Making company policy answerable with generative AI
Generative AI turns an HR policy library into something employees can query directly, answering in plain language and citing the document it came from. The gain is twofold: people get correct answers immediately, and the HR team stops relaying the same twenty questions every week.
A significant part of any HR team's week is spent answering questions that are already answered in writing. How much leave carries over. What the notice period is for this grade. Whether this expense is claimable. The policy exists, it is correct, and nobody can find it.
This is a search problem wearing the costume of a service problem, and it is unusually well suited to generative AI because the source material is text, the questions are repetitive, and the correct answer is verifiable.
Why the intranet did not solve this
Most organisations have already tried. There is a policy portal, the documents are on it, and people still email HR.
The reason is that a search box matches words while a person asks about a situation. Somebody wants to know whether they can carry leave into the next year after a period of sick leave. The policy covers it, in two separate clauses, in a document called "Absence and Leave Framework v4". Keyword search returns the document. The employee wanted the answer.
Closing that gap is the whole value. An answer in a sentence, with the clause it came from attached, ends the interaction. A document does not.
Citation is not optional
The single design decision that determines whether one of these systems is trusted is whether every answer names its source.
Without citation, the system is asking people to take its word on questions with real consequences, and the first confident error destroys the trust permanently. With citation, an answer is a shortcut to the relevant clause rather than a replacement for it. The employee can check, the HR business partner can check, and a wrong answer is caught as a documentation problem rather than experienced as a betrayal.
It also changes what the system is for. A cited answer is a navigation aid. That is a modest claim and it is one the technology can actually meet, which is why generative AI for human resources is most defensible when framed that way.
Drawing the line at individual records
Policy questions and personal questions feel similar to the person asking and are completely different systems behind the scenes.
"What is the parental leave entitlement" is a document question. "How much leave do I have left" is a record question, and it requires authenticated access to that individual's data. "How much leave does my team have left" requires a manager relationship. "How much does she earn" should return nothing regardless of who asks, unless the asker holds a role that permits it.
These distinctions have to be enforced where data is retrieved. A model instructed to decline politely is a request, not a control. If the retrieval layer will not return the record, the answer cannot contain it, and that is the only arrangement that survives contact with a determined employee.
Where it should refuse
Some categories should route to a person by design, not because the technology cannot produce text about them, but because the interaction matters more than the information.
Disciplinary processes, grievances, anything with a legal dimension, and anything touching an individual's health belong with a human being. A system that answers a grievance question correctly and coldly has still done harm. Configuring these routes at the outset is a small piece of work that prevents a category of incident that is very hard to recover from.
The workforce planning use
Separately from policy, the same interface over headcount, skills and attrition data changes how planning conversations run.
Questions about where the organisation is thin on a particular skill, how attrition varies by team and tenure, or how the current hiring plan compares with the shape of the workforce today are all assemblable from data HR already holds. They are asked infrequently because assembling them is a project, which means planning conversations often run on impressions.
The caution here is that small numbers are identifying. An answer about attrition in a team of four is a statement about individuals, whatever the intent. Aggregation thresholds should be set before anyone asks, not after.
Getting the document set right
The quality ceiling is set by the documents, and most policy libraries are worse than their owners believe. Superseded versions circulate as attachments. Two documents contradict each other on the same point. A policy references an appendix that nobody can locate.
Preparing for this kind of deployment surfaces all of that, which is uncomfortable and useful. The practical sequence is to establish one authoritative location, retire what is superseded, and resolve the contradictions before turning anything on. A system that answers from a contradictory corpus will answer inconsistently, and the inconsistency will be blamed on the AI rather than on the library.
A sensible pilot
Take the twenty questions HR answers most often. They are usually already known, and if not, a fortnight of triage will produce them. Confirm which document answers each one, fix what is wrong or missing, and start there.
Measure the reduction in repeat questions rather than usage. Usage tells you people tried it. Fewer emails tells you it worked.
The tone question
One thing that gets overlooked in these deployments is that HR communication has a register, and a generic assistant does not have it.
Policy answers should read the way the organisation writes to its people. If the handbook is warm and plain, an answer that comes back in clipped legalese is jarring and subtly damaging, because the employee reads it as the company's voice. This is worth reviewing explicitly with whoever owns internal communications, rather than accepting whatever register the system defaults to.
There is a related point about certainty. Where a policy genuinely does depend on circumstances, the answer should say so and route the person onwards rather than choosing an interpretation. An HR system that is confidently wrong about entitlement creates a commitment somebody then has to withdraw, which is worse than the original question going unanswered for a day.
To discuss this against your own policy set, schedule a demo.
Frequently asked questions
Is it safe to let staff ask HR questions of an AI system?
It is safe when answers are drawn from approved documents and cite them, and when anything touching individual records is gated by the same permissions as the underlying system. It is not safe if the model answers from general knowledge.
What happens when a policy is updated?
Answers should be generated from the current document at the time of asking, so an update takes effect immediately. This is a reason to keep policies in one managed location rather than distributed as attachments.
Will it give a wrong answer on something important?
Any system can. The mitigation is citation: every answer points at the clause it came from, so the person can check, and sensitive categories such as disciplinary or legal matters route to a human instead.
Does this help with workforce planning too?
Yes, though it is a separate use. Planning questions draw on headcount, skills and attrition data rather than policy documents, and they need the same care about permissions.