Sell the work, not the software: what the shift means
By Frontmark. Updated . 7 minute read.
Sell the work, not the software means an AI company delivers the finished outcome, such as a filed tax return or a processed claim, and charges for that result instead of licensing a tool to the people who do the job. The company runs the work itself with AI and takes responsibility for the quality.
By late 2026 this is the most repeated investor thesis in AI. It deserves a plain reading, because the argument is stronger than its slogans and its open questions are larger than its advocates usually say. What follows is that reading as of 16 November 2026.
What does it mean to sell the work?
A software company sells a tool. The customer still employs people to use it, and the customer still owns the result, good or bad. A company that sells the work takes the job itself. The customer buys a completed bookkeeping close, a placed insurance policy or a reviewed contract.
Three things change when a company makes that move.
The price is set against the cost of labor. The buyer is comparing the service with a salary or an outsourcing contract, and no longer with another software license.
The provider carries the risk of getting it wrong. A tool vendor can point to the user. A firm that delivered the work cannot.
The provider decides how the work is done. How much is handled by AI and how much by people is an internal matter. The customer does not have to adopt anything, retrain anyone or change a process. They receive the outcome the way they would from any outside firm.
Where did the idea come from?
It built up over several years, and no single firm owns it.
Sarah Tavel’s essay Sell the Work argued that AI startups should sell the work product itself, with EvenUp as its example. Emergence Capital published The AI-Native Services Playbook by Jake Saper and Rishub Nahar on 31 January 2025, advising founders to pair a domain expert with an applied-AI expert.
The statement that set the tone for 2026 came from Sequoia. In Services: The New Software, published 5 March 2026, Julien Bek argues that the next $1 trillion company will look like a services firm built on software. He says six dollars go to services for every dollar spent on software, which in his telling is the size of the opening.
Y Combinator then turned the argument into requests. In its Spring 2026 Requests for Startups, archived 3 February 2026, Aaron Epstein asked for AI-native agencies and said agencies can now use the software themselves, sell the finished product and reach software-like margins. In the Summer 2026 list, archived 18 May 2026, Gustaf Alströmer asked for AI-native service companies that do the work instead of selling software. He named insurance brokerage, accounting, tax and audit, compliance, and healthcare administration.
By November 2026, then, describing a company that delivers outcomes is standard language. It is no longer a novel claim, and nobody writing about it now should present it as a discovery.
What is the difference between a copilot and an autopilot?
The distinction is Sequoia’s. A copilot is a tool sold to the professional who does the job. The accountant, lawyer or broker works faster, and remains the one responsible. An autopilot sells the work to the customer who needs it done.
The two compete for different money. A copilot is paid for out of a software budget and is judged against other tools. An autopilot is paid for out of a labor or outsourcing budget and is judged against the cost and quality of people.
Bek’s advice on where to begin is specific. Start with work that is already outsourced and that he calls intelligence-heavy, then move toward work that is judgment-heavy. Intelligence-heavy work follows rules, however complicated the rules are. Judgment-heavy work depends on experience and on weighing things that are not written down.
Starting with outsourced work has a practical logic. The buyer already has a budget line for it and already accepts that an outside party does the job. Nobody inside the customer has to be replaced for the sale to happen.
This is related to the difference between designing a company around AI and adding AI to an existing one, covered in AI-native vs AI-enabled.
What is an AI-native services company?
An AI-native services company does what a service firm does, with AI carrying out most of the tasks from its first day. It has customers and deliverables like an agency or a practice. Its cost structure is meant to look different, because the volume of work does not rise in step with headcount.
The phrase full-stack AI company describes the same thing from another side. A full-stack company does the whole job for the customer instead of supplying one tool in someone else’s chain.
Y Combinator partners gave a sense of how common this has become. On the Lightcone podcast episode of 18 September 2026, they said the share of accepted companies doing full-stack, end-to-end work rose from about 10 percent to over 25 percent of the batch. That is the partners’ own description of their intake, and it is the clearest signal available of where new founders are heading.
For the underlying definition, see what is an AI-native company.
Why are investors making this argument now?
Three things lined up.
The first is the size of the budget. If Sequoia’s ratio is right, services are a far larger pool of spending than software, and until recently software companies could not reach it.
The second is capability. A summary by NYU Shanghai, dated 17 April 2026, of Anthropic’s 2026 Agentic Coding Trends Report describes single agents becoming coordinated teams, with tasks that run for hours to days. Work of that length starts to look like a job and less like a feature.
The third is capital. Venture firms are funding the acquisition of existing service firms as a faster route to the same end. That model is compared with building new companies in AI roll-up vs venture studio vs AI holding company.
What are the open questions?
The thesis is persuasive. Several parts of it are not yet shown.
Does the AI do the work? Emergence Capital warned in its playbook of “Mirage PMF”: revenue growth that hides a lack of real automation. A company can sell outcomes and deliver them with people behind the curtain. It will grow, and it will have the economics of an ordinary services firm.
Does the work get finished? A company that sells outcomes is paid for completed jobs. The research section of the State of AI Report 2026, published by Air Street Capital in October 2026, reports that on OSWorld 2.0 the best model earned 77.7 percent partial credit and fully completed 44.3 percent of computer workflows. The distance between those two numbers is the part a service business has to cover with people, checks or rework.
Are these venture-scale companies? Euclid Ventures, writing on 16 April 2026, is cautious on scope, market size and venture fit. It separates services used as a way into a market from businesses that are services companies throughout.
Will the margins hold? The margin figures that appear in these essays are targets and expectations. The sources reviewed for this article do not contain audited results that confirm them.
Who is accountable? Gartner predicted on 2 October 2026 that 40 percent of agentic AI projects will be cancelled by the end of 2027, and said human accountability cannot be removed. A company that sells the work is the accountable party by construction. It needs named people who answer for the result.
Who knows the domain? In Y Combinator’s Summer 2026 list, Tom Blomfield’s Company Brain request says the blocker to automating companies is now domain knowledge and no longer the models. Selling the work in insurance or audit requires knowing how that work is judged by the people who buy it.
What does this mean for someone starting a company?
A few practical points follow from the sources.
Choose work with a clear definition of done. If neither you nor the customer can say when the job is finished and correct, an agent cannot be held to it either.
Measure the share of the work that no longer needs a person, and measure it honestly. That number, more than revenue, says whether the company is what it claims to be.
Keep people on the decisions that need judgment. The sources that argue hardest for selling the work also describe a gradual move from rule-following tasks toward judgment. They do not describe removing people on day one.
Decide who owns the company that does the work. The essays mostly leave this open. A services business built on AI can be founder-owned and venture-funded, bought by a roll-up, or built and kept by a long-term owner. Each produces a different company over ten years.
Where Frontmark stands
Frontmark is an AI-native company-building institution. It builds new companies and keeps long-term ownership of them. The idea that a company can do the work for its customer is common ground by now, so Frontmark’s position is about how the work gets done and who owns the result. Agents do the volume of the work, human operators make the decisions that need judgment and stay accountable, and a collective of founders and operators supplies access and domain context. The model, Research. Form. Build. Back. Compound., is described on the approach page.
Common questions.
What does sell the work, not the software mean?
It means an AI company delivers a completed job and charges for the result, instead of selling a software license to people who then do the job themselves. The buyer gets a finished outcome, such as closed books or a placed insurance policy, and the provider decides how much of the work is done by AI.
What is the difference between a copilot and an autopilot in AI?
The terms come from a Sequoia essay published 5 March 2026. A copilot is a tool sold to a professional, who remains responsible for the job. An autopilot sells the completed work to the customer who needs it done. The copilot competes for software budgets, while the autopilot competes for labor and outsourcing budgets.
What is an AI-native services company?
An AI-native services company does the work of a service firm, such as accounting, compliance or insurance brokerage, with AI carrying out most of the tasks from the first day. It sells the outcome to the customer. Y Combinator's Summer 2026 Requests for Startups asked for companies of this kind.
What is a full-stack AI company?
A full-stack AI company does the whole job for the customer instead of supplying one tool in the chain. Y Combinator partners said on the Lightcone podcast on 18 September 2026 that the share of accepted companies doing full-stack, end-to-end work rose from about 10 percent to over 25 percent of the batch.
What are the risks of an AI-native services business?
The main risk is growth that hides a lack of automation, which Emergence Capital named in January 2025. If people are still doing most of the work, the business has the economics of an ordinary services firm. Other risks are unreliable task completion, uncertain margins and the need for real credibility in the customer's industry.