Working template · September 2026 · Version 1.2
AI pilot planning worksheet
Test one workflow and its AI agent action boundaries, human approvals, pause, revocation and recovery.
- For
- Process owners and teams moving from AI experiments to daily use.
- You leave with
- A bounded pilot plan with a baseline, review responsibilities and a scale, redesign or stop decision.
Choose your format
DOCX is editable in Word or a compatible editor. PDF is ready to print. No signup is required.
Inside this template
The full document follows these sections. The first prompts are shown below so you can check the fit before downloading.
Ada Studio · Working template
Choose the work
Define the workflow problem before selecting the tool.
Problem and desired outcome
Describe the task, current friction and the decision the pilot should support. Who benefits if it works?
Current workflow and baseline
Map the main steps and record current time, quality, corrections and review effort. State how the baseline was measured.
Participants and ownership
Name the business owner, pilot users, reviewer and the person making the final decision.
Boundaries
Define what the pilot includes and excludes, its duration and which data and tools are approved.
Design a fair test
Measure a useful outcome, including the work needed to review and correct it.
Proposed workflow
Describe where AI contributes, what stays with a person and what happens if the tool is unavailable.
Quality and human review
Define what must be checked, which source material the reviewer sees and when to correct, reject or escalate.
Success measures
Choose measurable targets for quality, total time including review, risk and actual use. Record a comparison method and sample.
Stop conditions and support
List unacceptable failures, the person who can pause the pilot and how users get training or help.
Review the evidence and decide
Complete this page at the pilot review, using observed results rather than impressions alone.
Results against the baseline
Record the sample tested, results for each measure and limitations of the evidence.
User and reviewer feedback
What helped, what caused friction and what work moved elsewhere? Include corrections and near misses.
Decision
Choose: scale, redesign, pause or stop. Record the reasons, approver and decision date.
Next actions and review date
Name owners and dates for process changes, guidance, training and the next review. Keep a record of what changed.
Benefits evidence and portfolio decision
Link this review to a stable use-case ID. Estimated benefits are not measured outcomes.
Baseline and benefits
Compare total work plus review effort, quality, usefulness and costs with the manual baseline. Record method, period and evidence.
Data and dependencies
Review data-quality gaps, access, integrations, suppliers and capacity. Name owners and readiness conditions.
Portfolio decision
Choose pilot, continue, change, scale or retire. Record evidence, rationale, authorised approver and conditions.
Handoff and follow-up
Update the use-case register. Assign implementation, supplier, training and incident-learning actions, with owners and review dates.
Action boundaries and recovery tests
Run in the approved test environment. Record expected and observed results, evidence, owner and date for each test.
Allowed and prohibited actions
Test permitted actions and attempts outside scope. In the fictional support example, read approved guidance and draft; verify that sending and customer-record changes are denied.
Approval boundary
Test the required approval and its absence. Change the proposed action, recipient or record after approval and check that it cannot execute under the old approval. Name the authorised reviewer.
Pause and delegated-access revocation
Pause new work and revoke the agent’s delegated access. Check queued and in-flight actions and retained logs. Record actual effects and any delay before access stops.
Recovery and reauthorisation
Validate restored state and manual fallback. Check uncertain outcomes before retrying to avoid duplicate sends or updates. Name who authorises restart; preserve failures, corrections and retest evidence.