AI HR: People Operations Automation

Status: not built. HR roles exist in our internal organisation experiment; no people operations system has been built or run. Nothing here is a product.

The line runs through "about a person"#

People operations splits cleanly into work that is about policy and work that is about a person, and the two need entirely different treatment.

Policy questions — how much leave do I have, what is the expense limit, how does the referral scheme work, what is the notice period — are answered from documents. They are asked constantly, the answers are written down, and answering them consumes a large share of a small HR team's week. This is the strongest case for retrieval in any business function we have looked at.

Anything about a specific person — performance, discipline, promotion, redundancy selection, compensation — is a decision about someone's livelihood, frequently with legal consequence. It needs a human decision, a documented rationale, and a route to challenge it.

Most bad outcomes in this area come from a system built for the first case drifting into the second.

Where it genuinely helps#

Policy question answering. With one non-negotiable condition: retrieval filtered by the asking employee's permissions, at query time. An HR corpus contains documents that must not be visible to everyone, and a manager's guidance is not the same as an employee handbook.

Onboarding. New starters ask the same forty questions, and asking them repeatedly feels uncomfortable, so people guess instead. A system that answers at 21:00 on day two without anyone having to ask a human is genuinely valuable.

Drafting. Job descriptions, policy updates, communications. Drafted by a system, reviewed by a person.

Structuring performance evidence. Helping a manager write a review that cites specific examples rather than impressions. Note the framing: it improves the human's input, it does not produce the rating.

Summarising engagement feedback — with care. Aggregating free-text survey responses into themes is useful; the risk is re-identification in small teams, where a summary of four responses can identify the person who wrote one.

Scheduling and process chasing. Unglamorous and consistently the thing HR teams actually want.

The confidentiality trap#

HR retrieval is the case where permission-filtered retrieval matters most, and where it is most often got wrong.

The documents involved are legitimately different for different audiences: an employee handbook, a manager's guide, a policy under consultation, an individual's record, a live grievance. A system that indexes "the HR folder" and answers questions from it will, eventually, tell someone something about a colleague.

That is not an embarrassing bug. It is a data protection incident involving special-category information in many cases, and it must be treated as one.

The requirement: filter at query time by the asking person's entitlements, before ranking. Filtering after generation is too late — the model has already seen the content and the answer can leak it even when sources are hidden. See RAG.

Where we would not go#

Any automated decision about an individual. Performance ratings, promotion, discipline, selection for redundancy. Automated decisions with legal or similarly significant effects on people are specifically regulated in several regimes, and this is a case where the ethical line and the legal line coincide.

Sentiment scoring of employees. Inferring engagement or flight risk from communications is surveillance regardless of the label, and it damages the trust that people operations runs on.

Monitoring that employees do not know about. Whatever the local legal position, the discovery is corrosive and permanent.

Answering about a named colleague. Even where permissions technically allow it, a system that will discuss individuals is one prompt away from a problem.

Where we would start#

The onboarding assistant. Policy documents only, permission-filtered, sources shown on every answer, an explicit refusal when the answer is not in the corpus, and a visible route to a human.

It is high-volume, low-risk, immediately useful to the people who need it most, and its failure mode is "I do not know, ask HR" rather than a decision about somebody's career. It also builds the permission-filtering discipline before anything more sensitive is attempted.

What to measure#

Deflection rate — questions answered without a human — but paired with satisfaction, or you optimise for people giving up.

Refusal accuracy on questions the corpus cannot answer. A system that always answers is unmeasured, not confident.

Permission correctness, tested with accounts at different entitlement levels, continuously rather than once.

Escalation rate to a human, and whether that route is used. If nobody escalates, either it is excellent or the route is hidden.

FAQ#

Do you offer HR automation?#

No. This is a problem study.

Is an HR chatbot a good first AI project?#

For policy questions, it is one of the better ones — high volume, documents that already exist, a clear benefit. The condition is permission-filtered retrieval, which is more work than it appears and is the part most projects skip.

Can it handle questions about my own record?#

Technically possible; we would not start there. Individual data raises identity verification, data protection and accuracy requirements that a policy assistant does not, and it changes the consequence of every failure.

What about using AI in performance reviews?#

To help a manager assemble specific evidence and write clearly, yes. To generate the rating, no. The distinction matters legally as well as ethically — an automated decision with significant effect on someone is regulated territory in several jurisdictions.

Does employees knowing it is AI matter?#

Yes, and in the EU transparency obligations for AI interaction have applied since 2 August 2026. Beyond compliance: people share different information with a system than with a colleague, and they should be able to choose knowingly.

What is the most common mistake?#

Indexing everything in the HR folder. It is the fastest route to a working demo and to a confidentiality incident, and the two arrive in that order.

How do we handle special-category data?#

Deliberately, and preferably by keeping it out of the system entirely. Health information, union membership, disciplinary records and similar carry heightened obligations in most data protection regimes. Confirm your position with counsel before indexing anything of this kind — this is general information, not legal advice.

Related Articles#

See RAG for permission-filtered retrieval, cyber security for the access controls underneath, and AI Recruiter for the hiring side of the same boundary.

What else is coming for AI HR