How to build an assistant that knows your own policies

Who this is for: governance leads, membership managers, anyone who fields "where's the policy on…" questions. What you'll need: Microsoft 365 with Copilot Studio, and your documents in SharePoint. How long: an afternoon for a working agent, longer to tidy what it exposes. What you'll have: an assistant your team can ask in Teams, answering from your own documents with sources attached.

The question this answers

Every membership organisation runs on a body of written knowledge nobody can navigate. Minutes going back a decade. Policies with three versions and no clear current one. Guidance notes, terms of reference, the paper that explained why the subscription structure changed.

When somebody needs an answer they ask a colleague who was there. When that colleague leaves, the answer leaves too. That's the institutional memory problem, and it's usually described as unsolvable.

It isn't. The documents exist. What's missing is a way to ask them a question.

Step one: pick a narrow library, not everything

The instinct is to point it at the whole tenant. Don't. An agent that has read everything gives worse answers than one that has read the right thing, because it has more chances to surface something superseded or irrelevant.

1. Choose one SharePoint site or document library to start with. Governance papers, or membership policy, or committee minutes. Pick one of them.
2. Open that library and look at what's actually in it. You are checking for the two things that will spoil answers: superseded versions sitting alongside current ones, and scanned PDFs that are images rather than text.
3. Move or archive the superseded versions now. This is the tedious part and it's the part that determines whether the agent is any good.
4. Note the site URL. You'll paste it in a moment.

A scanned PDF with no text layer is invisible to the agent. It will not say so, it will simply answer as though that document doesn't exist. If a key policy is a scan, run it through OCR before you start or you'll get confidently incomplete answers.

Step two: create the agent

1. Go to copilotstudio.microsoft.com and sign in with your work account.
2. Choose Create, then New agent.
3. Give it a name people will recognise in a menu. Governance Assistant, not Agent v2.
4. Describe what it's for in a sentence. This shapes its behaviour more than people expect, so be specific: "Answers questions about our governance policies and Council decisions using our own documents."
5. Skip the fancy options on the first pass. You want something working before you tune it.

Step three: give it your documents

1. In the agent, find Knowledge and add a source.
2. Choose SharePoint and paste the site or library URL you noted earlier.
3. Save, and wait. Indexing is not instant. Give it time before judging the answers. If it says it can't find anything in the first ten minutes, that's usually indexing rather than failure.
4. Add a second source only once the first is working well.

The important thing here, and the reason this is safe to do: the agent answers each person under their own permissions. Somebody who cannot open a document in SharePoint will not get its contents through the agent. You are not creating a bypass. You're creating a search that respects the walls you already have.

Step four: write the instructions, which is where the quality comes from

The default agent is a generic assistant. The instructions turn it into yours, and this is fifteen minutes that makes the difference between useful and annoying.

Paste into the agent's instructions and adapt. The fourth paragraph is the one that matters.

You answer questions about our governance and membership policies
using only the documents you have been given.

Always cite the document name and its date.

If more than one document covers the question, say so and give the
most recent, but mention the earlier one exists.

If the documents do not answer the question, say so plainly.
Do not fill the gap from general knowledge. Our rules are ours
and a plausible-sounding answer is worse than no answer.

Where a decision is recorded, say who took it and when.

Keep answers short. Offer to go deeper.

That instruction not to fill gaps from general knowledge is the whole game. Without it, asked about your conduct process, the agent will cheerfully describe a reasonable-sounding conduct process that is not yours.

Step five: test it with questions you already know the answer to

Do this before anyone else sees it, and be adversarial about it.

1. Ask five questions you know the answers to. Check the citations, not just the answers. A right answer citing the wrong document means it got lucky.
2. Ask something the documents genuinely don't cover. It should say so. If it invents something, strengthen the instruction and test again.
3. Ask about something that changed. This is the one that catches superseded documents still sitting in the library.
4. Ask a question in the sloppy way a colleague would actually type it, not the careful way you'd phrase a search.

Step six: publish it where people already are

An assistant on a website nobody visits is an assistant nobody uses. Publish it to Teams, where your team already is, and it becomes something they ask on the way to a meeting.

Start with a small group: the people who field these questions today. They'll find the gaps fastest and they have the most to gain. Widen it once they trust it.

Tell people plainly what it does and doesn't know. "It has read our governance papers and current policies. It has not read minutes before 2019." Expectations set once prevent a lot of quiet disappointment.

What it will teach you about your own documents

Within a fortnight this will stop being a technology project and become a records project, because the agent surfaces exactly how disorganised the library is.

You'll find policies with no date. Three versions of the same terms of reference. A decision everybody refers to that was never actually written down anywhere. Documents whose filename is the only clue to what they contain.

None of that is caused by the agent. It's been true for years and the agent is simply the first thing to ask the library a direct question. Fixing it makes the agent better and makes your organisation better, which is a rare combination.

Do this yourself, or have it done

Building the agent is genuinely an afternoon, and the judgement calls are yours: which library, which documents are current, how to phrase the instructions. You know which policy supersedes which. We don't.

Where we help is the part that turns it from a good demo into something the organisation runs on: getting the document library into a state where answers are reliable, connecting it to your membership records so it can answer about members as well as policies, and making sure what it says matches what your CRM does. That's Maintain and Evolve. And if somebody has already built one of these quietly and the team has started depending on it, tell us rather than tidying it away. A working thing people rely on is the best specification anybody could write.