Internal AI assistant: which documents it needs and how to keep it from making things up.

The idea is attractive: an assistant the team can ask “what's the requirement for X?” or “how do we reply when the client asks for Y?” and get a good answer, instantly, without interrupting the person who knows. The concern is legitimate too: what if it makes things up?

The difference between a useful assistant and a dangerous one isn't the AI model. It's what it reads, what it's allowed to say and who maintains it.

Which documents it needs

• One-page processes. How each thing is done, who decides, what information is needed.
• Decision criteria. When to approve, when to escalate, what is never promised. It's what the team consults most.
• Internal and external FAQs with their approved answers.
• Policies and terms: prices, deadlines, guarantees, what can be said to a client.
• Templates for replies, quotes and follow-up in the house tone.

What isn't written and approved doesn't go in. If the assistant needs an answer that doesn't exist, that's a signal of which document is missing — not an invitation to improvise.

The rules that keep it from inventing

• It answers only from the documents. If it can't find the answer, it says it doesn't have it and who to ask. That “I don't know” is a feature, not a failure.
• It cites where each answer came from. The team can verify in seconds and spot outdated documents.
• It doesn't decide or promise. It prepares the information; the person uses it. Especially on prices, deadlines and commitments.
• It logs unanswered questions. That list is next month's documentation plan.

Who maintains it

An internal assistant without an owner becomes obsolete in weeks. Someone needs to be responsible for reviewing unanswered questions, updating documents when a criterion changes and retiring what no longer applies. It's a small, recurring task, not a project.

When it's worth it

When the same questions repeat, the answers are already approved and someone can maintain them. If there are no documents yet, the first step isn't the assistant: it's capturing the knowledge. If there are, the assistant is the cheapest way for the whole team to use them without reading them in full.

A trustworthy internal assistant is boring on purpose: it answers what the company approved, cites the source and stays quiet when it doesn't know.

The first step

For two weeks, note the questions the team asks the person who “knows”. The ten most frequent, with their approved answer, are the assistant's first document — and probably the one that frees the most capacity.