AI assistant

How AI assistants use your company's documents safely

An AI assistant can answer questions from your own documents, but only if it is built carefully. Here is how retrieval works, where the risks are, and the questions to ask before you trust one.

BK-InfoTech4 min read

Article

Imagine asking a question like “What is our refund policy for orders over thirty days?” and getting a correct, short answer in seconds, with a link to the paragraph it came from. That is what a well-built AI assistant can do with your company's documents.

It is also easy to get wrong. This article explains how these assistants work in plain words, where the real risks are, and what a careful build does about them.

How the assistant finds an answer

Most business assistants use an approach called retrieval-augmented generation, or RAG. The name is technical but the idea is simple. It has two stages.

  1. Retrieve. When you ask a question, the system first searches your documents for the passages most likely to contain the answer. This works like a very good search engine that understands meaning, not just keywords.
  2. Generate. It then gives only those passages to a language model and asks it to write a clear answer based on them.

The important part is the order. The model is not answering from whatever it remembers from the internet. It is answering from your material, and it can point to it.

Where the risks are

Used carelessly, an assistant can cause real problems. These are the ones we plan for.

Made-up answers. A language model that has nothing relevant to read may still produce a confident reply. This is the best-known risk.

Showing people things they should not see. If the assistant can read every document, it may quote an HR file to someone who should never see it.

Your data leaving the building. Some assistants send every question and document passage to a public AI service. For confidential or regulated information, that may not be acceptable.

Out-of-date answers. If last year's policy is still in the index, the assistant will quote it.

Untrusted text. Documents and web pages can contain instructions aimed at the AI (“ignore your rules”). A careless system may follow them.

What a careful build does about each one

Answers that show their sources

Every answer should point to the passage it came from, so a person can check it. Good assistants also say “I don't know” when the documents do not contain the answer. An assistant that admits its limits is more useful than one that never does.

Access rules that follow the person

The assistant should only search documents the person asking is allowed to read. That means connecting it to your existing permissions, not building a second, separate set.

Private and open-source models

If your data must stay in-house, the assistant can run on open-source language models on your own servers, for example through a tool such as Ollama. In a private setup, documents and questions never leave your control. This is how we built Taxfolio, our tax research agent: a tax assistant that shows the filing behind every answer, and runs on local open-source models so information stays private.

Fresh documents

Documents change, so the index must too. A good setup re-reads updated files on a schedule and removes withdrawn ones.

Treating documents as data, not commands

A careful system keeps its own instructions separate from the text it retrieves, and limits what the assistant is able to do. An assistant that can only read and answer has far less to go wrong than one that can also send emails or change records.

Testing with real questions

Before launch, the assistant should be tested with real questions from the people who will use it, and each answer checked. After launch, reviewing what people ask shows where your documents have gaps.

A human fallback

For customer-facing assistants, hard questions should go to a person, with the whole conversation attached.

Questions to ask before you trust one

Whether you build or buy, these questions separate the careful from the careless:

  1. Where do my documents and questions go? Which servers, which companies?
  2. Does it show the source of every answer?
  3. What does it say when it does not know?
  4. How does it decide who can see which document?
  5. How are documents kept up to date?
  6. What can it do besides answer? Can it take actions?
  7. Can I see a log of what was asked and answered?
  8. How was it tested before I got it?

If a vendor cannot answer these plainly, keep looking.

Where to go next

If you would like to see how this could work for your own documents, read about our AI assistant service, look at the Taxfolio case study, or talk to us.

Share
All articles
  • Agentic ERP4 min read

    What is an agentic ERP?

    An agentic ERP has AI agents built in. They watch your business data, spot what needs attention and draft the next step, and you approve every action. Here is what that means in plain words.

    Read the article
  • ERP3 min read

    Signs your business has outgrown spreadsheets

    Spreadsheets are a great start. Here are seven honest signs that yours have stopped keeping up, and what to do next without a big-bang project.

    Read the article

Questions?

Let's talk about your business

Tell us how you work today. We’ll tell you, plainly, what we would build and what it would take.