An AI agent costs in two ways, and the two behave quite differently. Building it is a one-off. Running it is a recurring cost that follows usage. Conflate them and the budget is wrong at both ends.
Cost one: building it
This resembles an ordinary development project, and is driven by three things.
How many sources it must read
One source with a usable API is manageable. Four sources, two of them legacy systems without an API and one a spreadsheet somebody maintains by hand, is another matter. As in every integration project, this is where the time goes.
How messy those sources are
An agent reading an out-of-date document library gives out-of-date answers. Tidying the sources is often the largest single item, and it has to be done by somebody who knows the data, which means you.
How much judgement the task demands
"Fetch every company matching these criteria" is simple. "Judge whether this company is worth contacting" requires you to agree first on what a good answer is. That conversation takes longer than the building does.
Cost two: running it
This is the part people do not budget for, and it is straightforward once measured. The cost per run follows how much the agent has to read and how much it has to produce. Multiplied by how often it runs, that gives a monthly figure.
An agent running once a day over a bounded set of data costs little. An agent running every fifteen minutes and reading the whole document archive each time costs something else entirely, for the same benefit.
Set the ceiling before going live
We measure what a single run actually costs during a test period, and set a ceiling per run and per month before anything goes into production. That makes the recurring cost predictable, and it turns a bug in a loop into an alert rather than an invoice.
The arithmetic that decides
The question is not whether the agent is cheap. It is whether it is cheaper than the work it replaces, counting the fact that a human still reviews the output at first.
Frequency is therefore decisive. A task done every day has an entirely different arithmetic from one done once a quarter, even if the agent is technically identical. Done rarely enough, it never pays, and then we say so.
The hidden items
- The approval step. Where a decision has consequences, a person reviews it. That is time, and it belongs in the arithmetic.
- Logging and follow-up. To answer why the agent did something, everything has to be logged per run. Cheap to build in, expensive to retrofit.
- Changes in the sources. When a system it reads from changes, the agent needs adjusting. That is operations, not project.
- Models change. What was the right choice last year is not necessarily right now. Budget for a review each year.
How we do it
We start with one task, one source, and measure. Only once we know what it actually costs to run do we widen it. Everything runs in your own Microsoft tenant, on your data, with access that inherits the permissions of the user the agent acts for.
An agent we built does exactly this every day at a maritime customer. Here is how it was set up, or ask for an estimate for your task.
Common questions
What does running an AI agent cost per month?
It follows how often it runs and how much it reads each time. We measure that during a test period and set a ceiling per run and per month before anything goes into production, so the recurring cost is known in advance rather than discovered on the invoice.
Is an AI agent cheaper than doing the work by hand?
That depends almost entirely on frequency. A task repeated every day has an entirely different arithmetic from one done once a quarter. If the task is rare enough, the agent does not pay, and we say so.
What is the biggest cost people forget?
Tidying the sources the agent will read. An agent reading out-of-date data gives out-of-date answers, and that tidying has to be done by somebody who knows the data.
Does our data leave the environment?
No. The agent runs in your own Microsoft tenant, on your data, and the data is not used to train models. Access follows the roles you already have in Entra ID.