AI-assisted recruiting screening

Why this example

This is the hardest kind of request to handle well. Saying no protects the firm but leaves a real problem unsolved and teaches the requester that the program is an obstacle. The useful outcome is neither approval nor refusal — it is finding which part of the problem can be solved without putting a consequential decision inside a tool nobody can explain.

Walkthrough

1. Step 01 of 11

   The business problem

   Two recruiters, 900 applications, 40 openings, and a nine-day median response time.

   • Campus recruiting volume has grown 40% in two years with no change in recruiting headcount.
   • Median time from application to first recruiter contact is nine business days. Competing firms respond in three to five.
   • Recruiters estimate 11 hours a week each on interview scheduling logistics alone.
   • The requester's framing was 'we need AI to screen resumes.' The problem underneath it is capacity, and capacity is being consumed by two different activities — evaluation and coordination — that carry very different risk.

   Owner: Grace Lin, People & Culture Director

2. Step 02 of 11

   Intake submission

   Submitted through the standard form. Complete enough to triage on first pass.

   • Registered as UC-004 on 2026-03-04 with a named owner willing to own the outcome.
   • Data involved: applicant resumes, contact details, and interview notes. Classification: restricted.
   • Proposed capability as submitted: automated screening with a candidate fit score and a ranked shortlist.
   • Human review as proposed: recruiters review the ranked shortlist.
   • The intake form asked one question that shaped everything after it: does this affect decisions about individuals? The answer was yes.

   Owner: Requester, with the AI Program Manager

3. Step 03 of 11

   Triggered risk factors

   Three separate triggers, any one of which would have been enough.

   • Affects employment decisions — the tool would shape who gets interviewed.
   • Restricted data — applicant personal data belonging to people outside the firm.
   • Consequential ranking with limited explainability — reviewing a ranked list still inherits the ranking, because reviewers rarely reorder what a system has sorted.
   • The triage record shows all three with the answers that triggered them. The requester could see exactly why this was not going to move at the speed of a meeting-summary request.

   Owner: AI Program Manager, weekly triage

4. Step 04 of 11

   Required stakeholders

   People & Culture owns it. Risk concurs. Security reviews. The committee decides.

   • People & Culture — owns the recruiting process and any change to how candidates are evaluated.
   • Risk, Legal & Privacy — applicant data handling, retention, and employment-related exposure.
   • Information Security — vendor security posture and data flows.
   • Hiring managers from two practice groups — they live with whoever gets interviewed.
   • AI Governance Committee — approves or declines the pilot.
   • The program's job here was scheduling one review with all of them rather than four sequential ones, and preparing material good enough that the meeting could reach a decision.

   Owner: AI Program Manager coordinates; each function decides its own scope

5. Step 05 of 11

   Questions that had to be answered

   The review lived or died on evidence the vendor could produce.

   • Can the vendor produce adverse-impact testing for the scoring model? Answer: no.
   • Can the firm validate the model's effect on protected groups independently? Answer: not with current capability.
   • What is the applicant privacy notice, and does it cover automated processing? Answer: it did not.
   • How long is applicant data retained, and where? Answer: vendor default exceeded the firm's recruiting retention schedule.
   • If a candidate asks why they were not advanced, what can the firm say? Answer: nothing specific, which was the moment the room stopped debating.
   • Which part of the recruiters' lost time is evaluation and which is coordination? Answer: coordination is the larger share.

   Owner: Reviewers, with the program tracking answers and gaps

6. Step 06 of 11

   Pilot guardrails

   Screening declined. A reduced scope was approved that removed evaluation entirely.

   • No candidate scoring, ranking, filtering, or shortlist generation of any kind.
   • No resumes processed by the tool — scheduling data and structured interview notes only.
   • Every candidate advanced or declined by a named human reviewer, exactly as before.
   • 90-day limited pilot with 26 participants and a defined end date.
   • Immediate stop if scoring capability appears in a vendor release.
   • The reduced scope addressed the larger half of the capacity problem and removed the consequential decision from the tool. That is the outcome worth aiming for on requests like this.

   Owner: AI Governance Committee, decisions DEC-2026-04 and DEC-2026-05

7. Step 07 of 11

   Human-oversight requirements

   Written as named roles doing specific checks, not as a general instruction to review.

   • Recruiter reviews and confirms every interview note before it enters the candidate record.
   • Hiring manager makes every advance-or-decline decision, with no tool-generated recommendation present.
   • People & Culture and Risk review the monthly interview-slate composition against the applicant pool.
   • The People & Culture Director holds explicit authority to stop the pilot without committee approval.
   • 'Human in the loop' means nothing until you name the human, the check, the frequency, and the authority to override.

   Owner: Grace Lin, People & Culture Director

8. Step 08 of 11

   Success and harm measures

   Both were defined before the pilot, and the harm measures came first.

   • Benefit: median time to first contact, target under five business days from a nine-day baseline.
   • Benefit: recruiter hours on scheduling, target under six from eleven.
   • Harm: interview slate composition compared with the applicant pool, monthly, with no material divergence expected.
   • Harm: candidate experience rating at or above the 4.2 baseline.
   • Harm: any instance of the tool producing evaluative content, which triggers an immediate stop.
   • A pilot that measures only benefits will find only benefits.

   Owner: People & Culture with the program coordinating collection

9. Step 09 of 11

   The governance decision

   Two decisions on the same day: one declining, one approving something smaller.

   • DEC-2026-04: automated screening and ranking not approved. Reasoning recorded, including the three alternatives that were considered and rejected.
   • DEC-2026-05: reduced-scope coordination pilot approved with four conditions.
   • Both records name People & Culture as the decision owner, with Risk concurrence. The AI Program Manager prepared the material and wrote the record.
   • The requester left with something that solves most of her problem, and with a written reason for the part that was declined. That reason is what makes the next request easier rather than harder.

   Owner: Grace Lin with Risk concurrence; AI Governance Committee

10. Step 10 of 11

    Ongoing monitoring

    Monthly checks with named reviewers, and conditions tracked as open actions.

    • Monthly slate-composition review by People & Culture and Risk.
    • Biweekly pilot check-in covering scope adherence — specifically, whether anyone has asked for a ranking feature yet.
    • Vendor release notes monitored by IT & Security for scoring capability.
    • Two conditions tracked as open actions with owners and dates: the privacy notice update and completed training.

    Owner: AI Program Manager coordinates; owners named per condition

11. Step 11 of 11

    The next decision

    December 8, with all four dispositions genuinely available.

    • Scale: extend to experienced-hire recruiting if the coordination benefit holds and no harm indicators move.
    • Modify: adjust scope or oversight if the slate comparison shows anything unexpected.
    • Pause: if scope creep toward evaluation appears, or a condition lapses.
    • Retire: if the coordination benefit does not materialize, the tool goes away and the recruiters get a different kind of help.
    • Reconsidering screening would require vendor adverse-impact documentation and a defined validation approach. People & Culture owns that question, not the program.

    Owner: AI Governance Committee, 2026-12-08