Custom ChatGPT for your business, on your own documents

In most companies of twenty to two hundred people, knowledge is scattered. We see the same pattern almost everywhere:

• Five versions of the same procedureIn a shared folder, in last year's email, and in the head of the colleague who has been there ten years. Whoever searches finds the wrong one.
• New staff ask everything twiceThey do not know where it is, so they ask whoever knows. That colleague loses half an hour a day to it.
• Old quotes are a memory nobody consultsWhat did we charge last year for a similar job? It is somewhere, but nobody finds it fast enough.
• People paste into a public toolPutting a contract or a client file into a free chatbot means that, on the personal plans, your input can train models until you switch it off. What each plan does with your data is in ChatGPT pricing.
• Everyone sees everything, or nobody doesA payroll file in a folder called "General" is found by every assistant that reads that folder. Permissions are the quiet trap of any AI on company documents.

An assistant of your own solves this by doing three things together: it searches the latest version, shows the passage it relies on, and reads only what the person asking may read.

This is what an assistant on your own documents does reliably today, department by department.

In sectors with many files, see how this plays out in AI for law firms, AI for accountants and AI for HR. To show the same knowledge to the outside world, an AI chatbot for your website is this assistant's sibling; for your clients' emails there is AI customer service, and an assistant that also carries out tasks is an AI agent. For an assistant inside a platform of your own, see custom software.

Take a company where twenty staff look things up every day in procedures, quotes and product sheets. The figures are assumptions; run them again with your own team.

The sum: 20 times 10 minutes times 220 working days is 44,000 minutes, or 733 hours. At an assumed staff cost of 45 euros an hour that is 33,000 euros. The model: 3,520 questions times 4,000 tokens is 14 million input tokens at 2 dollars per million, plus 1 million output tokens at 10 dollars, together 38.72 dollars a month. On Claude Haiku 5.5 (0.10 and 0.50 dollars) it is about 2 dollars a month. Run your own case in the AI cost calculator.

For comparison, licences for the same twenty people: ChatGPT Business, listed at 21 euros per seat per month in Belgium, comes to 5,040 euros a year; Microsoft 365 Copilot Business at 15.60 euros per user to 3,744 euros a year on top of Microsoft 365. Those licences give your team a broad assistant; your own assistant is a narrow one that knows your documents and permissions. Many companies use both.

In this example an assistant of your own may cost up to about 31,000 euros to pay back within one year, before counting how much faster new staff find their way.

Your own ChatGPT does not always have to be custom built. This is how the three routes compare, at Belgian list prices; prices in your country can differ.

An assistant inside ChatGPT or Copilot is quick to set up. It runs out where your documents get large, where each department has different permissions, or where answers must come from your own software. Which language model suits which work is in ChatGPT vs Claude.

An assistant on your documents sees everything you give it. So we build it around three rules, which apply to any company with staff or clients in the EU.

Permissions per role. The assistant reads only what the person asking may open. It skips a document without access, even if the answer is in it. Clean up your folders before you start: a payroll file in a general folder is the first thing an assistant finds.

The GDPR. When your documents hold personal data, as staff files and client contracts do, every service that processes them is a processor. Article 28 requires a contract that sets out what, for how long and why. We keep documents and the search index in an EU environment. The language model runs through Amazon Bedrock with processing kept inside the EU; the model's maker does not see your prompts and answers there, and Bedrock lets you set per region whether anything is retained.

The EU AI Act. Since 2 February 2025, Article 4 requires that staff who use AI understand its limits: a short briefing at the start and an AI policy usually cover it. Since 2 August 2026, people talking to AI must know it (Article 50); for an assistant labelled as such on your intranet, that is obvious. If you use it to decide about employees, such as in hiring or performance reviews, that falls under the high-risk rules from 2 December 2027; the new timeline is explained in EU AI Act high-risk delay. Which duties your use case carries, see the EU AI Act checker.

For AMOSI, our own AI software that went live in July 2026, we built the reading of what a company supplies, with the source kept for every fact: the core of an assistant of your own. For Business Centrum, a payroll bureau in Ghent, we built a management platform with encrypted data, two-factor sign-in, a log of who did what, and an import that read 971 employees and 309 employers from their real spreadsheet in the trial. That is how we treat your documents too.

What you get from us:

• A trial on a real stack of documents, not a demo.
• Permissions per role that follow your existing accounts.
• The passage with every answer, and an honest "I don't know".
• Data in the EU, under a processing agreement.

The full offer is on AI for business; why a team in Belgium can be the right call is on nearshore software development. Show us the folder your team searches most: book a call.