What an AI Agent Really Costs After You Sign

The licence is the number the committee debates. The one that actually matters, what it takes to sustain the agent once it's live, never makes it into a proposal.

Carlos Andrés Ramírez · 2026-09-06

A COO told me a few months ago that the committee had approved an agent to triage support requests with one clear number in mind: the annual licence fee. Six months in, the compute bill and the hours his team spent reviewing flagged cases had already passed that number, and nobody had asked about either before signing.

That pattern shows up in almost every committee that approves a decisioning agent. The number everyone debates is the licence fee, because it is the only one the commercial proposal states. The real cost, keeping the system working well month after month, never shows up on a line item, because nobody selling the project has a reason to put it there.

And that cost is neither small nor optional. An agent that decides something, approving a refund, prioritising a case, drafting a reply, needs someone watching for when it drifts, someone reviewing the cases it flags as uncertain, and someone retraining it once the world it was built on stops matching the world it now runs in. The vendor does not do that work for the price of the licence. The team that bought the system does, or should, and almost nobody budgeted for it.

The symptom

How much does it actually cost to keep an AI agent running in production, beyond the licence?

The short answer is that almost nobody has measured it carefully yet. The long answer starts by admitting the cost is not a fixed number: it is a curve that grows with how many decisions the agent makes, and it grows faster than the licence does.

• The project budget covers the licence and the build, and stops there; nobody set aside a line for what comes after launch.
• The team reviewing the cases the agent flags grows every quarter, and nobody counts that as project cost, it gets folded into normal operating cost instead.
• The compute and storage needed to keep a full record of every decision, so it can be audited later, climbs quietly and gets buried inside the general technology bill.
• Nobody owns checking whether the agent started drifting, so it gets discovered through a customer complaint, not an internal alert.
• The vendor contract never states what happens to the price once usage volume rises, and volume rises exactly when the project starts working.

The problem underneath

The budget gets signed to switch the agent on, not to keep it running.

Switching an agent on is a project, with a start date and an end date, and projects get budgeted properly: scope, timeline, price. Keeping it running is an ongoing job with no end date, and ongoing jobs rarely go back through the same committee that approved the project. So the cost of sustaining it does not disappear, it just moves to a different budget, technology or whichever function runs the agent, where nobody ties it back to the original decision to buy or build.

The outcome is predictable. The project closes with a number that looked reasonable, the agent goes live, and a few months later someone in finance asks why the compute bill rose without anyone approving anything new. The answer is almost always the same: the agent is now handling more than it did in the pilot, and sustaining it at that volume costs more than sustaining it in the pilot, even though the licence fee never changed.

The licence tells you what it costs to switch an agent on. It does not tell you what it costs to keep it working well a year from now.

BECOME

The framework

What to budget before an agent goes into production.

These five line items get budgeted before signing the proposal, not after the first surprise invoice arrives.

Monitoring

Who checks, and how often, whether the agent is still deciding within expected limits. Without someone assigned to this, drift gets discovered through a complaint, not an alert.

Retraining

What it costs to update the model or the data feeding it once the business changes, a new product, a new policy, and the agent starts deciding against a version of the world that no longer exists.

Exception review

The human time spent looking at the cases the agent flags as uncertain. It grows with volume, it does not shrink over time, and it almost never gets counted as part of the project cost.

Audit trail

The compute and storage needed to keep a full record of every decision, so it can be checked if someone questions it. Without that record, the agent decides and nobody can reconstruct why.

Repricing

What happens to the price once usage volume rises. Almost no contract fixes that number upfront, and volume rises exactly when the project starts proving it works.

Add these five up and they usually pass the licence fee within the first year, and that is not a sign the project failed. It is the real cost of a system deciding every day instead of running in a demo. The mistake is not that the cost exists. It is budgeting for it only after it arrives.

Before approving the next proposal for an AI agent in production, ask the vendor or the internal team to estimate each of the five line items above, with a number next to each one. If nobody can put a figure on them, you do not yet know what this project will cost you, only what it costs to switch it on.

Let's budget what it takes to sustain your agent