What it costs to run an agent

Agent business cases are usually built on a saving against a build cost. The build cost is the easy part. Six things cost money every month after go-live, and together they often exceed it within the first year.

1. Consumption

Model usage is billed by volume. The agent’s cost therefore rises as adoption rises, which is the outcome the investment was meant to produce.

Forecast from expected transaction volume rather than from a licence count, and set a ceiling with an alert. The dangerous period is the first month, when usage patterns form and nobody is looking.

2. Keeping the source content current

An agent grounded in your documents is only as good as those documents. Policies change, procedures are revised, prices move. Somebody has to notice and update the source set.

Where nobody owns this, the agent does not fail visibly — the answers just drift out of date and people stop trusting it. Budget the time explicitly.

3. Evaluation

A set of real questions with known-good answers, run before each change, with results recorded. Without it there is no way to tell whether a change improved things, and every complaint becomes an argument.

This is a small ongoing cost that prevents a large one.

4. Exception handling

The work the agent hands back. This is a genuine workload and it is frequently counted as a saving rather than a cost.

The honest arithmetic: the volume the agent clears, times the time it saves, minus the time spent on exceptions, minus the time spent checking the agent’s work in the early period when trust is being established.

5. Support

People will report wrong answers, ask why something was not processed, and need access changed. Someone has to triage that and decide whether it is a content problem, a permissions problem or a genuine defect.

This usually lands on whoever built it, which works until they move on.

6. Change

Source systems get upgraded. Permissions change. A process is redesigned. A model version is deprecated and behaviour shifts slightly. Each of these needs attention, and none of them is optional.

What a complete case looks like

Build cost, then the six run costs above, against the current cost of doing the work. Stated over three years, not one, because year one is flattering — the content was just refreshed, the exceptions are still being learned, and the volume has not scaled.

Cases built this way are less exciting and considerably more likely to survive a finance review. They also make the right ideas more obvious, because the uses with genuine volume absorb these costs easily and the marginal ones do not.