Policy and Handbook Assistant

---

The problem

An organisation with a handbook and no search over it tends to have the same gap. Staff repeatedly ask HR about carryover rules, expense limits, and notice periods. The answers are already written down. The documents exist, are current, and nobody reads them, because finding a clause in a forty page PDF is slower than messaging a colleague.

The cost is not the individual question. It is that the HR or operations lead becomes a lookup service, and the answers they give informally start to diverge from the documents they gave them from.

Why retrieval fits this

The corpus is small, stable, and authoritative. Policies change quarterly at most, they are already written in prose, and there is exactly one correct answer to each question. That combination is what retrieval is good at, and it is rarer than it sounds.

Where this approach is the wrong answer: anything requiring calculation against a person's own record. "How many days do I have left" is not a policy question, it is a data question, and answering it from documents would be guessing. That belongs in a Power App over a leave table.

Architecture

The agent grounds on a SharePoint document library, retrieves the relevant passage, and answers with a citation back to the source document. Scoping the knowledge source to the library rather than the site keeps site scaffolding out of retrieval.

The constraint that drives the design

The model already knows the general answer, and that is the danger.

Ask about parental leave and a general purpose model can produce a confident, plausible, statutorily accurate figure that is not this employer's policy. An employee acting on it has been misinformed by a system that looked authoritative.

So the agent is constrained to answer only from the corpus. When a question has no grounded source, it says so, names the team that owns the answer, and stops. A correct refusal is the single most important behaviour in the build. It makes the assistant safe to put in front of staff, and it is the thing most demos never show.

Delivery considerations

Permissions must be honoured per user. The agent authenticates with Microsoft so SharePoint's own access rules apply. An assistant that reads a document the asker cannot open is a data exposure with a friendly interface.

Citations are not decoration. They let a sceptical employee verify the answer, and they let the policy owner see which documents are actually being consulted.

Document hygiene becomes the maintenance task. The agent's accuracy is a function of the corpus. Superseded versions left in the library will be retrieved and quoted, so archival discipline matters more after deployment than before.

Coverage gaps are findable. Reviewing questions that hit the refusal path shows which policies people need and the handbook does not cover.