PART 07

AI × HR & Hiring

AI can assist with low-risk hiring administration such as accessible job-description checks, candidate-authorized scheduling, and source-grounded onboarding drafts. People—not models—must assess candidates and make every employment decision.

No automated hiring, rejection, scoring, or candidate ranking

Accessible Administration & Human-Led Hiring

Start by mapping the lawful purpose, candidate notice, accommodations, retention, decision owners, and appeal route for each step.

BASELINE RISK

Unexamined hiring process

Role criteria are vague → sensitive applicant data spreads across tools → accessibility and accommodations are inconsistent → decisions, reasons, retention, and appeals are not documented.

CONTROLLED PILOT

Administrative assistant with a narrow boundary

Approve job criteria and candidate notice → limit AI to language, scheduling, and policy-grounded onboarding drafts → provide accessible alternatives → qualified people review all content and assess every candidate → record reasons and preserve appeal.

Evidence to collect

RIGHTS
Notice, choice, and appeal
Verify candidates understand where AI assists, can request an accessible human channel, correct their data, and challenge an outcome without penalty.
Artifact: candidate notice and appeal log
FAIRNESS
Test process and access
Review job language, scheduling access, accommodation handling, error patterns, and outcomes across relevant groups with qualified legal and HR oversight.
Artifact: scoped impact assessment
QUALITY
Source-ground every draft
Check onboarding content against current approved policy; measure unsupported statements, privacy leakage, accessibility defects, and reviewer corrections.
Artifact: versioned evaluation set
Vendor claims about speed, fit, or bias removal were removed. Hiring impact depends on the role, population, jurisdiction, process, access needs, and how humans use the output; human review alone does not cure a discriminatory system.

Reference architecture: administrative assistant only

Use tenant-isolated, least-privilege connectors and minimize applicant data. Keep sensitive and inferred traits out of prompts; validate schemas, encrypt records, enforce purpose-based access and retention, and log every use. The model cannot score, rank, reject, hire, set compensation, or provision access. Qualified HR owners approve communications and onboarding, with stop and rollback procedures.

Candidate-authorized schedulingApproved policy libraryHuman decision record
Before any pilot, obtain privacy, labor, accessibility, information-security, and qualified legal review for the applicable jurisdiction.

Pilot evidence package

  1. Jurisdiction-specific impact assessment, lawful-purpose record, candidate notice, accommodation path, retention schedule, and appeal procedure.
  2. Representative accessibility and fairness tests designed with qualified HR, legal, privacy, and affected-user input.
  3. Permission matrix proving the assistant cannot evaluate candidates or make employment, compensation, or access decisions.