An internal AI assistant for staff questions that stays accurate
Short answer: An internal AI assistant is only as good as the documents behind it. Pick one narrow job (how we do things, product details, policies), gather the current versions into one controlled place, make sure the assistant can only see what each staff member is allowed to see, and require it to cite the document it used. Test it with real staff questions before launch, and give every source document an owner who updates it.
"Just ask the AI" sounds like the end of interrupting the operations manager twenty times a day. It can be, but only if the assistant answers from your current procedures rather than an old PDF, a draft someone abandoned, or its own guesswork. The setup work is mostly about documents and permissions, not prompts.
Here's a fictional example from start to finish, then the setup steps in general form.
The worked example: a plumbing supplier
Hypothetical business: a 12-person plumbing supplies business in Geelong with a trade counter, a small warehouse and two delivery drivers. New counter staff take months to learn which fittings are compatible, how trade accounts work and what to do when a delivery is short.
The problem
The same questions come up every day:
- "Which compression fitting goes with this pipe brand?"
- "Can a trade customer pick up on account without a PO number?"
- "What's the process when a delivery arrives with items missing?"
- "How do I process a return on a special-order item?"
The answers exist, scattered across a shared drive, a laminated sheet at the counter, two long email threads and the operations manager's head.
Step 1: Narrow the job
The owner decides the assistant will answer three kinds of question only: product compatibility (from supplier spec sheets), counter procedures, and warehouse procedures. Pricing, staff matters and anything customer-specific are out of scope for the first version.
Step 2: Gather and clean the sources
The operations manager spends two afternoons on this, and it's the most valuable part of the whole project:
- Exports the current supplier compatibility charts as PDFs.
- Writes up the five most common counter procedures as one-page documents with numbered steps.
- Deletes or archives old versions so only one copy of each procedure exists.
- Puts a "Last reviewed" date and an owner's name at the top of each document.
Step 3: Choose where the assistant lives
The business already uses Microsoft 365, so Microsoft 365 Copilot is one option. Microsoft's privacy documentation states that Copilot only surfaces organisational data the individual user has at least view permission for, and stresses using SharePoint's permission models so the right people have access to the right content. That means the folder structure and sharing settings matter as much as the documents themselves.
An alternative would be a shared ChatGPT project. OpenAI's help page describes projects as keeping related chats, files and instructions together, with uploaded files used as context, project instructions that override personal custom instructions, and sharing subject to plan and workspace settings. It also notes that in a shared project, deleting a file removes it for everyone.
Either works for a small team. The deciding factors are where the documents already live and how carefully access can be controlled.
Step 4: Write the instructions
The assistant's instructions say, in plain words:
- Answer only from the provided documents.
- Name the document and section used for every answer.
- If the documents don't cover it, say so and suggest asking the operations manager.
- Never guess part numbers, prices or stock levels.
Step 5: Test before launch
The team collects 30 real questions over a week (written on a sheet at the counter), including a few tricky ones. They run them through the assistant and mark each answer as correct, partly correct, wrong, or "correctly said it didn't know". Wrong answers almost always trace back to a missing or ambiguous document, which gets fixed.
Step 6: Launch small and keep it current
Two counter staff use it for a month. Every time an answer is wrong, they flag it, and the document owner updates the source. Procedures get reviewed every quarter, and the "Last reviewed" dates make stale ones obvious.
The setup steps, in general
- Pick one narrow job. "Our procedures" beats "everything about the company".
- Create a single source of truth. One current version of each document, in one place, with an owner and a review date.
- Write documents for retrieval. Short sections, clear headings, numbered steps, one topic per document. Tidy documents are easier for an AI tool to quote correctly and easier for people to check.
- Lock down access. Make sure the assistant can't surface HR files, payroll or customer data to people who shouldn't see them. In tools that respect existing permissions, fix the permissions first.
- Require citations. An answer that names its source can be checked in seconds. An answer that doesn't is a rumour.
- Give it an "I don't know" path. Tell it to say when the documents don't cover something, and to point to a person.
- Test with real questions. Collect them from staff, not from your imagination.
- Assign owners and a review cycle. Every document has a name next to it. Quarterly review is a sensible default for procedures.
- Measure. Track questions asked, flagged wrong answers, and which topics come up that have no document yet. That last list is your documentation to-do list.
Common mistakes
- Uploading the whole shared drive. Old drafts and duplicates guarantee contradictory answers.
- No owner for sources. The assistant stays confidently out of date.
- Ignoring permissions because "we're a small team". The day you hire someone new, it matters.
- Skipping citations to make answers look cleaner.
- Treating it as finished. Procedures change; the assistant only knows what you tell it.
FAQ
Which tool should we use?
Usually the one that sits where your documents already are. A Microsoft 365 business might start with Copilot; a team working mostly in shared chats and files might start with a shared AI project. Check each provider's current plan details before deciding.
Will the assistant learn from staff questions automatically?
Don't rely on it. The reliable way to improve answers is to fix or add the source documents.
How do we stop it showing sensitive files?
Keep sensitive files out of the assistant's sources entirely, and in tools that follow your existing sharing permissions, check those permissions before switching it on.
How long does setup take?
For a narrow first version, the bulk of the effort is tidying documents. The tool setup itself is usually the quick part.
If your team's know-how lives in people's heads and old PDFs, tell us which questions you'd want answered first.