What an AI agent costs, and why not per seat

Budget €200 to €3,000 a month recurring for enterprise use, and €12,000 to €80,000 of implementation if the deployment needs custom work. A packaged agent that plugs into tools you already run sits well below that second range.

But those figures answer the question badly, because they assume two offers can be compared. They cannot: the billing unit differs, and it is the unit that decides what you will actually pay.

Per-seat pricing is broken for an agent

This is the most important point in the article, and it fits in one sentence: the better the agent works, the fewer seats you need.

Conventional software is billed per seat because it makes a person faster: ten people equipped are worth ten times one. An agent does not make a person faster, it does the work. If your agent absorbs the follow-up workload of three business managers, you do not need three licences: you need it running.

A vendor billing per seat is therefore paid for the product to stay a tool rather than become a colleague. Their commercial interest is in you needing plenty of humans at screens, which is precisely the opposite of what you are buying.

The market worked this out before buyers did: pure per-seat pricing fell from 21% to 15% of software offerings in twelve months. Hybrid: a platform fee plus consumption, has become the de facto standard at roughly 41%. And 43% of buyers say they prefer consumption billing, with 27% preferring outcome-based.

The four models, and what each hides

Per seat. Predictable, easy to budget, and misaligned for the reason above. Acceptable when the agent is really an assistant: somebody stays at the screen and uses it.

Per consumption. You pay for what the agent executes, in credits, tokens or tasks. It is aligned, because the bill follows the work delivered. The risk is visibility: with no cap and no alert, runaway usage is discovered on an invoice.

Per outcome. The most aligned model and the hardest to define. Zendesk charges $1.50 per committed Automated Resolution and $2.00 pay-as-you-go: a resolution being a ticket closed by AI with no human involvement, confirmed after 72 hours without reactivation. It works because “ticket resolved” can be defined. In a staffing firm, “successful placement” depends on thirty factors the agent controls none of; nobody can honestly bill on that.

Hybrid. A base covering access and integration, plus consumption that follows usage. This is the compromise that dominates, for a good reason: it gives the buyer a predictable floor and the vendor an alignment.

What actually moves the bill

Token cost is real: $1–5 per million input tokens and $5–25 output for frontier models, a fraction of a dollar for light ones. For a thirty-consultant firm it is almost never the decisive line.

Three other factors weigh considerably more.

The volume of delegated tasks, obviously, but it grows far more slowly than people fear, because adoption takes three to four weeks and then moves in steps.

Context length. An agent that re-reads an entire mission history on every question consumes ten times more than an agent with a memory. That is an architecture question, not a usage one, and it is invisible in a demo.

Round trips per task. An agent that searches, comes up short, widens the criteria and searches again does the right work and costs three requests. That is the price of adaptability, and it is what separates it from a workflow.

What a credits model changes in practice

Credits are the most legible form of consumption billing, and the one most agents adopt. The principle: a monthly allowance included, a top-up when it runs out, and a counter showing what was consumed by what.

Its real advantage is not commercial but political. A shared allowance makes visible, inside the company, who delegates and how much, without buying a seat per person. A business manager who tries it twice costs nothing; the one who has moved half their work across shows up in the counter.

Three things are worth checking before committing. Is the allowance shared or per person? A shared allowance lets usage distribute itself naturally across the team. Do unused credits roll over from one month to the next? If not, month-end pushes people to spend so as not to lose, which is an absurd incentive. Is the rate card per task type published? A vendor who cannot say what a shortlist costs cannot let you forecast your bill.

What the demo never shows

A demo runs on clean data, a well-phrased question and a single round trip. Three cost lines are invisible in it.

Failed tasks. An agent searching badly because the brief was vague consumes as much as one that succeeds. Budget a margin for the first weeks: that is the price of learning, and it falls.

Data catching up. If your ATS is out of date, the agent will look elsewhere, cross-reference more sources and consume more for a worse result. That is the cost of bad data, billed in tokens.

What runs unasked. A daily digest, a watch on roll-offs, monitoring for replies: those are recurring tasks, so a base cost independent of activity. Ask for it separately: it is the part of the bill that will not fall in a quiet month.

The calculation nobody makes

Comparing the agent’s price against the cost of doing the same task by hand is the natural reasoning, and it is wrong, because most of these tasks are not done by hand. Most of the time they are not done at all. The follow-up that never went out costs nothing in salary, and it costs a placement. The roll-off spotted at three weeks instead of eight saves no hours: it avoids bench time on a consultant costing around €5,480 a month.

Hence the useful calculation, which fits in three lines. How many tasks of that kind come up per month, what share of them is not done today, and what a single recovered one is worth. A single avoided bench month per quarter comfortably covers a bill of a few hundred euros a month.

What to ask before signing

Four questions, and they take ten minutes.

What is the billing unit, and what happens if we double usage? The answer should be a number, not “we’ll look at it then”.

Is there a cap and an alert? A consumption model with no cap transfers the entire budget risk to the buyer.

Can we see consumption per task and per person? If the vendor cannot show you, they do not know what their own product costs to run either, and their price will move.

What do we pay in a month we do not use it? The answer tells you everything about the model. A reasonable base is legitimate; full price for a month with no work means you are paying for access, not for work.

The cost that is not on the invoice

One remains, and it is the one that kills projects: the time it takes before the agent is useful.

Budget 10–20% of the initial spend per year in maintenance, but above all a break-even between 4 and 9 months. That delay has nothing technical about it. It comes from learning to delegate, from data that needs tidying, from the approval line that has to be written, and it is exactly why four projects in ten are abandoned while still working.

An agent at €300 a month that nobody adopts costs infinitely more than one at €1,500 that everybody does. That is the only price comparison that really counts.