The Economics of Autonomous AI Workers

A finance team can price almost anything, until someone asks them to put a number on a worker that has no seat, no salary, and no clock. The old columns don't fit — and the reason they don't fit is the whole story.
Somewhere right now a procurement analyst is staring at a spreadsheet with two columns that have served them faithfully for a decade, and neither one works. The first column is for software, and it prices things per seat: so many users, so many dollars a month, multiply and you are done. The second column is for people, and it prices things per salary: a role, a band, a fully-loaded cost with benefits and a ramp and a line item for the recruiter who found them. Every tool and every hire the company has ever bought slots cleanly into one column or the other. Now the analyst has been asked to price an autonomous AI worker that will handle a chunk of the accounts-receivable process, and it refuses to sit in either place. It is not a seat, because no human logs in behind it. It is not a salary, because it does not draw one whether or not there is work to do. The spreadsheet is not wrong about software or about people. It is simply missing the column for a third kind of thing, and the discomfort of that empty column is where most of the confusion about AI economics begins.
The instinct, faced with something new, is to reach for the nearest familiar frame and force a fit. So the market has spent two years pricing autonomous workers as if they were SaaS — a monthly fee per "agent," a tier, a seat with a robot in it — and reasoning about their value as if they were employees, asking how many people each one replaces and dividing the salary saved by the license paid. Both framings feel responsible. Both are quietly wrong, and the ways they are wrong compound, because a thing priced like a seat and justified like a salary ends up measured against the wrong baseline in both directions at once. To get the economics right you have to put down both columns and look at what an autonomous worker actually is: a cost that appears only when work appears, does the work, and disappears again.
A seat assumes a human behind it; a salary assumes a body in the chair
The per-seat model is not arbitrary. It encodes a real assumption that held for the entire history of business software, which is that behind every seat sits a person whose time is the actual scarce resource. You license a CRM per user because each license corresponds to a salesperson whose day the software is meant to make more productive; the seat is a proxy for the human it serves, and pricing it that way roughly tracks the value it creates. The moment you point that model at an autonomous worker, the proxy breaks, because there is no human behind the seat whose time the software is amplifying. The worker is not a tool a person uses. It is the thing doing the work, and charging for it by the seat is like charging for a delivery service by how many trucks are parked in the lot rather than how many packages arrive — a number that has stopped corresponding to anything the buyer cares about.
The salary frame breaks from the other side, and more subtly, because it looks so much more sophisticated. Pricing an autonomous worker against a human salary at least acknowledges that it does work rather than assists work, which is a step up from the seat. But a salary is fundamentally a payment for availability, not for output. You pay a person for roughly two thousand hours a year whether those hours are dense with high-judgment work or spent waiting for something to arrive in a queue, because a human comes as an indivisible unit — you cannot hire two-thirds of a collections specialist, and you cannot pay them only for the minutes they were actually resolving an account. Their cost is a fixed floor you commit to in advance and then hope to keep busy. An autonomous worker has no such floor. It does not need to be kept busy to justify itself, it does not degrade when the volume drops, and it does not sit idle at full cost between exceptions. Measuring it against a salary imports the one property of human labor — the fixed, presence-based cost of a whole person — that the new thing was specifically built not to have.
What it actually costs to run scales with the work, not the calendar
Strip away both frames and price the thing on its own terms, and a different shape appears. What an autonomous worker consumes to run is, at bottom, computation — the inference that the Reasoning Core burns to read an incoming message, gather the relevant context, decide what should happen, and act, plus the metered calls out through an LLM Gateway to whatever model does the reasoning. The defining property of that cost is that it is marginal. It is close to nothing when there is no work, and it rises in proportion to the work actually completed — not to the hours the worker was theoretically available, because availability is free when there is no body to feed and no clock to run. The cost curve of an autonomous worker looks nothing like a salary, which is a flat line you pay regardless of throughput, and nothing like a seat, which is a step function you pay per license. It looks like a utility bill: near-zero at rest, climbing with consumption, attached to volume rather than to time.
This is exactly the property that gets destroyed when the market prices autonomy as a fixed monthly fee per agent, because a flat fee re-imposes the salary's dead weight on a cost structure that was naturally variable. It is also, not coincidentally, where a great deal of the current disappointment comes from. Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value, along with what it calls "agent washing" — older tools relabeled as agents without the substance beneath. Part of what makes those costs escalate and that value stay unclear is a pricing model that charges by presence for something that only creates value by outcome, so the buyer pays a flat rate for a fleet of "agents" and then cannot connect the invoice to anything that got done. When the cost is variable but the price is fixed, the economics are illegible by construction, and illegible economics are how a project ends up on the cancellation list regardless of whether the technology worked.
The unit that means anything is the completed outcome
Once you see the cost as variable and attached to work done, the right unit of account stops being either the seat or the hour and becomes the completed outcome — the resolved ticket, the reconciled account, the screened application turned into a move-in, the AI Mission carried from the signal that triggered it to the human sign-off that closes it. This is the unit that everything else in the model wants to collapse toward. The cost to run attaches to it, because computation is spent per outcome rather than per hour. The value attaches to it, because what the business actually wanted was never a worker's presence but the outcome the worker produces. And the comparison the finance team should be running attaches to it too: not "how much cheaper is this agent than the salary it replaces," but "what did one completed outcome cost when a human carried the coordination around it, and what does it cost now that a Specialist Agent does, with a person kept in the loop only at the decisions that genuinely need one." That comparison is the honest one, and it is the only one that survives contact with a workforce that is no longer priced by the body.
Pricing by outcome also disciplines the technology in a way the other frames do not. A seat charges the same whether the agent resolves a case or escalates everything to a human; a salary pays out whether the day was productive or idle; but a cost that attaches to completed outcomes only accrues when an outcome is actually completed, which means the buyer stops paying for theater the moment the work stops finishing. This is why the outcome, and not the hour, is the unit that keeps everyone honest — the vendor cannot bill for an agent that does not complete work, and the buyer cannot fool themselves about value that never materialized. Platforms built on this premise, an Enterprise AI Platform running Autonomous AI Workers under Human-in-the-Loop gates, are legible precisely because the thing being counted is the same thing being paid for and the same thing being valued: work that got done.
The mental model worth carrying out of all this is that an autonomous worker is not a cheaper employee and not a more capable seat, but a new line item that neither existing column was built to hold — a variable, outcome-metered cost that behaves less like anything on the payroll or the software budget and more like electricity, billed for what it produces rather than for its presence. The companies that keep trying to force it into the seat column will overpay for idle capacity and underuse what they bought; the ones that force it into the salary column will measure it against the wrong baseline and miss the point of its cost structure entirely. The ones who get it right will do something the spreadsheet has never asked of them before: add a third column, price the outcome, and stop asking how many people the machine replaces — because the machine was never really priced in people to begin with, and the sooner the accounting admits that, the sooner the economics start telling the truth.
Discussion
No comments yet — start the conversation.