Service companies, run on AI

For two decades, the safe way to build a company on software was to sell the tool. Law firms bought it for their lawyers, accounting firms for their accountants, and the firms went on selling what they always had: the work, billed by the hour.

Some of the most influential investors in technology now argue that this order is turning round. Sequoia Capital and Y Combinator ask founders to build companies that sell the finished work, and other investors buy service firms to rebuild them around AI. Here is what they say, where the money goes and what is not yet proven.

Selling the work

Sequoia named the idea in October 2024. The cloud had turned software into a service, its partners Sonya Huang and Pat Grady wrote, and AI would turn services into software, for a market measured in trillions of dollars. Their example was Sierra, which answers customers for companies and is paid per resolved issue, not per seat.

In January 2026 the same partners declared 2026 the year of agents that can work on a task for hours, and wrote: “Remember all that talk of selling work? Now it's possible.”

In March, Sequoia partner Julien Bek made it a bet on services. The next company worth a trillion dollars, he argues, will be a software company that looks like a services firm. A company might spend 10'000 dollars a year on QuickBooks and 120'000 dollars on an accountant to close the books. The next great company will simply close the books. The reason is the budget: “For every dollar spent on software, six are spent on services.”

Copilots and autopilots

Bek draws one line through the market: “A copilot sells the tool. An autopilot sells the work.” Harvey sells to law firms, and the lawyer stays the customer and answers for the output. Crosby sells to the company that needs an NDA drafted, which buys the result directly.

Which of the two wins depends on the work. Bek separates intelligence, work that follows complex but learnable rules, from judgement, which rests on experience and taste. The more of a profession is intelligence, the sooner autopilots win, and he expects the border to move as AI systems learn what good judgement looks like.

His advice is to start where work is already outsourced. There the company already has a budget and buys a result, so an AI native provider is a change of supplier, not a reorganisation. His map includes insurance brokerage, accounting, tax, medical billing, legal work, IT services and recruitment. Management consulting, mostly judgement, has no clear candidate yet.

What Y Combinator asks for

Y Combinator publishes Requests for Startups, ideas it would like founders to take on. For the summer 2025 batch, Jared Friedman asked for full stack AI companies. If AI can do much of a law firm's work, he wrote, most founders sell an agent to law firms. The other way is to start your own firm, staff it with AI agents and compete. “Instead of selling to the dinosaurs, you could make them extinct.”

In spring 2026, Aaron Epstein turned to agencies, always hard to scale: low margins, slow manual work, and growth only by adding people. With AI, an agency can use the software itself and sell the finished product, such as legal documents written in minutes instead of weeks. The agencies of the future, he expects, will look more like software companies, with the margins of software.

For the summer 2026 batch, Gustaf Alströmer asked for AI native service companies. Services first became software, he wrote, then AI copilots. The next step is companies that sell the service itself: “Instead of giving you a tool, they just do the work.” His list of fields reads much like Sequoia's. The same list asks for a company brain, a living map of how a company works that AI can follow.

Buying the firms

Other investors buy the firms instead of waiting for new ones. General Catalyst backs founders who build AI for one service industry, then provides the capital to acquire and run businesses in it. It calls this the AI enabled roll-up, and its partners Marc Bhargava and Kate Bender sum it up in one line: “Customers care about outcomes, not products.” They put US service industries at more than 6 trillion dollars a year, against 370 billion for the entire software market, and say the firm has committed billions of dollars.

The deals have reached public markets. CNBC reported in June that General Catalyst and Trian agreed in December to take the asset manager Janus Henderson private for 7.6 billion dollars, and that Long Lake, backed by General Catalyst, agreed in May to do the same with American Express Global Business Travel for 6.3 billion. Forbes reported that Thrive Holdings, spun out of Thrive Capital, plans to put 1 billion dollars into local accounting firms and rewire them with AI.

Pricing follows. Ema, which runs teams of agents across HR, IT and finance and raised 77 million dollars in September, charges neither by seat nor by token but by completed tasks and business outcomes, its cofounder Surojit Chatterjee told TechCrunch.

What is not proven yet

Much of this is still a thesis, written by the people who invest in it. Forbes noted that the promise of AI roll-ups has so far outpaced reality. CNBC named two risks. Operating companies have historically returned 100 to 200 percent over a long hold, not the tenfold returns venture funds often promise. And unlike private equity, venture firms have not spent decades building teams to run the companies they buy.

Selling outcomes is also harder than it sounds. Ashu Garg and Jaya Gupta of Foundation Capital, who see 4.6 trillion dollars a year in salaries and outsourced services within reach of AI agents, warn that pricing by outcome fails when results vary too much between customers, because the result also depends on the customer's own business. They expect the shift to be messy.

And someone still answers for the work. In May 2025 the Solicitors Regulation Authority approved Garfield.Law, which helps businesses recover unpaid debts, as the first purely AI based law firm in England and Wales. The regulator noted that the system takes a step only when the client has approved it, and that named solicitors remain responsible for all its outputs and for anything that goes wrong. The machine may do the work. A person still signs for it.

Where crue stands

We made this move from the other side. We ran an agency before we built products, and stopped selling hours when the work no longer took the time we billed. Every crue product has a name, a result and a price, and nothing goes on sale that we have not run on crue first.

In Sequoia's terms, our products sit between copilot and autopilot. crue writes every offer and invoice through Kontor, and a person checks and sends each one. Tempo runs social media from strategy to posting, and a person approves every post. Twin, our company brain, holds how a company works as skills an assistant can follow.

That is human led, where every one of our products stands today. Agentic comes next: agents run whole routines, and people give the final yes. Autonomous comes after, and each stage is earned on our own work first. Helm, a vision we are shaping with our partner Eigna Labs, is where the line leads: a whole company run by one responsible founder with agents. It is not live.

The investors are right that companies want to buy the work, not the tool. The open question is who answers for it. Our answer stays the same: a person, with a name.