The service is the product

For twenty years, software companies sold tools and service firms sold hours. The buyer paid for a licence, then paid again for the people who ran it. AI agents break that split. When software can read an invoice, match a bank line or review a contract clause, the thing worth buying is no longer the tool. It is the finished work.

Y Combinator said so plainly in its Summer 2026 request for startups, which asks for AI-native service companies: firms that sell the completed work rather than software for someone else to operate. Sequoia makes the same argument from the investor side in its thesis, Services: The New Software. The margin that used to sit in labour can now sit with whoever delivers the outcome with agents.

This is not a small change in packaging. It changes who carries the operating burden. A licence leaves the buyer to staff, train and supervise the tool. An outcome leaves the buyer to approve the result. For a finance team of five, that difference is most of the working week.

Forbes reported in April 2026 that agencies are selling outcomes into labour budgets rather than software budgets. That matters more than it sounds. A software budget is small and argued over line by line. A labour budget is large, and a buyer judges it on one question: did the work get done?

The funding follows the same line. Crosby raised money to review contracts at a fixed price per document. Distyl built its business on outcome pricing. Neither sells a seat. Both sell a result with a number attached, and both carry the cost of producing it.

The pattern is not limited to startups. Established agencies and consultancies are rewriting proposals around deliverables and prices per unit of work, because their clients have started asking why an hour of preparation should cost the same when an agent does most of it.

A company with a controller and a small finance team does not want another tool. It wants the close done on time, with the matching finished and the exceptions explained. Buying software means someone on the team must learn it, configure it and keep it running. Buying the close means a provider carries that work and is paid when it is done.

This changes how to judge a vendor. Ask what they will deliver, by when, and how you will measure it. Ask who approves the work before it posts. Ask what happens to your data when you leave. A vendor selling outcomes can answer all three in writing. A vendor selling a platform usually answers with a demo.

It also changes where the risk sits. With a tool, the risk of it not working sits with you: you paid for the licence either way. With a fixed-price outcome, the provider carries the risk of a slow build and a hard dataset. That is why outcome pricing needs evals. Without a measured result, a fixed price is only a promise.

Outcome pricing is easy to say and hard to deliver. An agent that matches most bank lines still leaves exceptions, and someone has to clear them. A provider that hides exceptions to hit a number is worse than a tool, because the error arrives looking finished.

So the finished work has to include the approval step. The controller still signs off. The audit log still records what the agent did and who approved it. Selling the outcome does not remove the person. It removes the hours of preparation in front of them, and it makes the remaining judgement easier to see.

There is a second failure: pricing an outcome nobody defined. "Better reporting" is not an outcome. A pack delivered by the fifth working day, with commentary accepted without rewrite, is. The audit exists to turn the first kind of sentence into the second before anyone signs.

A buyer can tell within one call whether a vendor sells outcomes or tools. Ask them to describe the finished work in one sentence, including when it is delivered and how it is checked. Ask what happens when the agent is wrong, and who notices first. Ask what the price covers if your data turns out messier than expected, and what it does not cover.

Then ask about the end of the relationship. An outcome-priced vendor should be able to say what you keep if you leave: the runbook, the code, the configuration, and written confirmation that your data has been deleted. A vendor that goes quiet on exit terms is usually counting on you not leaving, which is the old software model with a new label.

We run our own studio on agents first, and we hold our own work to the pattern we sell: an agent does the preparation, a person approves, a log records it, and an eval tells us whether it is getting better. If a pattern does not survive inside our studio, we do not sell it to a client.

Client work is priced the same way. The audit fixes the scope and the price. The pilot delivers one workflow live on your data, measured on a period the agent never saw. The run retainer keeps it working and queues the next workflow. At no point does the client pay for hours, and at no point does the client need to learn a new platform.

The service is the product. For a buyer, the practical test is simple. If a vendor cannot tell you the outcome, the price and the approval step before you sign, you are still buying a tool, however the proposal is worded.