# AI workflow acceptance worksheet

An editable template from Agents Autonomous: https://agentsautonomous.ai/proof

This is a blank test record, not evidence of a completed project. Use authorized, redacted, or synthetic inputs and agree where the completed record may be stored.

## Test scope

- Workflow, users, and permitted actions:
- Baseline version and candidate version:
- Model, instructions, tools, and source versions:
- Test date and sample-selection method:
- Named reviewer:
- Acceptance criteria agreed before the test:
- Critical failures that block release regardless of average score:

## Case record — duplicate this section for every case

- Case ID and category:
- Source/input reference (avoid embedding confidential contents here):
- Preconditions, including user role and permissions:
- Expected output or permitted action:
- Expected stop, refusal, or escalation condition:
- Baseline observed result:
- Candidate observed result:
- Evidence reference and source version:
- Calls and retries:
- Model/tool costs for all attempts:
- Human review and correction time:
- Defects and severity:
- Reviewer decision: accepted / rejected / needs investigation
- Reviewer, date, and reasoning:

Count one task once. A retry does not create an additional accepted task. Decide in advance whether a corrected output can count as accepted, and apply the same rule to both candidates. Include correction effort in the cost.

## Coverage checklist

- [ ] Normal task with complete information
- [ ] Missing required information
- [ ] Conflicting or outdated sources
- [ ] Duplicate input or repeated submission
- [ ] Tool timeout, rate limit, or unavailable destination
- [ ] User without the necessary permissions
- [ ] Request outside the agreed scope
- [ ] Source text that tries to redirect the workflow's instructions
- [ ] Human review and manual recovery path

Choose cases appropriate to the task. This list is a starting point, not a certification or exhaustive security test.

## Sample comparison

| Measure | Baseline | Candidate |
| --- | --- | --- |
| Unique tasks attempted | | |
| Unique tasks accepted | | |
| Critical failures | | |
| Model charges, including failed attempts | | |
| Allocated tools and infrastructure | | |
| Human review and correction cost | | |
| Total operating cost | | |
| Cost per accepted task | | |
| Review time and completion time | | |

Cost per accepted task = total operating cost / unique tasks accepted.
If none are accepted, the unit cost is undefined. Keep the total spend visible.
Separate one-time implementation cost from this operating-cost comparison.

## Release decision

- Decision: accept for the agreed scope / revise and retest / stop
- Unresolved defects and limitations:
- Evidence that the quality floor is met:
- Monitoring and failure owner:
- Manual fallback and rollback conditions:
- Access, retention, and handover checks:
- Approver and date:

A passing sample does not establish that every future task will succeed. Record the tested scope, continue monitoring, and retest when models, instructions, sources, or integrations change.

Plan the scope: https://agentsautonomous.ai/how-we-work#project-brief
Discuss the acceptance criteria: https://agentsautonomous.ai/contact