AI-Drafted Reviews
Evidence-gathered AI drafts from goals, prior reviews, and work signals — with a hard human-accountability gate so the AI never submits.
Why It Matters
The blank page is why reviews are late, and recall bias is why they’re unfair. Kit’s AI drafter attacks both: it gathers real evidence — the reviewee’s goals and check-ins, their previous evaluation, their role expectations, their own notes, and (opt-in) their actual shipped work — and produces a structured draft answering the cycle’s questions. What it will not do, by design, is submit.
The model is AI drafts, humans decide. That’s not a slogan — it’s enforced in the submit path, logged in the evidence, and disclosed to the reviewee. Under GDPR Article 22, a decision with legal or similarly significant effect on a person may not be based solely on automated processing, and courts have read “solely” strictly: a human rubber-stamping machine output doesn’t count as human involvement. Kit’s gates are built so your review process never crosses that line.
Important
The human reviewer is accountable for every submitted review. An AI-assisted review cannot be submitted until the reviewer has materially edited the draft or explicitly checked “I have reviewed this AI draft and take ownership of it as my own assessment.” Both paths are timestamped into the review’s metadata. The AI never submits, never rates anyone on its own authority, and never bypasses the reviewer — if you submit it, you own it.
Generating a Draft
Open any review in your queue and press Draft with AI. Generation is synchronous — a few seconds later the form is filled in and ready to edit. The drafter gathers:
| Evidence | Source |
|---|---|
| Goals + check-ins | The reviewee’s goals: title, status, progress, latest note |
| Previous evaluation | The manager summary from the reviewee’s last finalized evidence record |
| Role expectations | The participant’s role summary for this cycle |
| The reviewee’s own notes | The Your own notes field from their self-review — tool-invisible work like mentoring and incidents |
| Work signals (opt-in) | Merged MRs and review activity fetched live from connected sources — only if you tick the per-draft checkbox |
The output fills the cycle’s question answers, a proposed overall rating (only if it fits the template’s scale), and a summary written as themes plus representative contributions. The review is flagged AI-assisted from that moment, and the flag follows it into the evidence record.
The prompt is narrative-only by construction: the drafter cites representative work, credits AI-direction and AI-enablement contributions, and is forbidden from producing volume metrics, rankings, or cross-person comparisons.
The Accountability Trail
Everything about the draft is recorded in the review’s ai_draft_metadata — encrypted, and carried into the frozen evidence:
- Generation — when it was drafted, who requested it, and a verbatim copy of what the AI produced.
-
Material edit — every save compares your content against the AI’s verbatim draft; the first real change stamps
edited_at. -
Explicit ownership — submitting an unedited draft requires the ownership checkbox, which stamps
owned_atand your name. - Signal disclosure — if work signals were used, the exact items (titles, links, dates — never stats) are stored and shown to the reviewee before finalization: “These items were fetched at draft time and are visible to the reviewee.”
The finalized evaluation record marks each review ai_assisted: true/false, and the evidence PDF prints it — full transparency to reviewee and auditor alike.
| Attempted action | What Kit does |
|---|---|
| Submit an untouched AI draft, checkbox unticked | Refused: “must be edited or explicitly owned before submitting — an unreviewed AI draft can’t be evidence” |
| Submit after material edits | Allowed; edited_at in the trail |
| Submit unedited with the ownership box ticked | Allowed; owned_at + name in the trail |
| AI submitting on its own | Impossible — no code path exists |
Credits and Failure Modes
Drafting runs on your account’s AI credits, metered like every other AI feature (see Integrations > AI Settings) — one LLM completion per draft. If credits are exhausted you get a dedicated message (“AI credits are exhausted for this account — write the review by hand or top up”) rather than a silent failure. If the model or a signal source is unreachable, the draft fails cleanly or degrades — a dead GitLab becomes a note in the draft context, never a broken form — and you can always write by hand.
Quick Checklist
- Ask reviewees to tend goals and Your own notes before you draft — evidence in, quality out
- Press Draft with AI; tick the work-signals box only with the reviewee’s awareness
- Edit the draft — verify every claim against your own judgment
- If you genuinely change nothing, tick the ownership checkbox deliberately
- Check the signal disclosure panel shows what you expect before submitting
Next Steps
- Work Signal Sources — connect GitLab so drafts cite real shipped work
- SOC 2 Evidence and Exports — how the AI-assisted flag appears to an auditor