Free field guide · worksheet included
AI workflow evidence kit
Document a workflow with fair comparisons, quality checks, and an honest record of setup, review, rework and failed attempts.
No signup required. The worksheet is a blank template. Examples are illustrative unless explicitly identified as observations.
Define the task, establish a fair manual comparison, log every attempt, and review both outputs against the same standard. A fast wrong answer is not a useful improvement.
Choose one repeatable task.
Start with a campaign brief, meeting summary, research outline, or another task you already understand. Define the intended reader and acceptance criteria before using AI. Use approved, sanitized or fictional inputs. Record the tool, model/version when visible, date, instructions, and relevant settings.
Keep the input, output, review and next decision together. An undocumented result is difficult to reproduce or teach.
Check quality and effort separately.
| Check | Record | Example criterion |
|---|---|---|
| Accuracy | Unsupported claims and corrections | Every action is supported by the source notes. |
| Completeness | Missing or unknown fields | Owner and deadline remain unknown when unspecified. |
| Usability | Can the reader act on it? | Decisions are separate from suggestions. |
| Effort | Setup, generation, review, rework and failures | Correcting the wrong owner counts. |
| Privacy | Approved input and sharing boundaries | Use fictional notes for practice. |
Run a fair comparison.
- Complete a representative task using your usual process. Record effort and quality.
- Use AI on an equivalent task. Include setup and failed attempts.
- Review both outputs with the same checklist. Keep unresolved errors visible.
- Repeat with different inputs, including an incomplete or difficult example.
- Record instruction changes and choose keep, change or stop.
Small samples describe your observations; they do not establish a universal productivity claim. Only compare effort when both outputs meet the same standard. If amortizing setup, state the number of repetitions and retain the full setup cost. Do not report a ratio when review, quality or comparison effort is unknown.
Try a complete fictional meeting-notes example.
All names and notes below are fictional teaching material. The goal is to preserve commitments, proposals and unknowns.
Source notes
Decision: keep the Monday intake check-in. Mina will send the draft agenda on Thursday. Jules suggested adding a FAQ page; no owner or deadline was agreed. October 28 was mentioned as a possible launch date, but was not approved.
Instructions to try in an approved AI tool
Use only the source notes. Separate confirmed decisions, confirmed actions, proposed actions, and unresolved questions. For each action include an exact source excerpt, owner and deadline. Write "not specified" when absent. Do not infer commitments or convert relative dates into calendar dates. Finish with questions for the human reviewer.
Expected reviewed output
Decision: keep the Monday intake check-in.
| Action | Status | Owner | Deadline | Source excerpt |
|---|---|---|---|---|
| Send the draft agenda | Confirmed | Mina | Thursday | Mina will send the draft agenda on Thursday. |
| Add a FAQ page | Proposed | Not specified | Not specified | Jules suggested adding a FAQ page; no owner or deadline was agreed. |
Questions: Which calendar date is Thursday? Will the FAQ proposal be approved, and who would own it? What launch date, if any, is approved?
“October 28” must remain an unapproved possibility. It is not a deadline for the FAQ. The expected output is an editorial teaching answer, not a claim about an AI model’s measured performance.
The free checker includes a deliberately flawed draft so you can see two text-check flags and practice correcting them.
Make the method teachable.
Write a short evidence note: task and audience; input boundary; manual process; AI-assisted process; review method; observations; limitations; next decision. Ask someone else to follow it on a new practice input. Where they need clarification, improve the instructions.
A team pilot needs clear criteria for appropriate inputs, unacceptable outputs, human approval, and a fallback to the original process. Recheck after meaningful task or tool changes. Publish sanitized examples only when authorized.
Use the existing practice tools.
Method and scope.
This is an original editorial worksheet, not a validated benchmark or certification. For broader evaluation guidance, see NIST’s AI Risk Management Framework. Apply your organization’s data and AI policies.