What is an AI assistant on your own documents?

What is an AI assistant on your own documents?

An AI assistant on your own documents is a tool you ask questions in plain language. The answer comes from your own files: manuals, work instructions, procedures, quotes or contracts. Below the answer it shows which document it comes from, so you can check it.

The difference from an ordinary search function is that you don’t need to know the exact search terms. For example, you ask: “How do we process a return from a business customer?” The assistant looks up the relevant passages and summarises the answer.

A short explanation of the term is under AI assistant. In this article you can read how such an assistant works, what standard software packages can already do and what to watch out for with your data.

What can you do with it?

You can use it to find knowledge that is currently in folders, mailboxes or people’s heads. A few made-up examples:

• A new employee asks how a particular machine is cleaned, and gets the steps from the work instruction.
• Someone in customer service looks up the conditions for a repair, with a link to the right document.
• A salesperson looks up how a similar job was described in an earlier quote.
• A service engineer searches a hundred-page manual for the error code he sees on the screen.
• A manager asks what the staff handbook says about requesting leave.

Each time it is about questions whose answer is already written down somewhere. The assistant doesn’t make up new policy. It helps you find what is already there.

How does an AI assistant on your own documents work?

The assistant searches first, and only then writes an answer. This approach is called RAG, short for retrieval-augmented generation. In plain language it works in four steps:

1. Preparing the documents. Your documents are split into small pieces and made searchable. This happens in advance, and again when a document changes.
2. Searching. When you ask a question, the system looks for the pieces that best match your question. It usually searches by meaning, not just by literal words.
3. Answering. A language model receives your question and those pieces, with the instruction to answer using only them.
4. Showing the source. Below the answer it shows which document and which part it comes from.

So the language model does not learn from your documents. It reads the relevant pieces again for every question. That means you can add or replace documents without retraining the model.

Is the answer correct, then?

The answer is based on your own documents, and you see the source. That helps, but it is not error-free. An AI assistant can go wrong in three ways:

• It finds the wrong piece. For example, an old version of a procedure.
• The document itself is no longer correct. The assistant doesn’t know that a way of working changed last month if nobody has updated the document.
• The language model misreads it or adds something that isn’t there. That is called a hallucination.

That is why the source reference is so important. Is it about something where a lot is at stake, such as a price, a safety rule or a contract? Then check the source. And make sure every document has an owner who keeps it up to date.

Is this the same as your own ChatGPT environment?

It looks similar, but there is a difference in administration. In ChatGPT, Copilot and Claude you can upload files in a conversation or a project. Then you ask questions about those files. For one person with a few documents, that works fine.

For a team it gets harder. Who makes sure everyone uses the same, latest version? Who may see which documents? And what happens when someone leaves who had the files in their own account? An AI assistant for your business arranges that centrally: one collection of documents, managed by your business, with permissions per person or group.

Can your office suite already do this?

More and more often, yes. Do you already work in Microsoft 365 or Google Workspace, or do you use a business AI plan? Then first look at what is already included. An overview, as of 7 October 2026:

• Microsoft Copilot (until recently Microsoft 365 Copilot). Copilot Chat is included with eligible Microsoft 365 licences and works mainly with the web and with files you add yourself. With a paid Copilot licence, Copilot can answer based on your mail, meetings, chats and files in OneDrive and SharePoint. It only shows what you are already allowed to see.
• Gemini Notebook from Google (NotebookLM until July 2026). You add sources and ask questions about them. It is a core service in a number of Google Workspace business editions. According to Google, your files there are not used to train the model, unless you give feedback.
• ChatGPT Business and Enterprise. With the company knowledge feature, ChatGPT searches connected apps, such as SharePoint and Google Drive. It respects the existing permissions and gives source references with the answer.
• Claude. With the Microsoft 365 integration, Claude searches SharePoint, OneDrive, Outlook and Teams. Here too: Claude only sees what you can already see. With a Team or Enterprise plan, an administrator turns the integration on.

Does a standard feature like this do what you need? Then that is often the simplest choice, and you don’t need to have anything built. Which of these packages suits you, you can read in ChatGPT, Copilot or Claude: which AI suits your business?

When is a custom solution better?

Having your own AI assistant built is better if your documents or your questions don’t fit into one standard package. For example, in these cases:

• Your documents are in different places. Partly in SharePoint, partly on your own server, partly in industry-specific software or a database.
• The answer has to appear in your own system. For example in your CRM, in a customer portal or in your service engineers’ app.
• You want fixed rules. Use only approved documents, show the source with every answer, and refer to a colleague when in doubt.
• Not everyone has a licence. For example, employees on the shop floor or in field service.
• You want to link an action to it. Such as a draft reply to a customer question, with the information from your documents. Then you are moving towards an AI agent.

Where is your data stored, and who can access it?

With an AI assistant, your documents pass through an AI service. That is why you ask three questions in advance.

Where are the documents and the searchable copy stored? With your own office suite, with an AI provider or with the builder of the assistant? And in which country?

Who can access them? An assistant may only give an employee answers from documents that employee is allowed to open themselves. Otherwise, for example, salary information leaks through a clever question. Microsoft and Anthropic write that their integrations follow the existing permissions. With a custom solution, you arrange that yourself.

What agreements do you have? If your documents contain personal data, the GDPR applies. If another party processes that data for you, a data processing agreement is mandatory. Is it about sensitive data, such as medical information? Then check whether you first need to carry out a DPIA.

In addition, use business accounts, not free accounts. With business services, the terms state whether your data is used to train models. Microsoft, for example, writes that prompts, answers and files in Copilot are not used to train the foundation models.

Which documents are suitable?

Suitable documents are ones that are correct, finished and not floating around in ten versions. Think of:

• Work instructions and manuals.
• Procedures and the quality manual.
• Product information and technical specifications.
• Standard texts, terms and conditions and approved quotes.

Less suitable are drafts, old versions and documents full of personal data that nobody needs for their work. Clean those up first or leave them out. The better your collection, the better the answers.

How do you start small?

Start with one collection of documents and one group of users. This is how you go about it:

1. Choose one collection. For example, the workshop’s work instructions, or the customer service procedures.
2. Appoint an owner who keeps the documents up to date and removes old versions.
3. Write down twenty test questions whose correct answer you know. That way you see what goes well and what doesn’t.
4. Have a person check along. In the first weeks, someone checks the answers and reports what isn’t right.
5. Then expand with a second collection or a second group.

What does an AI assistant on your own documents cost?

That depends on the route. With a standard feature you pay the provider, usually per user. Sometimes a basic version is already included in your current licence, such as Copilot Chat with eligible Microsoft 365 licences. Check the current price with the provider; it changes often.

If you have the assistant built, the price depends on the number of sources involved, how well organised your documents are and whether not everyone is allowed to see everything. With me, you start with a business analysis. In it I look at where your documents are, who needs them and whether a standard feature is enough. After that you get a proposal with a fixed price. It also states what is added per month for hosting, subscriptions and maintenance. That way you know in advance where you stand. Also read how your price is determined.

How do I help you with an AI assistant?

I first check whether a standard feature in your current package is enough. I receive no commission from suppliers. If a custom solution is needed, I build the assistant myself, on the documents and systems you already have, with the source with every answer. After that I explain to someone in your team how it works and how to add documents. Read more about AI in your business, or schedule a conversation.

Frequently asked questions

Sources: Microsoft Learn: Overview of Microsoft Copilot Chat, Microsoft Learn: Data, privacy and security for Microsoft Copilot (in Dutch), Google: NotebookLM is now Gemini Notebook, Google Help: Gemini Notebook with a work or school account (in Dutch), OpenAI Help: Company knowledge in ChatGPT, Claude Help: Microsoft 365 integration, Dutch Data Protection Authority (Autoriteit Persoonsgegevens): Data processing agreement (in Dutch) and Dutch Data Protection Authority (Autoriteit Persoonsgegevens): AI transparency requirements (in Dutch). Consulted on 7 October 2026.