do at the start of your career, chasing breadth and prestige early on is actually a good idea because it accelerates the exploration part of the journey. Breadth lets you see many potential paths, and prestige helps open the initial door to paths you might want to explore.

But they both rapidly turn counterproductive, and people with Generalist Disease don’t seem to notice. The kind of breadth you get in fields like consulting is so surface level that it can only tell you where to hunt, not cross anything off the list entirely.

The value of prestige asymptotes after one or two door-opening lines on a resume. Then it turns into a device that employers use to convince insecure people to keep working at things they probably wouldn’t otherwise.

Clearly some people work at traditionally generalist careers like consulting for a long time and love it. But let’s be clear: they are no longer generalists. They are specialists in the job of consulting. If that is what you want to do, that’s great. But if not, it’s important to get off the ride when your learning switches primarily from exploration to primarily learning the job of consulting itself, which for many people is around 2-3 years. 

You ultimately need to switch to chasing aptitude and enjoyment. If you can find something you are naturally pretty good at and enjoy enough to spend a lot of time on, ultimately you will become very good at it and learn to enjoy it even more. That creates a compounding advantage.

Getting off the ride and chasing aptitude and enjoyment usually means trading a career path that goes steadily (but slowly) up and to the right, for one that is choppier at first so that you can uncover opportunities to rapidly bend the curve upward.

Hard pills to swallow

Why doesn’t everyone just take this advice? Because the only two remedies are difficult to stomach for someone who is carefully curating an elite image.

First, you must be willing to take steps that feel at best lateral and often down in order to explore widely enough to figure out what you really want to do.

The peak of my own generalist hubris was working in private equity after BCG because it was more prestigious and paid more. My colleagues were incredibly smart, and I think some of them genuinely loved it, but it just wasn’t my thing. It was hard to figure out what to do afterward because everything looked like a step down.

I ultimately decided to take my first startup role at Thumbtack. I went from private jets and a private office at the top of the Transamerica tower to making a desk out of boxes in a supply closet so that I had a quiet place to take calls. But I ended up learning that I loved startups, and I wouldn’t have had a shot at all of the things I got to do afterward if I hadn’t tried it.

Second, you have to grind. Once you narrow in on the right field, you have to find the parts of it that are the “craft”: the unsexy, manual, and repetitive tasks that you just can’t excel without.

Of course there is nothing wrong with choosing a field like product management or VC. I know many people for whom these jobs are the source of maximum impact and happiness. But the best product managers I know don’t float above the team acting like the CEO of the product. They spend an incredible amount of time talking with customers to understand what they need, and with the team figuring out how to help them move faster. And the best VCs I know are out personally pounding the pavement every day to find the best founders to work with.

My own role has “strategy” in the title, a word that people with Generalist Disease often flock to. But almost none of the job is sitting back and pontificating. I don’t know how to be good at it without spending countless hours swimming in the data and customer feedback looking for patterns.

The best of both worlds

It is initially unsettling to take a job you’re not sure is right, or to put your head down and refine the craft rather than continuing to explore options. But it will certainly make you happier to work hard at something you love and are good at. And ultimately most people that do also seem to get more of the conventional success they wanted in the first place.

Credits

Thank you to Tom, Gabriela and Forrest for their thoughtful feedback on this essay.

The services industries should be home to massive marketplaces. They are enormous markets: annual freelance labor spend is 1.3T and home improvement spend is 600B in the US alone [1,2]. They have highly fragmented buyers and sellers that would benefit greatly from a better way to find and transact with each other.

So after thousands of attempts by some of the smartest teams in the world, where are they? You might argue that Uber is a service marketplace. But other than that, none of the ~10 US public marketplaces with market caps over $10B are in services. None of the top 10 private marketplaces are either [3,4].

This essay explores what is holding services marketplaces back, how they might get unstuck, and why this would produce some of the largest businesses in the world.

What is holding services marketplaces back?

You can buy a lightbulb in one click on Amazon, but hiring an electrician is still about as hard as it was 100 years ago. You don’t really know what you need, how long it will take, or how much it should cost, and you have to just start talking to electricians to find out.

Marketplaces dealing in physical things like products and properties solved this problem by bringing an incredible amount of information online. Everything they sell has a “SKU” - a unique identifier attached to descriptions, photos, reviews, and prices. In other words, they make everything they sell legible to customers in real time.

This was not a trivial undertaking. Amazon built its own classification system for products called ASIN (Amazon Standard ID number) and has used over 1 trillion of them [5]. Doordash pulled every menu onto its platform, first by literally driving around and picking them up, and later with the help of restaurants. Airbnb famously provided photography services to help bring new units of supply online for the first time.

This is much harder to do with services, because they are so diverse. Each one of those electrical projects (or web design or tutoring or almost every other service) is one of a kind, resisting categorization.

As I outlined in my last essay, marketplaces have evolved in four stages, which increasingly make it easier for buyers and sellers to transact, and usually get much bigger in the process:

• Lead gen marketplaces help sellers list their good or services and help buyers discover them

• Checkout marketplaces provide prices, terms, and reviews upfront, allowing for real time checkout

• Managed marketplaces bear the risk of something going wrong with guarantees 

• Heavily managed marketplaces directly participate in distribution 

Legibility is the great filter between Lead Gen and everything else. Without it, you can’t provide enough information for customers to be comfortable purchasing in real time, so the best you can do is make introductions between buyers and sellers and let them take it from there.

How services marketplaces can get unstuck

Purchasing services requires a lot of steps today: searching for suppliers, scoping the project, negotiating the price, making the hire, completing the job, and sending final payment. 

Today, everything after the search happens offline, usually over multiple days or weeks. Ultimately, everything except for the job itself must happen online, in real time.

Evolving to a Checkout Marketplace requires enabling a customer to hire in real time by surfacing sufficient information about the project, the potential suppliers, and their quotes. Today, buyers get this information through a messy chain of messages, phone calls, and in person visits that is highly costly for both them and potential suppliers.

Many marketplaces are trying to streamline this process by collecting as much information from suppliers in advance about what they offer, how they price, and when they are available, and then collecting as much information from buyers in real time through a series of scoping questions. For example, here are a few of the intake questions to post a web design project on Upwork:

Usually, these flows essentially make the marketplace more effective at search, by filtering the potential set of suppliers that buyers can explore. Occasionally it starts to do some of the work of scoping and negotiating, by delivering a price range or the ability to schedule a meeting. But very rarely does it result in a “hire now” button that a customer will be comfortable clicking. 

This is because the messy back and forth between buyers and sellers is a feature, not a bug. In most cases, the customer doesn’t actually know what they want, and talking to service providers is part of the process to figure it out. And exploring possible options is a decision tree with so many branches that you can’t write all of the if/then logic in advance.

What is really good at exploring a messy decision space in back and forth dialogue? Humans. And increasingly, LLMs trained on humans doing a lot of that in the past.

The solution is going to look more like that original messy back and forth, but instead of talking to a bunch of service providers, the customer will chat with an AI trained by the marketplace. 

For that electrical project, you might start by taking a video of your backyard and the closest electrical panel with some voiceover about where you want to install outdoor lighting. Instead of all of the questions above about the web design project, you might just list a few example websites with the aesthetic you’re aiming for and jot down the functionality you need.

In both cases the AI will then understand what information is missing and start asking you for it, until it can translate it into a detailed scope. Using information from suppliers about which jobs they are interested in and how they price, the marketplace will then deliver supplier options and quotes in real time. The customer journey is starting to look much more manageable:

From there, it’s less of a leap to evolve to a Managed Marketplace. You just need to finish the job of putting SKUs on services, by providing the final price upfront. 

As marketplaces get more data on which suppliers are hired, they can also build better feedback mechanisms for ratings and reviews. And as payments flow through their rails, they will see what the customer ultimately paid, not just the initial quote.

Information on supplier quality and pricing will allow marketplaces to deliver final prices before the job is completed. If the ultimate price is higher, the marketplace will make the supplier whole and eat the cost. But if it is lower, the marketplace will pocket it. Customers are likely to be willing to pay a small premium for the assurance that their price is final, giving marketplaces some buffer to play with.

The customer journey has now evolved to its final form: 

How this will create some of the largest businesses in the world

If you dig into the metrics of most services marketplaces, you quickly see two things: a high percentage of traffic coming from paid channels and SEO, and low retention rates which limit customer LTV and constrain their ability to acquire more. 

That is because there is one step in the customer journey we haven’t talked about: before customers search on lead gen marketplaces, they usually conduct a Meta Search, using Google (and friends and family) to figure out their options. This means the marketplace has to try to re-acquire most transactions.

Every mega successful marketplace becomes the default place that customers begin their search, bypassing aggregators like Google. The first place that most of Amazon’s customers search for a product is Amazon. This transforms them from a customer acquisition arbitrage business into a self-perpetuating flywheel that eats market share for breakfast.

The only way to become the default is to deliver a better experience so reliably that it becomes clear it is not worth the time to start anywhere else.

A better experience mostly means selection, value and convenience. You can’t beat Google at selection, because they have an index of the whole internet, including you. You can’t beat them on value, because if anything your commission drives up a seller’s cost and ends up in the price the customer pays. So you must win on convenience, and you must win by a lot. 

Services marketplaces aren’t doing that today, but there are two ingredients that would enable it.

First, they must have enough training data to automate the experience as we described above. Once they can reliably deliver a “hire now” button, it will be blindingly obvious that this is better than Google. It’s the same thing that has happened in other industries: if you can order a meal in a couple clicks on Doordash, why would you start by searching for restaurants on Google?

Second, they must have buy-in from suppliers to help them follow through on a great experience. Moving to real-time accomplishes this as well.

To suppliers, Lead Gen marketplaces are roughly interchangeable with other channels sending them leads, including ads or referrals from past clients. Their incentive is to just use whatever is cheapest, and try to disintermediate and pull customers into their own workflow whenever they can.

But when a marketplace delivers transactions instead of leads, it is now a new workflow that doesn't require spending time and money to vet and negotiate with customers. As they receive more demand, suppliers are incentivized to improve their performance on the marketplace, by integrating it into their workflows, creating the most up to date services offerings and prices, and delivering the most responsive service. 

Instead of a leaky funnel, services marketplaces that evolve to real time transactions will have a flywheel powering a better and better customer experience to aggregate demand:

Rearranging an industry in your favor

If you are thinking that the services industries have other issues we haven’t addressed, you are correct.

There are two in particular. First, relationships in services are often monogamous: customers find a supplier they like and stick with them. Because the primary role of a marketplace is helping create new matches, this reduces their potential. Second, purchases are often infrequent, making it hard to create a habit and be top of mind when customers need you.

Both of these are real handicaps, but they are also mostly downstream of the legibility problem and will get much better when it is solved.

Monogamous behavior is mostly a consumer response to high transaction cost. Even if you don’t love the plumber who helped you last time, you’re still more likely to hire them the next time to avoid having to spend hours chasing down three new ones to see who is best. If a marketplace removes this cost, customers are going to start picking whoever is best for their current job, right now. 

Solving legibility would also grow the market by a lot, especially for smaller jobs. For services like small repairs, personal assistance, or quick freelancing jobs, the cost of finding help is often higher than the price for the service itself, and removing that cost will increase demand substantially. These markets look a lot like taxis before Uber.

Most importantly, it would enable marketplaces to cover a broader range of services. Today, the only way to create a highly convenient offering is to custom build an entire experience around a narrow vertical. This has created businesses that work in their niche, but struggle to continue to scale, like Care.com and Rover. The ability to deliver a great experience across hundreds of categories will finally create the high frequency that all of the largest marketplaces have.

Predictions

The ability to cover many categories, in a rapidly expanding market, with a flywheel that continually improves the customer experience will create monster marketplaces. Three thoughts on how this may play out:

1. Low trust, low complexity services will go first

There are two dimensions that make it harder for customers to press the “hire now” button, and they will determine the order in which industries tip.

The first is the complexity of the job. It’s simply a lot harder to scope and price an HVAC installation in real time than a house cleaning. Solving this is mostly about sufficient data and ability to parse it, and the progress on these dimensions is astonishing. As models go multi-modal, we’re likely to see a step-change improvement, because AI can parse a lot of information from videos and photos that the customer doesn’t really understand or know how to communicate. 

The second is the level of trust a customer needs to have in the specific person completing the job. This is a different kind of information problem, primarily requiring reliable reviews and background information.

It seems that high trust, high complexity services may be stuck for a while in a world where a marketplace can deliver initial meetings, but not hires. However, people now buy cars online without driving them, and Airbnb got people comfortable with strangers staying in their homes. So it’s not clear to me that there is some fundamental blocker on either of these dimensions.

I expect categories will tip in roughly this order, informed in part by the fact that we’re already starting to see wins in the bottom two quadrants:

2. The first big winners will evolve out of the current lead gen marketplaces

While they have low penetration of their massive industries, the leading services marketplaces are already enabling millions of transactions. Data from messages, quotes, profiles, and uploaded photo and video will give them a big headstart in automating the process.

Some marketplaces are already actively working toward this vision. For example, here is a quote from a Thumbtack fundraising announcement: “We built a self-serve system used by our pros to capture billions of individual preferences across hundreds of categories so that we could generate quotes programmatically with the same fidelity that pros could manually.” [6]

Thumbtack already allows customers to instantly book services for less complex projects like TV mounting, or book an appointment to get a quote for more complex ones like painting [7]. Knowing the ambition of that team, I expect there is much more coming.

Who might be well-positioned to do the same in freelancing, or education, or healthcare?

3. This will pave the way for vertical integration

As we’ve seen in industries like e-commerce and food delivery, the trend is to take on more and more of the work that suppliers do to improve the quality and reduce the variance in the customer experience and better aggregate demand.

For tasks in which customers care a lot about the creativity of an individual supplier like wedding photography or architecture, it will be harder to vertically integrate because customers are looking for the unique vision of individual suppliers. (I wrote more about this here).

But in cases where the customer simply wants the job done like cleaning and appliance repair, we’re likely to see vertically integrated models win over the long term. Early attempts at this such as Handy failed because they had essentially the same cost structure as offline service providers, making it difficult to scale at venture pace without burning a lot of money. But if you simultaneously wipe out a large percentage of back office costs by automating transactions and drive up customer retention through a better experience, this becomes more feasible.

And finally, fully digital tasks like web design and online tutoring are likely to be solved entirely by AI. If the model can grok the task well enough to fully scope and price it, it can probably also just execute it.

Credits

Thank you to Casey Winters, Chen Peng, Chris Erickson, Forrest Funnell, Marco Zappacosta, and Zach Grannis for their feedback on this essay.

Sources

[1] Freelance Forward, 2022 Report

[2] Improving America’s Housing

[3] Top public US marketplace businesses with market caps >$10B at publishing: Amazon, Walmart, Uber, Airbnb, Doordash, Ebay, Expedia, Zillow, Etsy

[4] A16Z Marketplace 100; top 10 private marketplace businesses: Instacart, Valve, Viagogo, SeatGeek, Turo, StockX, RockAuto, GOAT, Whatnot, Evolve. The first service marketplace on this list is StyleSeat, at #19.

[5] The Story Behind ASINs

[6] Blood, sweat, tears and fundraising

[7] Thumbtack Launches Instant Book to Make Hiring Pros Even Easier