The software company has discovered a bold new business model: agree to do something for a customer, do it, and charge for the result. Several agencies would like to welcome the industry to their field.
Sarah Tavel’s 2023 essay argued that AI companies could sell work rather than merely tools that help people perform it. We are revisiting that thesis, not presenting it as a new May 2026 announcement. Its appeal is straightforward: customers may care more about getting a job done than adding another application to their workflow.1
The offer moves responsibility across the table. A software buyer once had to operate the tool. A buyer of an outcome can ask why the promised result has not arrived.
The invoice has crossed a boundary
If you sell software for preparing a report, the customer may accept responsibility for reviewing and delivering it. If you sell the completed report, a missing section is more likely to be your problem. The customer did not purchase the opportunity to discover it.
Human review belongs in the cost of delivery. Hide it in “customer success” and the margin may look better until somebody asks what that team does all day.
The exception has entered the chat
Imagine a hypothetical business that prepares supplier-comparison reports for small manufacturers. Most requests follow a familiar pattern. Data is available, requirements are clear, and the system produces a useful result.
Then a customer asks about a supplier with incomplete records, an unusual certification, or a delivery requirement nobody specified at the start. The system needs more information. Someone must decide whether to research further, ask the customer, narrow the answer, or decline the work.
An exception is not necessarily evidence of a failed product. It can be an opportunity to learn. But if every exception requires expensive human effort, the business should count that effort rather than describing it as a temporary inconvenience forever.
The contract has to say which work is included, what counts as acceptance, and what happens when the customer disagrees. The model cannot negotiate those terms retrospectively.
A narrow promise can be a large advantage
Yet a narrow promise may be easier to trust and easier to improve. The customer understands what is included. The company can evaluate completion. Both can identify when the request falls outside the agreed scope.
This clarity can be especially valuable when the work affects decisions. An output that is correct in most ordinary cases may still require careful handling of the cases where it is uncertain. A delivery process that surfaces those boundaries can be more useful than one that presents every answer with identical confidence.
The agency is not the insult
Technology culture sometimes uses “services business” as if it were a diagnosis. That is too crude. Many customers need service, and many valuable businesses are built by delivering it well.
Revenue becomes more attractive when the company can deliver it repeatedly without discovering a new exception in every account.
A company might deliberately use people to deliver early work, learn the process, and automate more of it later. That can be sensible. It should describe the current operation accurately and have evidence for the proposed improvement. A future automation plan is not a present gross margin.
Likewise, an established agency using AI well may become more productive without becoming a venture-scale software platform. It can still be a good business. There is no requirement that every improvement qualify for the most flattering valuation category.
The customer buys the boring part
The real promise of selling work is relief. The customer does not want to manage another tool, coordinate another process, or inspect a mountain of plausible output. They want a particular responsibility handled.
That relief must be earned through delivery. The polished result matters, but so does what happens when the request is unclear, the data is incomplete, or the first version is wrong.
Just remember that the outcome includes the email asking why it has not arrived.