AI-Assisted Work in Your Self-Review: What You Can Prove

Describe AI-assisted work through a contribution you can support: what you delivered, what you decided, and what the available evidence says about the result. A quickly generated draft may still require substantial checking and correction. A useful experiment may also end with a decision to stop. This guide gives you a worksheet for auditing a review claim, a worked comparison that includes setup and review time, and wording for results you have not measured. Use your employer's review questions and approved AI practices as the starting point. Every worked case below is fictional; its wording illustrates an approach rather than a result you can borrow.

Choose a claim your evidence can carry

Start with the work your reviewer needs to understand. Identify the AI-assisted step, your own decisions, the checks performed, and the delivered result or current status. Follow your employer's rules for approved tools, disclosure, and handling work information. Keep restricted material in the systems your employer permits.

METR's July 2025 randomized study found that experienced open-source developers took longer on the studied tasks with AI access despite perceived speedups. Its February 2026 follow-up reported signs of improved productivity, but participant and task selection, plus difficulty timing concurrent work, prevented a reliable estimate of the current effect. Both reports concern particular developer settings. Use them as a reason to measure your own workflow.

For each effect, use one of three labels. Measured means you retained observations with a defined unit and period. Estimated means you can explain an approximation and its assumptions. Unknown means the record cannot support a conclusion yet. A measured difference between periods can still have an unknown cause.

A record of prompts or generated drafts establishes activity. Use the activity-to-impact distinction to decide what additional evidence your review sentence needs.

Copy this claim-audit worksheet

Complete these fields for one proposed claim and keep them with the supporting record. This original editorial worksheet is intended to work alongside your company's review form.

• Proposed claim: What changed, for whom, and during which period? Name the result separately from the tool used.
• Work boundary: Where does the task begin and end? Define the completed deliverable, quality checks, unit of effort, and costs included.
• Contributions: Which step used AI? What did I decide, verify, or correct? What did colleagues contribute?
• Effects: Label time, quality, output, and downstream outcomes individually as measured, estimated, or unknown. Record the method behind each label.
• Comparison: What is the baseline? Are task difficulty, volume, acceptance criteria, and periods comparable? What else changed?
• Supporting record: Where are the approved work samples, timing notes, correction history, or feedback? Preserve context and permitted source access.
• Allowed wording and next check: What can I say now? Which qualification belongs in the review, and what evidence would let me revise it?

Apply the labels separately. You might have measured completion time while customer benefit remains unknown. A reliable time log also needs a sensible comparison: an easier project or a more experienced reviewer can change the result. Preserve those qualifications in the sentence your manager will read.

A worked time comparison that includes the whole task

Fictional worked case; all figures in this section are illustrative. An employee prepares weekly documentation-change digests for Support. The comparison covers active team effort from an approved change list to an accepted digest. Both periods use the same approver and require source checks, compatibility warnings, and corrections before acceptance. Waiting time is tracked separately from active effort.

• Illustrative baseline: four September digests required 240 minutes of team effort, including drafting, checking, approval, and correction.
• Illustrative AI-assisted period: four October digests required 72 minutes for input preparation and drafting, 92 for checking and approval, and 36 for correction. Total recurring effort: 200 minutes.
• Illustrative setup: choosing the approach and preparing the reusable instructions required another 80 minutes. October's total pilot effort was therefore 280 minutes.

In this fictional comparison, recurring effort was 40 minutes lower across the four October digests. Including setup, the pilot required 40 minutes more than the baseline. Both observations belong in the record. Counting only the faster draft stage would miss much of the work needed to produce an accepted digest.

Quality and comparability still matter in this fictional case. The digests met the same acceptance checklist after correction, but the months contained different changes. Record the observed effort difference and leave its cause unresolved. Later customer comprehension and missed issues remain unmeasured.

Illustrative review wording for this fictional case: I introduced an AI-assisted drafting step for Support digests and retained source checking and approval. Four October digests took 200 minutes of recurring team effort, compared with 240 for four September digests. Setup added 80 minutes, so the pilot has not recovered its setup effort. I will check whether the recurring difference persists on comparable work.

For your own comparison, include colleagues' checking time and corrections discovered after handoff. Count setup once, and state whether you report it separately or allocate it across completed uses. Keep active effort distinct from elapsed turnaround when people work in parallel. Avoid projecting an observed difference across a year before you know the workflow will continue.

Write honestly when savings are unmeasured

Fictional example with unknown time savings: I used an approved assistant to propose themes from approved feedback excerpts. I checked the themes against the source material, removed unsupported interpretations, and revised the summary with the research lead. The team used the final summary to choose follow-up questions. We did not record comparable preparation times, so I cannot quantify an efficiency improvement.

This fictional wording describes a completed contribution and its observed use. It leaves the missing time comparison visible. If your baseline is reconstructed from memory, label it an estimate and explain its basis. A document timestamp shows when a file existed; it does not necessarily measure the effort spent producing it.

When the useful result is qualitative, describe the accomplishment without inventing a metric. A concrete decision or accepted work sample can support a narrower claim.

Include a stopped pilot and shared contributions accurately

Fictional stopped-pilot wording: I tested an approved assistant for drafting internal onboarding guides. Review found missing prerequisites and references to retired steps. I stopped the pilot before rollout, documented the failure cases, and kept the existing drafting process. The trial helped define acceptance checks for any later attempt; it did not establish a productivity gain.

Explain a stopped pilot in proportion to its importance. Retain the acceptance criteria, what failed, and the decision taken. Avoid presenting every rejected output as a major achievement. The useful contribution may be a well-supported recommendation to stop a particular approach.

Fictional shared-work wording: Our group completed the service-migration handover. I used an approved assistant to organize source notes, checked each proposed instruction, and coordinated revisions. The platform engineer validated rollback conditions, and the Support lead checked the handover against the team's needs. The final document reflects that shared work. We did not isolate the tool's effect on delivery time.

Keep the qualification when you shorten the review

Read the finished paragraph against the worksheet. Preserve the period, the people involved, and any limitation that changes the meaning of the claim. A pilot result should remain a pilot result when copied into your summary. If a number cannot survive that shortening honestly, explain the contribution in words and keep the calculation available for discussion.

Save the work, checks, and unresolved questions together while you can still explain the comparison.

Choose an AI-assisted contribution because it helps explain your work this review period. Complete the claim worksheet before polishing the sentence. Keep measured observations separate from estimates, and leave missing effects unknown. Include setup, checking, correction, and colleagues' contributions when those belong inside your comparison. A carefully bounded account can explain a delivered result, an unresolved question, or a pilot you stopped. Keep the supporting record so your next review starts with the facts you actually have.