Custom Copilots & Knowledge Assistants
Internal Chatbot Governance: Permission-Aware Retrieval and the Questions Legal and Risk Teams Must Ask Before Launch
· By AIHQ Team
The demo that always works — and the question it never answers
A department head opens the internal chatbot in a review session, types "what is our retrenchment benefits policy", and gets a clean, confident answer in four seconds. Everyone in the room relaxes. Then someone asks the question that should have come first: whose documents did it just read?
That is the moment internal chatbot governance actually begins. Not at the policy document stage, and not after launch when the first incident lands. It begins when a business manager realises the retrieval layer, not the chat window, is where the risk sits.
Most internal chatbots today are built on retrieval-augmented generation: the model searches an index of your documents, pulls the closest matches, and writes an answer from them. The model is rarely the problem. The problem is that the index was built by crawling a shared drive, and the permission model of the shared drive did not come with it.
Failure modes worth rehearsing before go-live
None of these are hypothetical. Each one is a predictable consequence of building retrieval for convenience rather than for entitlement.
- Inherited oversharing. A folder was set to "anyone in the company with the link" three years ago during a restructuring exercise. The chatbot inherits that setting faithfully and now answers compensation questions for anyone who asks.
- Metadata blindness. Documents carry a classification label in a footer, but the index only stores filename, owner and text. The model cannot filter on a field that was never extracted.
- Summarisation leakage. Empty search results do not stop a model. It may answer from adjacent documents, or from general knowledge, and present it in the same confident tone as a cited answer.
- Aggregation inference. No single document is confidential, but a sequence of three permitted answers reconstructs a salary band or a client's contract terms. This is the failure mode most permission models miss entirely.
- Stale entitlement. An employee moves department, or a contractor's engagement ends. The directory updates; the chatbot's cached permission snapshot does not.
Tool comparison: three retrieval patterns and what each one can guarantee
Before approving anything, know which pattern you are actually buying. The table below sets out what each does with permissions.
| Retrieval pattern | How permissions are applied | What it reliably prevents | Where it still fails |
|---|---|---|---|
| Naive RAG (index the drive, no filters) | None at retrieval time | Nothing structural | Inherited oversharing, stale access, all of the above |
| Pre-filtered RAG (ACL checked before search) | Permissions pulled from the document store, applied at query time | Retrieval of documents the user has no entitlement to | Aggregation inference; cached ACL drift |
| Document-level permission-aware RAG (ACL plus classification plus re-check on serve) | Entitlement and classification both evaluated per chunk, refreshed per session | Most direct leakage; supports classification-aware refusal | Policy design gaps; weak logging makes it hard to prove |
Off-the-shelf assistants vary widely in which of these they implement, and vendor documentation often describes the third while the deployment behaves like the first. Ask for a live test with a test account, not a slide.
The control questions to put to the build team
Treat this as a review checklist. The answers should be demonstrable, not described.
1. Document-level permissions
- Can you show me a query, run as a user without entitlement, where the system returns a refusal rather than a summary?
- Are permissions evaluated per chunk or per document? A single document can contain a public section and a restricted annex.
- Where does entitlement data come from, and how often is it refreshed? Same-session or nightly?
- What happens to a document that has no permission record at all — default deny or default allow?
2. Citation and audit trails
- Does every factual sentence carry a source reference the user can open, subject to their own permissions?
- Can an answer be reproduced later from the logged query, the retrieved chunk IDs and the model version?
- Is there a clear visual distinction between "answered from your documents" and "answered from general knowledge"?
3. Retention and logging
- How long are queries, retrieved chunks and generated answers retained, and who can read them?
- Do the logs themselves create a new access problem — for example, an admin console that shows every query in the company?
- Are logs exportable for an internal investigation without a fortnight of engineering time?
4. Confidentiality classification
- Which classification levels exist, and is each one mapped to a retrieval behaviour?
- Who can change a document's classification, and is that change logged?
- Are personal data categories treated differently from commercial confidential material? Malaysian organisations should confirm their own obligations under the Personal Data Protection Act 2010 with their legal advisers rather than relying on general guidance.
5. Escalation paths
- What does the chatbot do when confidence is low, sources conflict, or the question is about an individual's employment, pay or disciplinary matter?
- Is there a named human owner for each escalation route, with a response expectation?
- Can a user flag a wrong answer, and does that flag reach someone who can fix the index?
6. Human sign-off
- Which use cases require a human approver before the answer is released to the asker?
- Who signs off the chatbot at launch, and on what evidence?
- What is the review cadence after launch — quarterly, or after every flagged incident?
Who is affected, and where the load lands
Impact is uneven, and department heads should plan for the uneven parts.
- HR and People teams will field the most escalations, because policy questions and personal-data questions look identical to a chatbot. Expect a defined triage rota, not goodwill.
- Legal and compliance carry sign-off on classification and retention. Their real work is agreeing what "confidential" means in practice, document by document.
- IT and data owners own entitlement refresh, index rebuilds and the test evidence pack. If they cannot produce a demonstration on request, the control does not exist.
- Department heads own adoption: making sure their teams know which questions to ask the bot and which to route to a person.
- Line employees need one clear habit — check the citation before acting. That habit belongs in role-based training, not in an email announcement.
If you are budgeting for this, the governance and integration lines usually surprise finance more than the licence does. A line-item view is set out in this AI implementation cost in Malaysia budget model and the recurring run costs in this AI tool pricing comparison.
What good looks like in practice
A workable pattern, drawn from implementation work rather than theory:
- Classify before you index. Do not crawl a drive and classify afterwards. Start with the document sets you actually want answerable — HR policies, SOPs, product documentation — and leave everything else out of scope for version one.
- Default deny. Any document without a permission record is invisible to retrieval.
- Test with personas. Build five test accounts across levels and departments, then run a scripted set of twenty queries including deliberately out-of-scope ones. Keep the output as launch evidence.
- Publish the boundaries. Users should know what the assistant is for and what it will not answer.
- Review in the first month. Expect index gaps and misclassifications; fix them while usage is still low.
This sits alongside the wider controls in our AI agent governance for public sector and GLCs framework, which covers accountability owners and staged rollout gates. Where the assistant also takes actions — filing tickets, updating records — the review requirements tighten considerably.
Where this ends and a custom build begins
An off-the-shelf assistant that connects to your document store is a reasonable starting point for a bounded scope. It stops being reasonable when you need per-chunk permissions, classification-aware refusal, retention rules that match your records schedule, or an escalation route into an existing case system. At that point the question is no longer which tool to buy but what to build.
This is the territory of custom AI solutions — an internal copilot scoped to specific document sets and permission sources, with logging designed for audit rather than analytics. AIHQ has worked across corporate, public sector, professional and regulated environments, and the governance questions above tend to surface early in scoping rather than late in testing.
None of this makes an internal chatbot hard to justify. It makes it hard to launch casually — which is the correct outcome.
A five-line summary for the approval memo
- Retrieval, not the model, decides what leaks.
- Permissions must be evaluated per request, per user, from a refreshed source.
- Every answer needs a citation and a reproducible log entry.
- Classification and retention are policy decisions, not engineering settings.
- Someone senior signs off, and reviews again after month one.
For teams that want to work through the checklist with their own document sets and directory structure, we can run that review with you.
FAQ
What is permission-aware retrieval in an internal chatbot?
Permission-aware retrieval means the chatbot checks what the asking user is entitled to see before it searches or returns any document content. Rather than indexing everything and filtering afterwards, entitlement is applied at query time against a permission source that is kept in step with your directory and document store.
Is a pilot with twenty test queries enough evidence for go-live?
It is a reasonable minimum, not a guarantee. The useful practice is to script queries that deliberately cross department, seniority and classification boundaries, run them as several test personas, and keep the outputs as launch evidence. Plan a review in the first month, because index gaps and misclassifications show up under real usage.
How long should chatbot logs be retained?
There is no single answer that fits every organisation. Retention should follow your records schedule and your obligations, and should be agreed with legal and compliance rather than set by the engineering team. Malaysian organisations should confirm requirements under applicable law, including the Personal Data Protection Act 2010, with their own advisers.
Does an off-the-shelf assistant meet internal chatbot governance requirements?
Sometimes, for a narrow scope such as a single document set with simple permissions. It becomes harder when you need per-chunk entitlement checks, classification-aware refusal, retention matched to a records schedule, or escalation into an existing case system. Those are the usual signals to consider a custom build.
Who should own sign-off on an internal chatbot?
Assign a named business owner who is accountable for the use case, with legal, risk and IT contributing evidence rather than co-owning the decision. Without one accountable person, escalation routes and post-launch fixes tend to stall.