Sheet 07 · ART-08
Human-oversight plan
Turns 'a human reviews it' into a named role performing a specific check at a defined frequency with the authority to override. Without those four elements, oversight is a claim rather than a control.
Artifact details
- Format
- Document
- Owner
- Business owner
- When it is used
- Required for structured and elevated pathways before pilot approval.
- Retention
- Life of the use case.
Contributors: AI Program Manager, Functional reviewers, The people doing the reviewing
Downloads
The blank template is a standalone document with every field and its guidance, ready to fill in or print. Printing uses this page’s print stylesheet, which drops the navigation and the download controls.
Required fields
The guidance matters more than the field name. A field labelled “risks” with no guidance gets filled in with “standard AI risks” and the artifact stops being worth reading.
| Field | Guidance | Required |
|---|---|---|
| What the tool produces | Specifically, and what it must never produce. | Required |
| Who reviews | A named role, not 'the team'. | Required |
| What they check | The specific failure modes they are looking for. | Required |
| How often | Every output, a defined sample, or on a schedule — with the reasoning. | Required |
| Authority to override and stop | Who can stop this without asking permission. | Required |
| What is recorded | So the review can be shown to have happened. | Required |
| Escalation | What the reviewer does when something is wrong. | Required |
Worked example
Filled in against the sample record set, so the template can be judged on what a real entry looks like rather than on its headings.
PIL-004 — Human-oversight plan
- The tool produces: interview scheduling proposals and structured note templates. It must never produce a score, ranking, comparison, or recommendation about a candidate.
- Recruiter reviews and confirms every interview note before it enters the candidate record. Checking for: content that characterizes a candidate rather than recording what was said.
- Hiring manager makes every advance-or-decline decision, with no tool-generated recommendation present in the interface.
Frequency, authority, and record
- Every output reviewed — no sampling, because the volume is manageable and the consequence is not.
- Monthly: People & Culture and Risk compare interview slate composition against the applicant pool.
- Grace Lin, People & Culture Director, may stop the pilot immediately without committee approval.
- Recorded: confirmation on each note, monthly composition comparison, and any instance of evaluative content, which triggers an immediate stop.
Blank template
The same fields with nothing in them. Print this page, or use the download above to get a standalone file.
What the tool produces
Specifically, and what it must never produce.
Who reviews
A named role, not 'the team'.
What they check
The specific failure modes they are looking for.
How often
Every output, a defined sample, or on a schedule — with the reasoning.
Authority to override and stop
Who can stop this without asking permission.
What is recorded
So the review can be shown to have happened.
Escalation
What the reviewer does when something is wrong.