AI agent cost tracking for accepted work

AI agent cost tracking measures the resources used by a complete workflow and connects them to its result. The useful question is not only “What did the model call cost?” It is “What did it cost to produce an accepted result?”

Track three levels of cost

Keep the levels separate so a cheap call does not hide an expensive workflow.

Input cost

Input cost covers the resources used by one execution step. Depending on the worker, this can include model tokens, tool calls, compute, and paid services.

Workflow cost

Workflow cost covers the full run. It can include input cost plus context preparation, retries, review, human waiting, coordination, and rework.

Accepted-result cost

Accepted-result cost connects the workflow total to a result that passed the required quality and review bar. It lets teams compare different routes without counting a failed or rejected output as equivalent to accepted work.

Define the denominator first

Choose what counts as accepted before comparing costs. The acceptance event should identify the artifact, reviewer or evaluator, decision, and time.

A basic period metric is:

Cost per accepted result = total recorded workflow cost ÷ accepted results

Report coverage beside the number. If only some runs have recorded cost or acceptance evidence, the metric must not imply complete coverage.

Itemize the full workflow

A practical cost record can include:

• model input and output;
• tools and external services;
• context preparation;
• retries and repeated steps;
• automated and human review;
• time waiting for a decision;
• coordination between systems or workers; and
• rework after a failed check or review.

Some items are direct currency values. Others are time or indexed values. Keep the unit clear. Do not silently mix measured spend and modeled labor cost.

Join cost to quality evidence

A route change can reduce spend and reduce quality at the same time. Compare alternatives only when they face the same acceptance bar.

For each candidate, retain the workflow version, worker or model route, test or evaluator result, review result, final status, and cost. This makes a cost change open to review instead of treating the cheapest call as the best result.

Separate measured and illustrative values

Measured values come from actual run records. Illustrative values explain a method with a fixed example. Label the difference on the page and in exports.

Aestus uses this rule in its public optimization demo: its indexed ten-run curve is marked as illustrative. The product record uses the team's runs and its own acceptance evidence.

Review the failure paths

Failed and canceled runs still spend resources. Keep them in total spend and report them separately from accepted results. Also record missing cost coverage and missing quality evidence.

Useful companion measures include:

• acceptance rate;
• first-pass acceptance rate;
• retry and rework rate;
• failed-run cost;
• human wait time; and
• cost coverage.

Together, these measures show whether a lower cost came from real improvement, missing data, or a weaker result.

Build a baseline

Start with one workflow and a fixed period. Record each step, its cost source, elapsed time, retries, review, and acceptance decision. Do not optimize until the baseline is complete enough to explain where the total came from.

Then change one part at a time. Reuse context, move a predictable check earlier, or test a lower-cost route. Keep the candidate only when it clears the same acceptance bar.

Read what AI agent operations is for the full work-record model. The accepted-result walkthrough shows how Aestus separates an accepted artifact from a run that merely finished.