How do we get AI to answer from our own policies?

Give AI your approved documents and a short set of rules. Start with the policies, procedures, bylaws, and guides your staff already rely on. Then decide what AI should answer, how answers should look, when to say which document an answer came from, and when to pass a question to a person.

Staff already bring their own documents

In GovAI, half of all messages start with staff pasting in their own documents. That tells you what staff want: answers grounded in their organization’s own material.

Pasting works for one question. It doesn’t work for the twentieth person asking the same question about leave, or a new hire who doesn’t know which policy is current. Setting AI up to answer from a shared, approved set of documents solves both.

Source: GovAI messages from US and Canadian local governments, to August 2026.

Why the document behind the answer counts

On its own, AI answers from general knowledge. That’s fine for rewording an email. For “How many vacation days do new staff get?” or “What’s the setback for a garage in a residential zone?” the answer has to come from your documents, and staff need to see which one.

In local government, a wrong answer from general knowledge can be worse than no answer. A resident told the wrong setback builds the wrong garage. An employee told the wrong leave rule files a grievance.

Two parts: the documents and the instructions

• The documents are what AI reads: policies, procedures, bylaws, standards, and guides.
• The instructions are how AI behaves: what it helps with, how it answers, what tone and format it uses, and what rules it follows.

The documents give AI the facts. The instructions shape how it uses them.

A zoning help desk, start to finish

A planning department gets the same twenty questions every week, by phone and email: setbacks, fence heights, home businesses, accessory dwelling units.

1. Gather the documents. The current zoning bylaw, the fence bylaw, and the department’s one-page guides. Nothing older, nothing draft.
2. Write the instructions. Answer only from these documents. Name the section. Use plain language for residents. If the question involves a specific property or a variance, say a planner will follow up.
3. Test it. Run the twenty real questions, plus five it should hand off. Check every answer against the bylaw.
4. Fix and rerun. If it answers a fence question from the old bylaw, the old version is still in the documents. Remove it and run the test again.
5. Put it to work. Counter staff use it to answer callers, with the section number to quote. A planner reviews a sample each month.

The planners still own the hard questions. The twenty easy ones stop eating their mornings.

Where to start in your organization

Pick one set of questions that staff or residents ask again and again:

• HR: leave, benefits, and policy questions from staff
• IT: help desk questions about passwords, remote access, and new staff setup
• Planning: zoning and bylaw questions
• Clerk’s office: questions about meeting procedures or board packets

Gather only the documents that already answer those questions, to start.

Set a few rules for the answers

• Answer only from the approved documents.
• Say which document and section the answer came from.
• Say so when the documents don’t cover the question.
• Pass anything sensitive, unusual, or high-stakes to a person.

Keep the documents current

An answer is only as good as the document behind it. Name someone to own each set of documents, remove anything out of date, and add new versions when policies change. When you delete a document in GovAI, it’s removed from what AI searches. When a policy changes, rerun a few test questions to make sure the answers changed too.

Decide who can see what

Not every document belongs in front of every staff member. Before you add anything, decide which staff or departments should see each set of documents. HR investigation procedures and draft budgets usually stay with the teams that own them.

Try this week

Ask one department for the ten questions it answers most often, and the documents it uses to answer them. If the documents disagree with each other, you’ve found your first fix, before AI is involved at all.

Once it’s answering, test it properly: How do we test whether an AI Agent is doing the job well?