Mastery checklist: responsible HR AI

A hiring manager wants the model to rank applicants by resemblance to the company’s ten highest-rated employees.

Response: Decline that design. Historical ratings and employee similarity are not reliable definitions of job capability. They may encode unequal opportunity, manager bias, tenure, educational privilege, or demographic proxies. Return to the role analysis, define observable criteria, and choose direct assessment. If AI assists, constrain it to a validated task such as locating source evidence, not predicting who resembles past insiders.

An application contains no example for one essential criterion. The model labels the candidate unqualified.

Response: Correct the label. “No evidence located in this application” does not establish lack of capability. Follow the approved assessment plan: the criterion may require clarification, interview evidence, a work sample, or a consistent non-progression rule based on clearly requested application information. A human verifies the document and records the reason.

A vendor offers video analysis that scores confidence and honesty from facial movement and voice.

Response: Do not assume these traits can be validly inferred or are job-related. The method creates scientific, disability, cultural, language, privacy, and discrimination concerns. Use structured questions and representative work evidence instead. Human approval of an emotion score does not repair an invalid signal.

The organization has wellbeing survey comments and wants to identify employees likely to resign.

Response: Existing access does not authorize a new purpose. The proposal involves sensitive free text, power imbalance, inferred personal states, and possible adverse treatment. Assess necessity, lawful basis, transparency, governance, and safer alternatives with qualified specialists. Aggregate organizational improvements may be more appropriate than individual prediction. Do not quietly link voluntary feedback to personnel decisions.

A recruiter sees an overall recommendation but not the underlying passages. They can override it by filing a support ticket.

Response: This is not meaningful review. Show the approved criteria, original source, extracted evidence, uncertainty, and limitations. Give the recruiter time, training, authority, and an easy way to disagree before any action. Require evidence-based reasons, escalation, and a rollback path. Remove unnecessary ranking that anchors judgment.

Use this checklist before piloting, deploying, or materially changing an HR AI workflow.

Purpose and necessity

• The legitimate HR outcome is specific and documented.
• The team considered a non-AI method and can explain why AI assistance is necessary or useful.
• The task is classified as drafting, retrieval, extraction, analysis, recommendation, decision, or action.
• Prohibited uses and outputs are explicit.
• Consequences, affected people, scale, reversibility, and power imbalance are understood.
• Applicable legal, contractual, policy, consultation, and employee-representation requirements are assigned to qualified owners.

Job relevance and assessment

• Role criteria describe observable work capabilities or lawful requirements.
• Essential criteria are separated from preferences and trainable skills.
• Criteria were approved before candidate evidence was reviewed.
• Every screening or interview element traces to a criterion.
• Direct work evidence is preferred over weak proxies.
• Nontraditional career paths and equivalent evidence are supported.
• “Not evidenced” remains distinct from “cannot perform.”
• Accommodations preserve the capability assessed without demanding identical interaction.

Data and privacy

• Inputs are limited to fields necessary for the stated purpose.
• Sensitive data and inappropriate inferences are excluded.
• Source, prompt, output, logs, integrations, exports, backups, and deletion are mapped.
• Provider training use, location, subprocessors, access, retention, and change controls are understood.
• Approved accounts, least privilege, security controls, and incident routes are active.
• Notices are accurate and understandable.
• Source-of-truth, correction, access, retention, and deletion responsibilities are named.
• A new purpose triggers a new review rather than automatic reuse.

Prompt and tool behavior

• Instructions define allowed evidence, output schema, uncertainty, and prohibitions.
• Candidate or employee content is treated as untrusted data.
• The model does not infer protected traits, personality, emotion, honesty, health, or motivation.
• The output avoids unsupported ranking and pseudo-precision.
• Exact passages or source locations support consequential claims.
• Missing or unreadable input produces a safe, visible state.
• Model, prompt, parser, and integration versions are recorded.

Quality and fairness

• Evaluation data represents actual roles, formats, languages, career paths, and difficult cases.
• Qualified reviewers establish an evidence-based reference set.
• Tests measure omissions, unsupported claims, unreadable inputs, and consequential errors.
• Performance and outcome patterns are examined across relevant groups where lawful and responsible.
• Small samples and uncertainty are reported honestly.
• Counterfactual tests are supplementary, not treated as proof.
• Human overrides, complaints, accommodation failures, and reversals are monitored.
• Pause thresholds and reevaluation triggers are defined.

Human judgment and recourse

• A named reviewer sees original evidence, criteria, uncertainty, and limitations.
• The reviewer has enough time, training, authority, and freedom to disagree.
• Interface defaults do not encourage bulk approval or anchoring.
• Reasons are tied to job evidence and recorded proportionately.
• Stop, escalation, fallback, correction, and rollback paths are tested.
• Consequential communication or system writes occur only after the gate.
• Affected people can seek accommodation, correct data, and challenge process through an appropriate route.
• Review of a challenge is not merely another run of the same model.

Operations and ownership

• Product, HR, hiring, privacy, security, accessibility, legal, procurement, and employee-relations responsibilities are clear.
• Vendor claims are verified with evidence and contractual controls.
• Staff know how to report an error without concealing it.
• Incidents include downstream correction and communication duties.
• A safe manual process works when the tool is unavailable.
• Monitoring continues after launch and after every material change.
• The workflow has a retirement condition and an owner who can stop it.

Do not ship because a demonstration looks efficient. Ship only when the organization can show representative evidence of task quality and fairness, a justified data flow, trained reviewers, useful recourse, safe failure behavior, and accountable ownership.

For low-risk drafting, controls may be lightweight but still real. For recruitment, performance, compensation, discipline, termination, health, monitoring, or other consequential contexts, use heightened scrutiny and specialist advice. Some use cases should be redesigned or declined. The ability to buy or build a system is not evidence that its use is legitimate.