How to track the cost of each AI agent

Tag every model call and action with the agent and capability that made it. Add the people costs the bill never shows: approver and reviewer time from the held queue and sampling logs, agent operations and certification effort, and rework. Report cost per completed item by capability and depth each month, signed by the agent owner and finance.

Why is the cost of an agent hard to see?

Because most of it is not on the invoice. The model provider's bill is visible and easy to discuss, but it is often a small share of what an agent really costs. The rest is spread across teams: the approver who reviews held items, the reviewer who samples quality, the analyst who maintains the brief, the certification runs, the incident that took a day to unwind. The argument is made in full in the cost of an agent is not the model bill.

Tracking cost per agent matters for two decisions. Whether the agent is worth running at all, compared with the work done by people. And which capability to change: an agent whose cost is mostly approver time needs a different fix from one whose cost is mostly model use.

The agent owner owns the cost report. Finance owns the allocation method and the link to budgets. Agent operations provides usage and effort data.

What are the steps, and what evidence does each leave?

1. Finance and agent operations: define cost categories. Model use; platform fees; infrastructure and connectors; people time for approval, review, sampling, maintenance and certification; incidents and rework; set-up costs spread over an agreed period. Evidence: the category list and method.
2. Agent operations: tags every call and action with the agent and capability identifiers, at the moment it happens. Evidence: tagged usage records.
3. Finance: sets the allocation rule for shared costs, such as a platform fee covering several agents. Usage share is the usual basis. Evidence: the allocation rule, versioned.
4. Agent, depth 1: records time spent on its held items, from the moment an approver opens an item to the decision. Evidence: held item timings.
5. Reviewers and analysts: log time on sampling, brief maintenance and certification against the agent, in the same system they log other work. Evidence: time records.
6. Agent operations: records incident effort and rework, linked to the agent and capability. Evidence: incident and rework log.
7. Agent, depth 1: compiles the monthly figures: total cost by category, cost per completed item by capability and depth, and the trend.
8. Agent, depth 2: drafts the cost report, including comparison with the baseline cost of people doing the same work.
9. Agent owner and finance: review and sign. Evidence: the signed report, using a format such as the AI agent cost tracking template.

Why report cost per completed item by depth?

Because depth changes the cost structure. At depth 2, every item is drafted by the agent and reviewed by a person; you pay for both. At depth 3, most items are completed by the agent, and people only sample. An agent at depth 2 can easily cost more per item than people working alone, because review is not free. Moving it to depth 3, when the evidence supports it, often changes the economics completely.

Reporting by depth makes this visible. It also shows the cost of caution: a capability held at depth 2 for months because no one revisited it.

Where does people's time come from?

The held queue is the best source. Each held item has a time the approver opened it and a time they decided, which gives an approval cost per item. Teams that work through the held queue each day generate this data as a by-product. Approvers who keep items open for hours while doing other things inflate the figures, so use the median and note outliers.

Sampling, maintenance and certification time must be logged by the people doing it. Keep the categories few, or no one will log accurately.

What should you measure?

• Total cost per agent per month, by category.
• Cost per completed item, by capability and depth.
• People time as a share of total cost. A high share points to review effort as the thing to reduce.
• Rework and incident cost, which should fall as the agent stabilises.
• Comparison with baseline: what the same work cost when people did it alone.

A full approach to what goes into these numbers is in how much it costs to run an AI agent.

Where does it usually go wrong?

Only the bill is tracked. The model cost is reported and the agent looks cheap. Six months later, a team asks why it needs two more approvers.

Tags are added later. Usage from the first months cannot be split, and the trend starts from the middle. Tag from day one.

Shared costs are ignored. The platform fee is held centrally, and each agent looks cheaper than it is. Allocate.

Set-up cost is forgotten or overloaded. Some teams ignore it; others load it all into the first month and declare the agent expensive. Spread it over an agreed period.

Cost is cut in the wrong place. Reducing review time by approving faster lowers cost and quality together. Look first at reducing model use and rework; ideas are in how to reduce the running cost of AI agents. Any change to depth to save cost still goes through train, shadow, certify, widen.

Where Pryme fits

In Pryme, every action and every held item is filed with the agent, capability and verdict, so usage and approval time can be split per capability without extra tagging work. The cost report draws on the same record as the performance review.