Operations
The real cost of running an AI agent.
Model usage is one line in the budget. A useful cost model follows the whole task through review, retries, and completion.
Measure what it takes to get the work done to the standard you need. The useful number is cost per correctly completed task. Count failed attempts, retries, fallback calls, and the time people spend checking or fixing results.
Choose the unit you want to improve
Suppose a team uses an agent to draft a weekly supplier update. Count the job as done when the update meets the team’s standard and is ready to use. One job may need several model calls.
One update may need a document search, model calls, paid data and a person’s review. A failed draft may need another attempt. Keep all of that work under the same task so its full cost stays in view.
Choose a period to measure. Divide the cost of that workload by the number of updates that meet the standard. Include spend on failed attempts in the total cost. Show failed and unfinished tasks too. A neat average should not hide a drop in the amount of work finished.
Figure 01Unit economics
Measure the cost of an accepted result
Count every attempt.
Failed runs still consume resources.
Count the human work.
Review, corrections & upkeep.
Include the work left for people
Keep time spent drafting separate from time spent checking. An agent may collect the files but leave the reviewer with more facts to check. Measure both before deciding that it has saved time.
If you put a price on that time, state the hourly labor rate you used. Time freed for another task is useful. It does not always reduce payroll or bring in revenue. Show what changed in the work, then explain its value.
Count the upkeep too: fixing source material, tracing failures and keeping tools connected. Keep the one-time build cost apart from the cost to run the system. Then choose how long it should take to earn back that build cost.
For the wider project budget, see how we scope and price an AI project.
Does the other 30% erase the saving?
A workflow that completes 70% of tasks without help can still be useful. That depends on catching the remaining tasks and routing them for review. A 70% automatic completion rate is different from quietly letting wrong answers through.
Take a hypothetical batch of 100 tasks. Checks confirm that 70 meet the standard; people fix the other 30. All 100 are then correctly completed. Model and tool charges across every attempt total $20. Assume staff time costs $60 an hour. Doing each task by hand would take two minutes at the same standard, or $200 for the batch.
If each of the 30 fixes takes two extra minutes, the repair work costs $60. Add the $20 machine bill: $80 in total, or $0.80 per correctly completed task. If each fix takes eight minutes, repair costs $240. The total becomes $260, or $2.60 per task, above the $2 manual baseline.
These are example assumptions, not client results. They include no extra human review of the 70 automatic completions and no hosting or build costs. Add those costs when your process needs them. Time freed for other work is not automatically cash saved.
Count a repaired task once. If the other 30 remain unfinished, only 70 belong in the completion count. Keep the cost of all attempts and any failed repair work in the total. Use measured review and repair time before deciding whether the workflow is worth expanding.
Try your own review and correction costs in the agent cost calculator.
Look for repeated work in the trace
Before changing models, inspect a few complete tasks. Does the agent fetch the same supplier profile more than once? Does a formatting error make it repeat the research? Does every section of the report receive the entire document collection? These are specific things an engineer can investigate.
Split the bill into searches, models, paid tools, hosting and review. Show how many attempts each useful result took. Compare routine tasks with costly ones. The average may hide both.
Change one source of waste at a time. Keep the same pass rules. Reusing a research result may avoid another lookup. Sending only needed source text may cut input cost. Test either change: a missing fact can make a cheaper draft take longer to fix.
For Thesis Finance, we added caching between model calls and trimmed options data sent to the model. A recorded offline replay of 11 captured production options turns computed a drop from $2.09 to $0.90 per options turn, about 57%. That is model cost in the replay. It excludes human review and is not a live production before-and-after.
Read the Thesis Finance replay case for the method and its limits.
Figure 02Execution trace
Retry the step that needs work
The acceptance criteria stay the same.
Compare complete tasks when testing a model
A model with lower token prices can still cost more per correctly completed task. It may write more text, make more calls, or need retries and fallbacks. Try a candidate on one clear subtask, such as assigning an update to a category. Use the same inputs and pass rules, then compare the whole cost.
A more capable model may justify its cost on a difficult synthesis step if it reduces expensive rework. It may add no useful value to a straightforward extraction. The choice depends on the behavior you observe in that part of the workload.
Keep a record of the setup, inputs, pass rules and results. An offline test can justify a small rollout. Check the effect on live work before treating that estimate as a saving.
Set a budget the system can enforce
Decide what the agent should do when a task exceeds its expected effort. It might stop with a partial draft and flag the remaining research for a person. The output should clearly identify what is unfinished.
Set limits on run duration, paid tool usage, and repeated attempts. Review alerts against the volume of work: an increase in total spend may simply reflect more accepted tasks, while a flat bill can hide a decline in completion.
Show cost per correctly completed task beside quality, completion rate, review time and speed. Read those measures together. Use them to decide what to improve, what to expand and what still needs work.
Budget limits work alongside the recovery and operational controls in our production-readiness article.
Need help with your agent’s costs?
See how Olypsis benchmarks the work, tests changes and implements cost fixes.
Explore LLM cost optimization