Support people; do not score their worth
Workforce data reflects unequal opportunities, incomplete context, and human judgment. AI must not make consequential employment decisions.
A model must not determine or propose any person’s hiring, rejection, schedule, pay, promotion, discipline, monitoring, or dismissal. Do not infer emotion, honesty, disability, health, protected traits, or “culture fit” from messages, video, voice, biometric, or behavioral data. Employment decisions require qualified HR and legal review under the laws that apply.
Any production implementation additionally requires approved redacted inputs, rights and lawful basis, data minimization, least privilege, tenant separation, managed secrets, audit logs, representative, adversarial, and regression evaluations, proposal-only behavior, human approval, monitoring, a stop condition, an incident path, and tested rollback.
Safer assistance is narrow and procedural
Appropriate executive questions
- Are goals clear, job-related, achievable, and supported by the resources people actually received?
- Does a team-level pattern remain after checking workload, territory, tenure, leave, channel mix, and data quality?
- Could a metric reward harmful behavior or disadvantage a group through a proxy?
- Who can inspect, correct, contest, or appeal the information and the eventual human decision?
- What would cause the analysis to stop, be withdrawn, and be investigated?
A fictional support team sees a lower synthetic resolution rate in one queue. The system does not rank agents. Reviewers find that the queue receives more complex issues and an outdated knowledge base. The proposal is to improve routing and guidance, then re-measure at team level. No individual employment action is generated.
A safe analysis prompt
Exercise: audit one people metric
- Choose a team-level metric and document what it was designed to measure.
- Map collection purpose, provenance, access, retention, and relevant legal or policy requirements.
- List confounders, proxy variables, affected groups, and accessibility needs.
- Ask HR, legal, managers, and affected representatives to challenge the interpretation.
- Design a process-support experiment, not an individual decision, with appeal and rollback.
People-impact gate
Use the NIST AI Risk Management Framework as a general risk structure. It does not determine compliance. Obtain jurisdiction-specific employment, privacy, labor, and discrimination advice before any real use.