How to build an internal AI assistant that answers your team's questions

It's eleven in the morning and someone types in the team chat: "How do I file a travel expense?" Nobody answers right away, so they ask someone else, and then someone else. An internal AI assistant is meant to answer that question in ten seconds, without bothering anyone.

The questions that come back every week

Every company has its own repertoire: how to request holiday, where the budget template lives, which policy applies to a discount, who gets told when an order goes wrong. The answers exist. They're in a PDF, a shared folder, an email from two years ago or, most often, in the head of whoever has been around the longest.

So the veterans end up acting as a human search engine, and newcomers take months to learn who to ask. A McKinsey Global Institute study estimated back in 2012 that knowledge workers spend nearly 20% of the week looking for internal information or tracking down a colleague who can help (Chui et al., 2012). More than a decade later, anyone working in a mid-sized company knows the feeling.

(Chui, M., Manyika, J., Bughin, J., Dobbs, R., Roxburgh, C., Sarrazin, H., Sands, G., & Westergren, M. (2012). The social economy: Unlocking value and productivity through social technologies. McKinsey Global Institute.)

How an assistant that knows your company works

This isn't a generic chatbot you can ask anything. It works only with what you decide to give it: manuals, policies, procedures, FAQs, templates. When someone asks a question, it searches those documents, finds the relevant passage and answers with that information, saying where it came from.

That last detail matters a lot. A good assistant doesn't make things up. If the answer isn't in your documents, it says so and points to the right person. That's the gap between personalised AI built for your company and one that answers confidently about things it doesn't know.

Where to start without overcomplicating it

You don't need to load the whole company on day one. Pick a small area, such as HR policies or the purchasing process, and gather the ten or fifteen documents that generate the most questions. Ask the team which doubts come up most often and test the assistant with those real questions.

Two things are worth sorting out early. First, that the documents are up to date, because the assistant will answer with whatever it finds, even if it's outdated. Second, that someone is responsible for keeping them current, which should already be the case for any important document.

What the team gains

The person who just joined stops depending on other people's goodwill to understand how things work. The operations lead stops answering the same question three times a day. And knowledge stops living in the heads of two people.

There's a less obvious benefit too. When you see which questions keep coming back, you find out which processes are poorly explained, or badly designed altogether. The assistant shows you where internal communication breaks down.

At NappAI, this kind of assistant is built with AI agents for companies that work from your own documents and your business context, with no code. If a process changes, you update the document and the answer changes with it.

Your team already knows what it needs to ask. It just needs someone to answer the first time.