Genius answers questions about your organisation's documents, in conversation, citing the documents it used.
This chapter is shorter than you might expect, and it points backwards a lot. That is the honest shape of the subject: almost nothing that makes Genius correct is configured on the Genius screen.
Ask it something
Open Genius in the left menu and ask a question about the cases you have filed:
Which motor claims are waiting for a driving licence?
What did the repair estimate say on claim 2024-0184?
Which supplier certificates expire in the next sixty days?
The answer comes back with citations — the documents it drew on. Click one and you are in the case.
An addendum is cited by its full number — Dossier #123/620 — and a question about Dossier 123 covers its addenda too, each one only if you may open it.
Read the citations, not just the answer. They are the part you can verify, and the habit of checking them is the difference between a useful assistant and a confident one.
What is actually happening
It helps to be plain about this, because the marketing language around AI assistants is not.
An assistant answering from your documents retrieves some passages and writes fluent prose about them. The writing is the easy part and it is very good. Everything that decides whether the answer is right happened before the model was involved:
| What decides the answer | Where you configured it |
|---|---|
| Which documents exist, and whether they are current | The process, and the dossiers made from it — chapters 3 and 4 |
| Which passages retrieval considers | The process's Genius step |
| Which values the model may read | Include in RAG context, on each field |
| Which values it may never read | Is sensitive, on each field |
| Who may see what | Permissions — chapter 8 |
The three ways an answer goes wrong
Two of the three are not fixed on this screen.
1. It retrieved the wrong passage
Semantic search finds text that resembles the question, and resemblance is not relevance. A question about a cancellation clause finds the paragraph that talks about cancellation most fluently, which is not necessarily the one that governs.
The fix is narrower retrieval, not a better model. Search inside the case, the process, or the document type the question belongs to. This is the failure you tune on the Genius screen.
2. The reader was not allowed to see it
This is not a quality problem. It is an incident.
Permissions are applied when the passages are selected, not when the answer is displayed. Filtering the output is too late — the model has already read the document and will paraphrase it.
What makes this dangerous is how quiet it is: the answer looks helpful, there is no error, no warning, and nothing in the transcript says a boundary was crossed. It can run for a long time before anybody notices.
If a field must never reach the assistant, mark it sensitive on the process's Fields step. That is the switch, and it is not on this screen.
3. The document was no longer current
A superseded policy. An expired certificate. A renegotiated contract. Still well written, still on topic, still the best semantic match.
This is where answer quality stops being an AI problem and becomes a governance one — and it is why the Requires expiration date switch in chapter 3 was worth turning on. An assistant reading a document store with no notion of currency will cite last year's terms with complete confidence.
Several assistants, not one
Genius is not a single assistant. You configure as many as you need, each with its own scope:
| Setting | What it does |
|---|---|
| Name, description, icon, colour | How it appears in the list |
| System prompt | The instructions it works under |
| Model | Which model answers |
| Permission scope | Whose permissions retrieval runs as |
| Organisation scope | Which parts of the organisation it covers |
| Confidentiality rule | Which confidentiality levels it may reach |
| Retrieval strategy | How candidate passages are gathered |
| Allowed sources | Which sources are in scope |
| Allowed processes | Which processes it may answer about |
The one you get by default is Total search — everything you are allowed to see, unfiltered.
Narrower assistants give better answers. An assistant restricted to claims is answering a much easier question than one searching everything, and creating two or three scoped assistants is the single most effective thing you can do to improve answer quality. It beats prompt tinkering, and it beats changing model.
Start with: one for claims, one for compliance, one for procurement. See how much better they are than Total search on their own subject.
What you should have now
- At least one answer with citations you followed back to a document.
- At least one narrower assistant, and a comparison against Total search.
- Every field that must never be read by AI marked sensitive on its process.