An AI assistant that knows your own documents: how it works

ChatGPT knows a lot, but nothing about your company. It doesn't know your procedures, your price agreements, your product documentation or your customer files. Yet that's where the real time savings are for most teams: quickly finding the right answer in their own information.

A custom AI assistant, also called an AI co-pilot, does exactly that. In this article we explain how it works and what to watch for.

What does such an assistant do?

Your team asks a question in plain language, in Teams, Slack or a web app. The assistant looks for the answer in your own documents and gives a short answer, with a link to the source. Some examples:

• "What's our return procedure for a B2B customer?"
• "Which warranty terms apply to product X?"
• "Summarise the latest contract with customer Y."
• "How do I request parental leave?"

A new employee no longer has to interrupt three colleagues. And an experienced one finds in seconds what used to take fifteen minutes of searching.

How does it work technically?

The principle is called retrieval-augmented generation, or RAG. It sounds complex, but the idea is simple:

1. Your documents are made searchable. Manuals, procedures, contracts and knowledge base articles are indexed, so the assistant can search by meaning, not just by keywords.
2. When a question comes in, the assistant first finds the relevant passages.
3. A language model writes the answer, based only on those passages, and cites the source.

So the language model is not trained on your data. Your documents stay in your own environment and are only consulted when someone asks a question.

Why not just use ChatGPT?

• It doesn't know your company. Without access to your documents it stays generic.
• Permissions. A custom assistant respects who may see what. Someone in sales doesn't get answers from HR files.
• Sources. Every answer links to the document it came from, so you can check it.
• Privacy. You decide where the data lives and which agreements apply with the AI provider. More on that in our article on AI and GDPR.

What should you watch for?

Good sources are the foundation. An assistant is only as good as the documents it relies on. Outdated procedures give outdated answers. So start with the sources that are already up to date.

Start with one team or one topic. For example, support with the product documentation, or HR with the staff handbook. That way you quickly see what works.

Make it honest. A good assistant says "I don't know" when the answer isn't in the sources, instead of making something up.

Integrate where your team already works. One more tool that nobody opens delivers nothing. In Teams or Slack it does get used.

How long does it take?

We usually build an AI co-pilot with a limited number of data sources, integrated into Slack, Teams or the web, in 4 to 6 weeks. After a first version, we refine it based on the questions your team actually asks.

Want to know which other processes in your company are a good fit? Read which processes an SME can automate today, or tell us what your team searches for most.