Measurement · Operating decision guide
What will an AI workflow cost at normal and peak volume?
Otomat editorial · Revised
Budget for operating costs per accepted AI task. Include model and tool charges, retries, human review and support, then test a peak month before rollout.
Choose the denominator finance can use
A model call is not a completed business task. One invoice can require extraction, lookup, a second attempt and a reviewer before it is accepted. Record incoming items, completed items, rejected items and carried backlog separately. Divide cost by accepted tasks only after reconciling the unfinished work. Comparing cost per call with a manual end-to-end cost gives the cheaper-looking option an unfair advantage.
Observe the expensive cases
Sample ordinary cases and the exceptions that create extra retrieval, longer documents, tool calls and review. Capture the actual provider charge where available, plus reviewer minutes and recurring support. Keep sensitive content out of the cost log. A fixed subscription, a minimum commitment and a metered API each behave differently at low volume; put those terms in the budget rather than assuming every cost scales smoothly.
| Check | Evidence to bring | Decision or stop condition |
|---|---|---|
| Variable service cost | Provider receipts and tool charges by task type | Include retries and exceptions, not only successful first attempts. |
| Human review | Minutes per task, loaded rate and available monthly hours | Hold when the planned queue exceeds the reviewer’s capacity. |
| Fixed operation | Licenses, monitoring, support and minimum commitments | Check that included allowances are not charged twice. |
| Peak response | Approved cap, alert owner and manual fallback | A spending alert must lead to a decision someone can execute. |
Stress volume and review together
Model a normal month, a peak in arrivals and a month with a worse case mix. Higher volume can increase review queues even when model cost is modest. Show the available reviewer hours beside the cost estimate. If the queue exceeds staffing, reduce scope, add approved capacity or hold the release. A low average task cost is not acceptance of an unmanageable service backlog.
Set a budget response in advance
Name who can approve additional spend and who can stop the workflow. Use an alert threshold below the maximum authorized spend, with a manual path for urgent work. Track implementation costs separately from monthly operation and include them when evaluating the investment. The calculator below is a planning exercise; replace its assumptions with observed volumes and actual commercial terms.
Stress-test the monthly operating budget
Illustrative assumptions in US dollars. Replace them with your task mix and provider terms. Your entries stay in this browser page; they are not stored or submitted.
$18,165.00 / month in modeled operating costs.
- Model and tools: $1,040.00
- Review: 125 hours · $15,625.00
- Fixed operation and support: $1,500.00
- Cost per accepted task: $1.82
Include review of rejected work in your average. Do not count included provider allowances twice. Implementation, transition, tax and any unentered charges are excluded; price them separately. This is a cost estimate, not savings or a staffing recommendation.
Questions before the decision
Are token prices enough for an AI budget?
No. They omit retries, tools, integration, human review, support and fixed commitments. Use the total cost of accepted work for a useful comparison.
Does the calculator predict savings?
No. It estimates operating costs and reviewer hours from your assumptions. A savings case needs a comparable baseline, observed results and finance review.
Take the evidence sheet to the meeting
Download the blank CSV, assign an owner to each check and record the evidence. Keep completed sheets in your own approved system.
Download the evidence sheet (CSV)Use the working tools
Blank templates are ungated. Adapt them to the workflow; inputs are not submitted.
Sources and limits
These are Otomat’s proposed operating methods and original decision artifacts. External references inform risk management; they do not certify this method or establish client results. Linked sources were selected for this review; their live revision dates may change.
- US GAO: AI accountability framework and evaluation context
- UK government: AI procurement guidance (public-sector context)
For anonymous documentation and proposal patterns, see delivery evidence and its boundaries.
Related decisions
Bring a workflow to the discussion
Share the business goal and constraints through the mandate process. Do not send sensitive production data in the initial brief.
Book a Return Review