Personalized AI infrastructure
Cost Per Task for AI Agents: How to Track Model Spend by Workflow
Learn how to attribute model and tool spend to each business task, so you can see which AI agent workflows justify a more capable model and which do not.

Cost per task for an AI agent is the total model and tool spend needed to complete one business outcome, such as answering an order question or preparing a refund recommendation for a person to approve. To work out superintelligence cost per task, tag every model call and tool call with the workflow and task it served, add up the spend per completed outcome, and set that figure against what the outcome is worth to the business. That comparison shows whether a workflow deserves a more capable model or whether a smaller one does the job.
Key takeaways
- Measure cost per completed outcome, not cost per API call or per vendor invoice.
- Tag every model and tool call with a workflow and task identifier before you analyze anything.
- Count retries, escalations and human corrections as part of the task, because they are part of the work.
- A more capable model is justified only where a correct result is worth more than the extra spend.
- Change model routing one workflow at a time, and review the figures on a fixed schedule.
Why the monthly invoice does not answer the question
Vendor invoices show how much you spent on each model, API key or tool. They do not show which business activity consumed that spend. Two workflows can run on the same model, and one of them can cost far more per completed outcome because it retries often, reads long documents or calls a paid data service at every step. Without attribution, the usual responses are guesses: move everything to a cheaper model and hope quality holds, or keep the expensive model everywhere because nobody can point to where the money goes.
The question a leader needs answered is narrower. For this task, what does a finished outcome cost, how often does the agent get it right without help, and what does an error cost when it happens? Those three answers support a decision. None of them can be read from a vendor dashboard.
Define the task and tag every call
Start with the unit of work. Write down the trigger that starts the task, the state that counts as done, the person who accepts the result, and the failure that would matter most. A task is finished when its outcome is accepted or handed to a person, not when the last model call returns. A run that ends in an escalation is a different outcome from a resolved case, so the two should not share a denominator.
Then make the tags machine-readable. The OpenTelemetry generative AI conventions define shared attribute names for things like the requested model, the responding model and token counts, which helps you keep the same fields across providers. Your own records add the business fields that no provider knows about. A practical tagging set for each run looks like this:
- A stable workflow name that matches how your team describes the work, such as order status or refund review.
- A task identifier that stays the same across retries, tool calls and handoffs.
- The model name, version and token counts for every model call, with a flag where the provider reports cached input.
- The tool provider, the action name, and whether the call only read data or also wrote it.
- The final outcome: completed, escalated to a person, rejected or abandoned.
- Any human correction, with a reason code recorded by the reviewer.
Hidden spend usually sits in three places: retries after a tool error, retrieval that pulls in more context than the answer needs, and loops where an agent reconsiders a step it has already settled. Log each retry as its own event so it appears in the total instead of disappearing into an average.
How to calculate cost per completed task
For each workflow and period, add the model spend and tool spend for every run tagged to it, including runs that failed or were escalated. Divide that total by the number of outcomes accepted without rework. Keep a second line for the cost of escalated runs, since a person's time belongs to the task even when the agent did useful work first. Then compare the result with the value of the outcome in your own terms, such as the margin on an order, the staff hours saved, or the revenue kept from a customer.
Illustrative scenario: a small wholesale distributor runs an agent that drafts replies to order-status questions for a staff member to send. Tagging shows that whenever a shipping lookup fails, the agent retries and then passes the message to a stronger model, and those messages account for most of the rewriting staff do afterward. The team first repairs the lookup connection, which was the real cost driver. They then keep the stronger model only for messages that mention a disputed invoice. The scenario is illustrative and does not describe a specific company.
Decide which workflows justify a more capable model
Before you change a model, answer a short set of questions for that workflow using the tagged data:
- What does a wrong answer cost in money, time or customer trust, and would a person catch it before it leaves the company?
- Does the task need reasoning across several records, or does it mostly retrieve, classify and format?
- Does a cheaper model fail in ways your checks can detect, such as a missing field or a clearly wrong category?
- Has a test set built from your own past cases shown the difference, rather than a public benchmark score?
- Would the extra spend come back through fewer escalations, shorter handling time or fewer reversals?
Public benchmark results rarely settle this for your workflow. Your own cases, your own tools and your own definition of a correct outcome do. The guide to which superintelligence benchmarks predict business results explains how to separate a useful signal from a headline score. Anthropic's writing on evals for agents also points out that a run should be judged by the state it reaches, not only by the wording of individual replies.
How Be Human handles cost attribution
Be Human's managed approach fixes the attribution fields during the first engagement, when each agent role is defined with its memory, tools, approval boundaries and success criterion. Every run then reports spend against a workflow the agent was built to perform. Monitoring keeps the tags complete, recurring evaluations test whether a model or routing change still meets the success criterion on current cases, and permission reviews confirm which tools an agent may call and which writes need approval. Model changes are handled as controlled improvements with a recorded reason, so any cost decision can be traced back to the evidence behind it.
Review cost per task on a fixed schedule
Run the same review each period so the numbers stay comparable. Keep the meeting short and focused on decisions:
- Compare cost per completed outcome with the previous period for each workflow.
- Read the escalation and rework reasons, and assign each one to a fix in the tool, the data or the model.
- Check whether new tools or retries added spend without a matching gain in completed outcomes.
- Change at most one routing rule per workflow, then measure its effect before making the next change.
Review also tells you when the agent's permissions no longer match its workload. If a workflow has grown to include new write actions, the permission review steps for AI agents show how to check those changes before the next cycle. Teams that capture these findings each period build a record that improves later decisions, as described in feedback loops that make company AI better every quarter.
If you are not yet sure which of your workflows should be the first to receive this kind of tracking, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
How do I calculate superintelligence cost per task for one workflow? Add the model and tool spend for every run tagged to that workflow in a given period, including runs that failed or were escalated. Divide the total by the number of outcomes accepted without rework, and record escalated runs on a separate line. Compare the result with the value of a correct outcome in your own terms.
Do I need a more capable model for every task? No. A more capable model is worth its extra spend only where an error is costly, the task needs reasoning across several sources, and a cheaper model fails in ways your checks cannot catch. For simple retrieval, classification and formatting, a smaller model usually meets the success criterion at lower cost.
Which costs should I count besides model tokens? Include every tool call that the agent makes, including calls that only read data, because each one adds latency and sometimes fees. Count retries, long context retrieval, and the time staff spend correcting or escalating an output. Those costs often exceed the model bill for workflows that touch several systems.
How often should cost per task be reviewed? Review it on a fixed schedule, such as monthly for active workflows, so the numbers stay comparable from one period to the next. Change at most one routing rule per workflow between reviews, so you can see what each change did. Rerun your evaluation set after any model or prompt change before relying on the new figure.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


