"Can the AI answer using our documents?" It is one of the questions we hear most often. The short answer is yes. The long answer is that it works very well for some questions and poorly for others, and it pays to know which before you start.

How it works, without jargon

The assistant does not learn your documents by heart. When someone asks a question, the system searches the documents for the most relevant passages and gives them to the AI model together with the question. The model answers by reading those passages, and it can show which document the answer comes from.

This technique is called RAG (Retrieval-Augmented Generation). The practical advantage is big: when you update a document, the assistant uses the new version straight away, with no training involved.

Where it works well

  • Manuals and procedures. "How do I register a return?", "Which documents do I need for accreditation?". The answers are written down somewhere, but nobody wants to look for them.
  • Contracts and terms. "What is the notice period with this supplier?". The assistant finds the clause and quotes it.
  • Repetitive questions from customers or colleagues. The same ten questions, every week. The assistant answers at once, and a person steps in only for new cases.
  • Long documents. Summarising an 80-page tender or comparing two versions of a set of rules.

In Media Accreditation, our product for press accreditation, the assistant answers journalists and staff from the event documents: times, rules, the press room map. The questions are always the same, and the answers are already written.

Where it works poorly

  • Sums and figures across lots of data. "How much did we invoice Lombardy customers in March?" is not a question for documents: it is a question for your management system. That needs a direct connection to the data, for example through an MCP server.
  • Old or contradictory documents. If the folder holds three versions of the same price list, the assistant may pick the wrong one. Tidy up first, then connect.
  • Poor-quality scans. A crooked photo of a PDF is hard to read, for people and machines alike.
  • Guaranteed zero errors. A good assistant is rarely wrong and makes it clear when it is unsure, but it is not infallible. For important decisions, check the answer against the source it quotes.

Privacy first

Company documents often contain confidential data. Before starting, you need clear answers to three questions:

  1. Where do the documents go? For Media Accreditation we use Claude models on AWS Bedrock in a European region: the data stays in Europe.
  2. Are they used to train the model? With professional services the answer must be no, written in the contract.
  3. Who sees what? The assistant must respect permissions: someone who cannot open a document must not receive its content through the AI either.

How to start without wasting money

You don't need to connect the whole company. Our method is simple:

  1. Choose one set of documents with a clear purpose, for example the support procedures.
  2. Collect 20 real questions, the ones people actually ask.
  3. We test the assistant on those questions and measure how many answers are right and complete.

Within a few weeks you know whether it works and how much time it saves, with your own numbers rather than a demo.

If you have a folder of documents people keep asking about, tell us which one: we'll tell you whether AI can help and where to start.