Vice President · TheGreyMatter.ai
In brief
Divide the full cost of the workflow by outputs that meet a stable acceptance standard.
Key takeaways
- Use accepted business work as the denominator.
- Include failed attempts and human review in the numerator.
- Keep cost, quality, volume, and realized business benefit visible together.
A low cost per model request can coexist with an expensive operating workflow. Retries, retrieval, reviewer time, and unresolved cases all consume resources. Cost per accepted task connects those resources to a result the business can use. The calculation below is an editorial measurement approach, not a pricing forecast or a claim about the return any particular AI product will produce.
1. Define the unit before opening the invoice
Choose a completed unit with a clear acceptance standard: an approved support response, a reconciled report, or an account brief ready for its intended meeting. Define when the unit enters the workflow and when it is accepted. A draft and its correction are stages of the same task, not two successes. Count duplicate outputs once. Record an appropriately escalated case separately when escalation is the intended outcome.
The FinOps Foundation's unit economics guidance connects technology spending with business value and distinguishes resource metrics from business metrics. That distinction helps explain why cost per token is useful for engineering diagnosis but insufficient for the operating decision. Both measures can remain on the dashboard. They answer different questions: how efficiently the infrastructure runs and what it costs to complete useful work.
2. Build the cost numerator
For a defined period, collect model usage, retrieval, storage, orchestration, monitoring, and other direct service costs. Add the time people spend preparing inputs, reviewing outputs, correcting failures, and handling exceptions. Use an agreed internal labor-cost method and identify estimates. Keep initial implementation work separate from recurring operation, then show how including that investment changes the business case over the intended period of use.
Allocate shared infrastructure using a documented driver such as usage, task volume, or reserved capacity. No allocation is perfect; consistency and visibility are more useful than false precision. Avoid counting the same shared bill in multiple workflows. Retain an unallocated category when the evidence is insufficient. This makes a missing attribution visible instead of quietly treating the cost as zero or distributing it in a way that flatters one project.
3. Calculate an illustrative result
The basic formula is total attributed workflow cost divided by accepted tasks in the same period. Suppose a hypothetical pilot attempts 120 tasks and produces 90 accepted results. It incurs 180 cost units in services and 420 in preparation, review, correction, and support. Its operating cost per accepted task is 600 divided by 90, or about 6.67 cost units. These are illustrative units, not vendor prices.
Dividing by all 120 attempts would report 5 cost units and hide the effect of rejected work. Keep those attempts visible through a separate acceptance rate: 90 divided by 120, or 75 percent. If the 30 failures later require a manual completion, include that cost when measuring end-to-end delivery. Define the period and status rules so a task waiting for review cannot quietly become either a success or an excluded expense.
4. Compare equivalent work
Measure the manual baseline with the same acceptance criteria, task boundaries, and labor-cost assumptions. Split easy and difficult cases when the mix differs. A team may route straightforward work to automation and retain difficult exceptions manually; that can be sensible, but comparing the two raw averages would misrepresent the result. Report each segment and the combined workflow so readers can see where the economic improvement actually occurs.
Run a sensitivity check before projecting scale. Change the acceptance rate, review time, task volume, and shared capacity assumptions. A workflow can look economical at full utilization and costly when demand is intermittent. Conversely, a modest service bill may be dominated by unnecessary review steps. Use the calculation to identify the next experiment, such as improving evidence quality or narrowing scope, rather than reflexively choosing the cheapest model.
5. Keep value and cost in the same conversation
A lower cost per accepted task does not establish a revenue increase, cash saving, or valuation effect. Record what changed operationally: shorter backlog, faster response, improved coverage, or capacity made available for other work. Verify those effects separately. Recovered minutes can be valuable without becoming a budget reduction. State which type of benefit the organization actually observed and avoid adding unlike benefits into one impressive total.
Review the metric beside total spending, accepted volume, error severity, and time to completion. Set a decision rule before the next trial: expand if quality holds and the relevant economics improve, repair if one cost driver dominates, or stop if the workflow is not worthwhile. Preserve the calculation and assumptions so a later model or process change can be compared fairly. The purpose is a better operating decision, not a cheaper-looking chart.
Common questions
Should failed AI requests count in the cost?
Yes. Failed attempts, retries, and their review consume resources. Include them in the attributed workflow cost while counting only accepted work in the accepted-task denominator.
Does time saved equal money saved?
Not automatically. Time can improve capacity or service without reducing spending. Report the observed operational use of that time and reserve cash-savings claims for changes supported by financial evidence.
Sources and evidence
AI assisted editorial analysis. Primary sources checked October 10, 2026. The procedures are editorial recommendations, not vendor instructions, reported customer results, or claims of personal experience. Numerical examples are hypothetical.
One useful next step: Choose one idea from this note and test it at the smallest scale that could teach you something this week.
