Software sold by the seat assumes someone is sitting in it

Per-seat pricing encodes a belief about how work happens: that value scales with the number of people logged in. It was a reasonable belief for twenty years. It is becoming an odd one.

In August 2025, Gartner predicted that 40% of enterprise applications would be integrated with task-specific AI agents by the end of 2026, up from less than 5% at the time of writing. In the same release, its best-case projection had agentic AI driving roughly 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from 2% in 2025.

Forecasts are forecasts. But the direction has a consequence that does not depend on the exact figure: if the thing doing the work in an application is not a person, then a price per person is measuring the wrong noun.

What a seat was actually for

A seat was a proxy. Nobody ever wanted to buy access for a named human; they wanted the output that human produced with the tool, and headcount was the most convenient thing to count. It worked because the relationship held. More people meant more work done, so more seats meant more value, and everyone could live with the approximation.

Agents break the proxy in a specific way. The work no longer scales with the number of people who have logged in, so a company can double its output without adding a seat, or add fifty seats and change nothing. Once that link is broken, per-seat pricing stops describing value and starts describing a headcount that happens to sit near the software.

The category question, not the pricing question

The pricing debate is a symptom. The real question is what kind of thing you are buying. Software is a tool: it waits, it presents options, it does what it is told, and someone competent has to be available to drive it. That is why buying more of it does not reliably produce more output. You also need the hours, and the hours are the scarce thing.

A coworker is a different purchase. You are not buying a surface to operate; you are buying a job that gets done, with a name attached, with an explicit boundary around what happens without asking you, and with a report at the end. The unit of value is the finished work, which means it can be priced as the finished work.

Nobody has ever wanted software. They have wanted the outcome the software was standing in front of.

What changes for the buyer

The most useful shift is what the evaluation looks like. A tool is assessed on features, because you are estimating whether a capable operator could get what they need out of it. A coworker is assessed on completed work, because that is the deliverable, and the demonstration either produces some or it does not.

• —Ask what it finishes without you, not what it can do with you. The second question is about the tool; the first is about the outcome.
• —Ask what comes back for a decision, and how often. A system that returns everything has moved the work rather than done it.
• —Ask what it writes back into, and where the record lives. Work that stays inside the vendor is a demonstration, not an operation.
• —Price against the job, not the login. If the number only makes sense when you count people, you are still buying a tool.

We are not neutral here, obviously. Evvolv sells named coworkers rather than seats, which is a position with an interest attached. But the argument stands on its own arithmetic: once output stops tracking headcount, a price that tracks headcount is measuring something that no longer matters.

The next generation of enterprise software probably will not look like better software. It will look like staff.