AI Startups Are Selling Labor, Not Software Seats

Generative AI changes what buyers are purchasing, according to Elad Gil. The older SaaS model sold seats, tools, and workflow systems. The new AI model sells “human labor equivalents”: work hours, labor hours, or “bits of cognition.”

That shift can reopen markets investors and founders had learned to avoid. Gil points to legal software as the example. The dogma had been that selling into law firms was a bad business: slow, difficult, and structurally unattractive. Harvey worked, he argues, because it was not merely another tool sold into a resistant buyer. It offered work product and augmented lawyers’ output.

The Harvey interface shown alongside the discussion reinforced the point: a prompt window connected to sources such as Vault, EDGAR, LexisNexis, and iManage, plus a workflow builder that could collect a supply agreement, ask whether the user represents the buyer or supplier, branch on that answer, and generate buyer- or supplier-favorable analysis. One visible prompt asked Harvey to research the strongest evidence for a client in a lawsuit and draft an email summarizing allegations and favorable evidence. The product was framed less as a passive database than as a system for producing legal work.

For Gil, Harvey is “a fundamentally different type of product from what people were selling before.” The market did not open simply because lawyers became more willing to buy software. AI made a different thing available to buy.

Elad Gil describes investing against dogma as hard because failed attempts often look definitive until they suddenly do not. A founder or investor may believe in a category, back or build a company in it, and watch it fail because the timing is wrong. Five years later, another company may win in what appears to be the same market. The practical question is what changed.

He lists several possible answers: technology can get good enough, regulation can shift, a market can change, an incumbent can stumble, or buyer willingness can move. The question is not just whether a team is excellent, but whether the market is open to innovation at that moment.

Gil frames this as a disagreement about entrepreneurial supply. One school of thought, which he attributes to Y Combinator, is that the world is founder-limited: more founders would mean more big companies. The alternate view is market-limited: only some markets are open at any given time, and those are the ones where very large companies can be built. If a market is not open to change, he argues, founders cannot build much there regardless of effort.

AI is unusual, in his view, because it has opened “tons and tons of markets that were closed for a long time.” Part of that comes from model capability. Part of it comes from buyer psychology: “every CEO is asking themselves, what’s my AI story?” Gil says he has never seen more openness to trying new things.

The implication is blunt. In the current AI market, he says, if an AI company is not growing explosively and quickly, “something’s fundamentally broken.” His reasoning is not that every AI idea is good. It is that buyer openness is so unusually high that strong products should be able to grow at rates that were previously rare. He cites OpenAI and Anthropic as examples, saying their ramps to tens of billions are the fastest ever to that scale.

Tim Ferriss asks how Gil distinguishes a good market from a great one, pointing to event-based investment logic: if a large company exits or declines to build in a space, that may create room for startups. Elad Gil’s answer begins with the venture question “why now?” The important issue is what has shifted to make an old ambition newly possible.

Regulation is one answer. Gil uses Samsara, the fleet-management company, as an example. He says Samsara benefited from regulation around in-cab driver monitoring, creating a clear wedge: cameras in trucks watching people so they do not fall asleep while driving on the road. From there, Samsara could build a broader software suite. The visuals shown with the example placed the Samsara brand next to a mobile interface with navigation and hours-of-service timers, then dashcam footage labeled “Driver yawning” and “Actual Footage,” with the driver’s face blurred.

Technology is another answer, and Gil argues that AI is especially powerful because foundation models plugged directly into existing commercial surfaces: enterprise data, information, email, white-collar work, and code. Language is used across enterprises and consumer products, so language models had an immediate commercial runway.

Robotics is different. Gil says robotics is interesting and will be important, but even with the world’s best robotic model, the submarkets with existing robotic hardware are relatively small. Unlike language AI, robotics does not automatically plug into nearly every enterprise workflow. The key issue is whether a new capability has an existing commercial surface to plug into.

Competitive shifts can also open markets. Gil gives the example of Infisical, a security company he is excited about, which competes in part with Hashi. Hashi’s acquisition by IBM, he says, creates startup opportunity because “anytime you get bought by IBM you slow down a lot usually.” A fast-growing underlying market can do the same: Coinbase benefited from crypto adoption and the proliferation of token types.

Across these examples, Gil is looking for a specific change that creates timing: regulation, technology, buyer behavior, incumbency, competitive disruption, or rapid market expansion.

Elad Gil treats market size as necessary but easy to fake. He defines TAM, after Ferriss prompts for the term, as total addressable market: the market a company is actually in. The danger is that founders often describe their market at a level so abstract that it stops being useful.

His example is a company claiming to facilitate global e-commerce. If global e-commerce were, in his hypothetical, $30 trillion a year, the founder might claim that capturing a tenth of a percent would produce $300 billion in revenue. Gil rejects that reasoning. If the company has built “this little optimization engine for SMB websites,” then global e-commerce is not its market.

At the same time, he argues that redefining a market can legitimately expand ambition. His example is Coca-Cola. Coke and Pepsi had long thought about market share in soda, where they were roughly neck and neck. Then, as Gil tells it, one Coke CEO reframed the market as share of liquids or drinks sold rather than share of soda. That changed Coke’s perceived position from roughly 50 percent share to 0.5 percent, and helped explain moves into products such as Dasani and other beverage categories.

The distinction is important. A fake TAM uses a huge adjacent economy to inflate the apparent scale of a narrow product. A useful redefinition changes the company’s strategic scope: Coke was not only in soda pop; it could understand itself as being in drinks.

Tim Ferriss asks whether there is an AI dogma that resembles the old payments belief that fraud would kill any company in the category — a commandment that may soon prove false. Elad Gil does not offer a long list. He mentions skepticism around the return on AI capital expenditure and whether the spending will ever be paid back, but only says he thinks that concern is “probably off.”

His stronger point is about posture. There are moments when it is smart to be contrarian, and moments when the consensus view is the smartest one. Right now, Gil says, AI is a consensus moment where the consensus is “very right.”

He warns against overcomplication. Investors or founders may try too hard to find the non-obvious angle — for example, deciding they should do hardware because it sounds more contrarian. Gil’s answer is simpler: “no, you should just buy more AI.” The unusually open market, the breadth of language-based commercial surfaces, and the speed of adoption make this a time when the obvious opportunity may also be the correct one.