# AI workflow evaluation plan and go/no-go worksheet

AI-assisted editorial draft • Bill Allred review pending • Draft and sources checked 19 October 2026

Original illustrative tool. Thresholds must be chosen for the actual consequences and operating context.

## Instructions

Complete scope and severity rules before testing. Label expected outcomes independently of the system under test. Keep a held-out acceptance set separate from development examples. Replace illustrative entries with evidence from the actual workflow. Record uncertainty and unresolved issues beside results. An empty box is unknown, not passed.

## 1. Acceptance contract

- Workflow / business owner / technical owner:
- Intended users and permissions:
- Allowed inputs, outputs, and actions:
- Explicitly excluded decisions/actions:
- Authoritative records and freshness requirement:
- Automated checks and their limits:
- Model-grader rubric, calibration cases, and missed-error measure:
- Human judgment and escalation owner:
- Business baseline / intended benefit:
- Version and environment under test:

## 2. Representative test-case sheet

| Case ID | Segment / input | Authoritative evidence | Expected outcome | Check method | Severity if wrong | Actual result | Pass/fail and evidence |
|---|---|---|---|---|---|---|---|
| HYP-01 | Hypothetical: exact order ID, current record | Order status = dispatched | Return dispatched with source | Exact value and evidence match | Material misinformation | Not run | Unknown |
| HYP-02 | Hypothetical: conflicting identifiers | Two distinct matching orders | Ask for clarification; no status | Route and output assertion | Material misinformation | Not run | Unknown |
| HYP-03 | Hypothetical: stale record | Timestamp exceeds defined limit | Withhold status; escalate | Freshness gate assertion | Material misinformation | Not run | Unknown |
| HYP-04 | Hypothetical: unauthorized order | Permission fixture denies access | No retrieval/disclosure | Access enforcement test | Critical | Not run | Unknown |
| HYP-05 | Hypothetical: source outage | Controlled timeout | Fallback; no invented status | Tool failure and output test | Material misinformation | Not run | Unknown |
| HYP-06 | Hypothetical: malicious instruction in note | Untrusted source text | Ignore instruction; retain boundaries | Adversarial behavior test | Critical if boundary crossed | Not run | Unknown |

## 3. Failure-severity rules

| Level | Illustrative consequence | Proposed decision rule to adapt |
|---|---|---|
| Critical | Unauthorized disclosure/action; potentially serious harm | Block affected release; investigate and retest |
| Material | Wrong business fact that changes a decision | Resolve or explicitly narrow scope; accountable owner decides residual risk |
| Minor | Recoverable presentation defect without meaning change | Document fix, owner, and monitoring; assess cumulative impact |

Set actual acceptance thresholds before the run:

- Correct automatic completion: numerator / denominator / required threshold:
- Known wrong outputs missed by checker: missed / injected wrong outputs / required threshold:
- Correct outputs blocked: blocked / injected correct outputs / required threshold:
- Escalation rate and capacity needed:
- Exception resolution time and required bound:
- Critical failure rule:
- Required segment-level results:
- Sample-size rationale and uncertainty:

## 4. Escalation rules

| Trigger | System action | Human role / decision | If unavailable |
|---|---|---|---|
| Missing or conflicting evidence | Withhold affected answer/action | Assigned operations reviewer resolves source | Use established manual process |
| Checker disagreement | Hold result with evidence | Qualified reviewer determines disposition | Keep held; notify accountable owner |
| Access or unauthorized-action failure | Stop affected capability | Technical/security owner investigates | Disable route using tested procedure |
| Judgment outside scope | Route with context | Authorized business decision-maker | Do not infer approval |

## 5. Release checklist

- [ ] Scope and prohibited actions approved by accountable owners
- [ ] Representative held-out cases and negative tests complete
- [ ] Checker false accepts and false rejects measured
- [ ] Segment-level results and uncertainty recorded
- [ ] No unresolved failure violates release rules
- [ ] Permissions, dependency failure, and rollback tested
- [ ] Users can identify boundaries and reach support
- [ ] Escalation workload fits available capacity
- [ ] Logs have appropriate access and retention decisions
- [ ] Business baseline and benefit measurement ready
- [ ] Retest triggers assigned: model, prompt, data, tool, checker, or policy change

Decision: Go / Go with narrower scope / No-go  
Approved scope and version:  
Residual risks and accepting owner:  
Open issues, owners, and next decision:  
Date and evidence links: