What an AI agent costs in a company

Three lines, and the most visible is the smallest

When a company asks what an agent costs, it thinks first of the engine, meaning what is billed each time the agent handles a case. That is the line most discussed, and by far the smallest.

An agent budget reads in three lines.

• Building: framing, tuning, connections, go-live. A one-off cost, sitting on average between 3,000 and 15,000 euros, like any custom tool. That average has no limit at either end.
• Usage: what the agent consumes each time it handles a case. A recurring cost, proportional to volume.
• Supervision and upkeep: knowing what the agent handled, what it passed on and why, and adjusting the rules. A recurring, fixed cost, in the region of 10 to 20% of the build cost a year.

Very low offers almost always cover the first line alone, with an agent running while nobody knows what it produces. Very high offers often include a process redesign nobody asked for.

So the right way to compare two proposals is to bring them to the same format: build cost, monthly usage cost at the expected volume, monthly supervision cost.

Usage cost is not counted in words

You often read that an agent's cost is calculated from the volume of text read and produced. That is part of the answer, and misleading if you stop there.

An agent does not merely read and write. It checks stock, queries a history, creates a record, sends a message. Each of those actions is an extra round trip, and several come with another pass of reading. An agent doing three things therefore costs markedly more than one doing a single thing, at identical text volume.

The only unit that makes sense is therefore the case handled, end to end, actions included.

At that scale the orders of magnitude are reassuring. On sorting messages or extracting fields from a document, you stay in cents per case. The monthly usage cost of an agent handling several hundred cases a day is counted in tens of euros, not thousands.

That is a question to put to your supplier exactly as it stands: what is the full cost of one case handled, actions included, and how was it optimised?

What moves the bill by a factor of ten

Five things shift that unit cost considerably, and those are what to check on your own case.

• Document length. A contract of several dozen pages re-read at every step weighs far more than a ten-line message.
• The number of actions per case. An agent that checks three systems and re-reads itself before acting does five times the work of one that handles a case in a single pass.
• The nature of the content. An image or a scanned document to interpret costs appreciably more than equivalent text.
• The rework rate. Every case reworked or retried is paid for twice.
• The accuracy required. Raising the bar implies extra verification, and therefore extra passes.

Conversely, three design decisions commonly reduce that cost by a factor of five: passing only the information relevant to the decision, reading reference documents once rather than at every step, and reserving the heaviest handling for the cases that warrant it.

That is why "what does an agent cost" has no answer out of context. It has a very precise one as soon as you know the volume, the length of the content and the number of actions expected.

Knowing whether it is worth it, before quoting

The payback calculation happens before you ask for a quote, and it needs only two figures: the time the task takes today, measured rather than estimated, and its real frequency over a representative month.

Multiply the two. If the result is counted in working days per month, an agent earns its place. If it is counted in hours, look elsewhere: a better-designed internal tool, or a classic automation, will cost less and be simpler to supervise.

Stay conservative about the share the agent will take on. A reasonable first-year assumption sits well below the whole flow, because some cases will keep going to a human, and that is by design.

Two benefits do not enter that calculation and yet explain why deployed agents stay in service: shorter delays, which your customers notice immediately, and consistency of handling, which reduces rework. They should not justify a project, but they count.

The concrete steps of a deployment: how to deploy an AI agent in a company

The costs quotes leave out

Four lines are regularly missing from proposals, and each catches up later, for more.

• Supervision. Without it, an agent that stops on a Thursday evening is discovered through a customer call on Monday.
• Upkeep of the material. Rules and examples are alive: regular review of the set-aside cases maintains quality, its absence degrades it slowly.
• Business changes. A new range, a new kind of request, a changed procedure: the agent has to follow.
• Handover. The rules, the examples and the configuration must be yours if you change supplier.

For all of these, count an annual budget in the region of 10 to 20% of the build cost. That is the same order of magnitude as maintaining custom software, and for the same reasons.

Five questions before signing

Five questions are enough to judge a proposal, and they are asked before signing.

• What is the full cost of one case handled, actions included, and how was it optimised?
• What exactly does supervision cover, and how will I be told about a stoppage?
• What level of reliability is targeted, and on which cases was it measured?
• Who owns the rules, the examples and the configuration?
• What happens in year two, and at what cost?

A supplier who answers those five precisely has already demonstrated most of what matters. And if you want an order of magnitude for your own case, two pieces of information are enough to start: the monthly volume of the task and the time it takes today.