Template
AI Trial Tracking Sheet
Record each trial's results, review time and next action.
Choose a review period that fits your work
Use this sheet for individual practice or a team pilot. Review weekly if useful; there is no required twelve-week schedule. Record events as they happen and keep only information you are allowed to share and retain.
For individual practice, record your task and next step. For a team pilot, also name the reviewer and person authorized to approve changes to shared systems. A proposed action is not an approved decision.
Compare work and quality together
| Period / task | Attempts / accepted outputs | Work time | Quality and exceptions | Resources | Next action / owner |
|---|---|---|---|---|---|
| [Dates and scope] | [Count and criteria] | [Baseline / with AI] | [Checks, corrections, unfinished cases] | [Setup time, tool costs, unknowns] | [Proposal or approved decision, due date] |
Include preparation, changes to prompts or Skills, checking, corrections, failed attempts and handoff. Itemize one-off setup separately so it is not silently excluded or counted twice. Use comparable inputs and the same timing boundaries for the baseline and AI-assisted trial.
Label each figure measured, self-estimated, assumed or unknown; record its source. Course completion is not evidence of workplace use. Record who used the method, how often and whether help was needed.
Worked entry: E1
Illustrative data from E1, not an observed run:
| Item | Entry |
|---|---|
| Scope | Four cases; independent practice: 2/3 participants |
| Time | Baseline 40 minutes; AI-assisted 32 minutes |
| Quality | Baseline 4/4 accepted; AI-assisted 3/4 |
| Resources | API 0.24 USD; setup/testing 30 minutes; other costs not supplied |
| Unresolved issue | E04 stopped without an accepted output; returned data needs verification |
| Proposal | Keep the affected route paused; IT verifies returned data, repairs only a confirmed defect; Lan retests; business owner decides resumption |
The 20% recorded time difference does not establish a productivity gain: the completed work is not equivalent. Do not infer ROI or cash savings.
Record safety checks separately
A refusal required by an access rule can pass a safety test without producing a completed work output. An incomplete or unverified output is not accepted. Define both criteria before testing.
| Event / date | Confirmed facts / unknowns | Scope paused | Investigator | Retest and resumption decision |
|---|---|---|---|---|
| [Reference] | [What happened; what is not established] | [Affected route] | [Owner] | [Evidence, authorized decision and status] |
An unauthorized attempt alone does not prove disclosure. Check evidence before claiming a leak, policy breach or successful containment.
Copy for the next review
Period / task / personal or team scope:
Prompt, Skill or workflow version; sources where relevant:
Attempts / accepted outputs / acceptance criteria:
Safety cases / expected behavior / actual behavior:
Baseline and AI work time, same measurement boundaries:
Setup time; review, correction and failed attempts:
Tool costs / estimates / unknown costs:
Measured, self-estimated or assumed / source:
Actual users, frequency and help needed:
Confirmed issues / unknowns:
Proposal or approved decision / approval status:
Next action / owner / due date / check:
At review, check comparability, quality, total effort, costs and exceptions before deciding to continue, change, pause or stop. Carry these records into the closing report; do not reconstruct results from memory.