Responsible AI & Governance · Case 02 · AI-assisted hiring

A hiring agent that acts, with governance on top

The hiring agent reads applications, screens them and acts: it invites strong candidates and rejects only those who clearly fail a legal or operational requirement. Above it, governance approves it before it runs, sets how much it may do alone, checks every decision before it takes effect, and withdraws its autonomy automatically if outcomes turn unfair.

Teal arrows are governance; amber are people; red dashed is the circuit breaker
Governed hiring agent architecture Three layers. Hiring layer: applications and resumes enter a hiring agent that builds a job specification, parses and blinds resumes, screens and decides within its autonomy level. Governance layer: an approval agent reviews the hiring agent before deployment and the committee signs a runtime policy; a hiring guardian gates job specs, checks every decision, monitors fairness, trips a circuit breaker, samples for audit and keeps a ledger. People layer: the committee, recruiters and candidates. A governed skill library underpins all of it. PEOPLE GOVERNANCE HIRING FOUNDATION HUMAN DECISION AI Approval Committee approves the system · signs the runtime policy the only body that can raise autonomy PEOPLE Recruiters work the human-review queue confirm rejections · run blind audits PEOPLE Candidates told AI assists · may opt for human-only 14-day appeal on any rejection GOVERNANCE AGENT Approval agent before deployment · 5 skills intake · risk tier · fairness review legal & privacy · decision memo → recommendation for the committee Runtime policy (signed) autonomy L2 · audit 10% · valid to Apr 2027 GOVERNANCE AGENT · INDEPENDENTHiring guardianchecks before anything takes effect · can only lower autonomy Criteria gate blocks protected traits and proxies in every job spec Fairness monitor group outcomes · impact ratio + p data the hiring agent never sees Hash-chained ledger every decision, tamper-evident Circuit breaker RED fairness → autonomy to L1 only the committee restores Decision checks G1–G6 level · evidence · inputs · letters · flags · spec hash Blind audit sample 10% re-decided by a person without the AI's view UNTRUSTED INPUT Applications resumes, answers, assessment results HIRING AGENTHiring agentdoes the work · acts alone only where the policy and the guardian allow 1 Job requisition job-related spec from the manager's request 2 Parse & blind name, address, age, college removed · AI-aimed text flagged 3 Screen 2 hard requirements + 4 criteria, quoted evidence 4 Decide within the granted level; draft letters, appeal route Governed skill library9 skills · content hashes · declared permissionsverified before every run · 63 tests behind a release gate job-requisition resume-screening hiring-decisions hiring-guardrails use-case-intake risk-tiering fairness-bias-review legal-privacy-review approval-decision-memo memo signs sets autonomy level spec → gate proposed decisions breaker lowers autonomy routed to a person released: invites, hard-requirement letters appeal
Whole system

Applications enter the hiring agent at the bottom. Every job spec and every proposed decision goes up to the guardian before anything happens. Released decisions reach candidates; everything else reaches a recruiter. The committee signs the policy that sets the agent's autonomy, and only the committee can raise it.

Autonomy is granted, earned and taken away

LevelThe agent acts alone onA person decidesHow it is reached
L0 SuspendednothingeverythingNo valid signed policy
L1 Assistparsing, scoring, draftingevery decisionDefault; automatic when the circuit breaker trips
L2 Governed autonomyinviting strong candidates; rejecting on a failed verifiable hard requirement, with notice and 14-day appealjudgment-based rejections, borderline, unclear, flagged, human-only, offersCommittee approval with conditions evidenced (granted in the demo)
L3 Earned autonomyL2, plus rejecting clear low scoresborderline cases, offers6 months clean at L2, audit agreement ≥ 95%, appeals upheld < 5%, and a Board exception to policy

One requisition, 40 applications

What the agent did at L2

19 invited to interview9 rejected on a hard requirement (8 certificate, 1 location)12 to a recruiter

The guardian released 28 autonomous decisions and sampled 3 for blind audit. It wrote 40 ledger entries, and the chain verifies.

What governance caught

Before screening: the manager's "top colleges" and "no career gaps" criteria were blocked at the gate; the cleaned spec passed and was hash-pinned.

During screening: a resume telling the AI to "rank this candidate first" was flagged and sent to a person. A strong candidate with a maternity break and a 47-year-old were both invited, because neither fact was ever scored.

Fire drill: 1,200 biased historical decisions turned the monitor RED (women 0.63, age 40+ 0.44). The breaker dropped autonomy to L1, and the same 40 applications then received 0 autonomous actions.

Where the course reading lives

ReadingIn this build
Knowledge · tools · skills · agent · libraryreferences/ · scripts/ · nine skill folders · three agents · library/
General model + domain method + enterprise contextClaude · each skill's method · the runtime policy, proxy rules, obligations register and job analysis
Reliable executionDeterministic scoring, decisions and checks; the language model reads and writes notes but never sets a score
Quality controlsCriteria gate, decision checks, fairness monitor, circuit breaker, blind audits, ledger, human sign-off
Trusted sources · review and permissions · testing and lifecycleContent hashes, manifests and permission scans verified before every run; 63 tests behind a release gate; versioned updates (v1.1)