# Pilot evaluation templates

These editable JSON templates accompany Plexnova AI's pilot guides. All sample policies, identifiers, values, and incidents are synthetic. They are planning examples, not customer results or a validated benchmark.

## Use the templates

1. Save a copy in your team's approved working environment. Enter the workflow, source, and review owners and the configuration being evaluated.
2. Replace or extend the examples with representative, approved material. Keep the expected evidence and access rules with each case. The document template's text fixtures do not test OCR or scan quality; use representative approved files for that evaluation.
3. Separate development cases from the cases used for the release decision. Include routine work, unsupported inputs, permission failures, unavailable integrations, and exceptions.
4. Agree acceptance criteria and measurement rules before running the pilot. The release thresholds are deliberately empty: the examples do not establish suitable targets for your business.
5. Record a comparable baseline using the current workflow. Leave unmeasured values null.
6. Execute each case against your pilot. Save the actual output and observations, then have the assigned reviewer score every expected check.
7. Mark a case pass only if all its required checks pass. Record failed, blocked, and not-run cases separately. A blocked or not-run case is not a measured zero.
8. Populate the metrics using their stated definitions and actual sample counts. If a denominator is zero, leave the metric null and explain why.
9. Investigate any access disclosure, unauthorised action, or unowned exception independently of average quality or speed scores.
10. Record the release decision and its evidence. A decision to narrow scope or improve the source material can be a useful pilot outcome.

## Interpret results carefully

The six illustrative cases in each file cover selected failure patterns. They do not establish statistical performance, production readiness, or expected savings. Extend the set to reflect the work and conditions you actually intend to support, and report results by request type, document family, or source/access boundary where relevant.

For time and cost comparisons, keep the workload and observation period comparable. Include correction, review, retry, and operating work. Report unanswered or unsuccessful tasks separately rather than silently dropping them.

These downloads are static files. Your team chooses where its evaluation records are stored and who may access them.

## Templates and guides

- Knowledge Assistants: knowledge-assistant.json and https://plexnovai.com/blog/knowledge-assistant-pilot-checklist
- Customer Support AI: customer-support.json and https://plexnovai.com/blog/customer-support-ai-pilot-checklist
- Document Automation: document-automation.json and https://plexnovai.com/blog/document-automation-pilot-checklist