Worksheet · LC-R01
Workflow Cost Limits Worksheet
Record what one workflow is allowed to spend or repeat, where each limit is enforced and what happens when it is reached.
Use this before a workflow can retry, call tools or run in the background. Start with one workflow, including repeated submissions and background work.
By Sunil Mathew, co-authored with Claude.
The complete worksheet is below. Copy the table into your working document or spreadsheet and keep the units and currencies visible.
What this helps you decide
A budget estimate describes expected cost. An enforced ceiling limits further work at a defined boundary. Record both the reason for your limit and the component that can stop the workflow.
The worksheet has five starting boundaries and eleven columns. Add a row when your workflow needs another boundary; keep different currencies in separate rows.
Worked example
Illustrative arithmetic only. The amounts make the calculation visible. They are not provider prices, a rate card or measurements of an actual system.
A task uses USD 0.10 of model work per attempt. Two attempts use USD 0.20 before retrieval, other tools or human review. If neither attempt succeeds, those attempts alone cannot establish cost per successful task. Keep failure costs in the record alongside completed work.
For this example, the design permits at most two attempts. At the limit, it saves the unresolved outcome and offers a review route. Two attempts is a teaching choice, not a universal recommendation. Concurrent requests and retries must share the intended accounting boundary.
Complete the ceiling table
For each boundary, explain the proposed limit, unit, reason, enforcement point, stop behaviour, owner, test and evidence. Include model calls, tool calls, elapsed time, run cost and aggregate cost per person or period where relevant.
| Review question | Your decision should explain |
|---|---|
| What counts as one run? | How retries and repeated submissions are recognised |
| What work is charged? | Model, retrieval, tool, background and review costs |
| Where is the limit enforced? | The component that can stop further work |
| What does the user experience? | Saved state, unresolved outcome and next step |
| What happens concurrently? | How simultaneous requests share the remaining allowance |
| When do we review it? | Changes in task, traffic, pricing or acceptable risk |
The ceiling table
Empty cells are for your own workflow. Each row needs a decision; the worksheet provides no score or recommended budget.
| boundary | limit | unit | currency | reason | enforcement_point | stop_behaviour | owner | boundary_test | evidence | review_trigger |
|---|---|---|---|---|---|---|---|---|---|---|
| Attempts per run | ||||||||||
| Tool calls per run | ||||||||||
| Elapsed time | ||||||||||
| Cost per run | ||||||||||
| Aggregate person or period cost |
The file is the table above, with the same column names. There is no PDF: printing this page, or saving it as a PDF from the print dialogue, gives you the whole of it as a form.
Test the boundary
Exercise a run at the limit, repeated requests, concurrent requests and a delayed tool result. Check that completed actions are recognised before a retry and that unresolved work remains visible.
Account for provider billing delays. A local counter may differ from the final billed cost. A hard stop also needs a defined response for work already in progress.
This worksheet does not choose a safe budget or guarantee the final bill. Use measured workload and current prices, then review the design with its owner.
Next experiment: What evidence would justify one of the proposed limits?
Related work
- Use the Evaluation-gates worksheet to connect a requirement to evidence and a release decision.
- Use the Deployment checklist to name the people and recovery paths needed around a release.
- The Run-cost model helps examine the assumptions behind a cost estimate.
Learn to reason about operating limits
Architecting Agentic AI Systems provides live work across a scoped learning project's design, evaluation and deployment. If this worksheet leaves a specific team decision unresolved, bounded advisory is a separate route for discussing it.