An AI knowledge assistant for your documents and data.

02 / 07

How an AI assistant for company documents works

The assistant works in two steps: first I connect and index your sources, then for every question it looks up the relevant passages and builds its answer from those alone. It searches by meaning, so it finds the right paragraph even when it doesn't use the same words as your question. The technical term for this is RAG (retrieval-augmented generation). In practice, it's an internal AI chatbot built on your company data, for your own people only.

• Sources: documents, email, your CRM, shared spreadsheets or an internal wiki, connected through APIs, exports or direct database access
• Every answer links to the exact document and the passage within it, so you can check it
• If the answer isn't in your data, the assistant says so and doesn't fill the gap with a guess

03 / 07

Access permissions and company data security

The assistant answers each person only from the documents that person has access to. A salesperson can't use it to look at payroll, and warehouse staff won't see investor agreements. Before launch, we agree on which sources the assistant may read at all, and I either carry over the permissions from your existing systems or we set them up by role. Everything the assistant does is logged and traceable.

Through the AI agency Eturial, I wrote the security layer with Row Level Security rules for the custom CRM and client portal of DAU, a tax and accounting firm, and rebuilt the portal and dashboard. These rules decide, inside the database itself, who can see which record. I'll build the assistant on the same principle: documents a person isn't allowed to see never reach the language model when that person asks a question.

04 / 07

Where the assistant runs and what happens to your data

By default, I run the assistant on my own accounts (Vercel, Supabase) for a monthly fee, and a move to your own accounts can be arranged. Company documents almost always contain personal data, so we sign a data processing agreement under Article 28 of the GDPR before any documents are transferred, even for the prototype. In it, I act as the processor and list the sub-processors: the providers of hosting, the database and the AI models.

I call the language models through their paid business APIs, not through a consumer chat app. According to its documentation, OpenAI doesn't use data sent through its API to train its models unless you explicitly opt in, and Anthropic's commercial terms prohibit it from training models on customer content. Not training doesn't mean nothing is kept, though: OpenAI, for example, keeps abuse monitoring logs for up to 30 days by default, and they can include prompts and responses.

If your documents and their processing must stay in the EU, I'll put both the database and the app's server side in EU data centres, since Vercel runs server functions in the US by default. I then call Claude through Google Cloud or Amazon Bedrock, set up to run only in EU data centres, and do the same with the models for semantic search and meeting transcription, which Google Cloud also offers in the EU. OpenAI offers processing in Europe only to customers it approves.

05 / 07

Meeting transcription with summaries and tasks

From a meeting recording, the assistant produces a transcript, a short summary of what was agreed and a list of tasks with names and deadlines, if those were mentioned in the meeting. If the tool you use to track tasks has an API, I can send the tasks straight into it. The summary is saved alongside your other documents, so six months later you can ask the assistant why you made a particular decision. Before launch, we agree on which meetings get recorded and where the recordings are kept.

06 / 07

What to prepare before we start

For the first consultation, all you need to know is where your documents are and who is allowed to see what. If an old and a new version of the same procedure sit side by side, the assistant may quote the old one, so they need to be separated before the sources are connected. Before the prototype, I'll need:

• A list of the places where you keep documents, emails and meeting notes
• The questions that come up most often in your company
• An overview of which document versions are current and what is archived
• Access to the systems that need to be connected
• Someone who knows the access permissions and can test the answers

07 / 07

How the project runs and what it costs

The project runs in four steps, the same as the other custom AI systems and software I build for businesses. In a free consultation (30 to 60 minutes), we find out whether an assistant is worth it for your company. Then you get a quote split into parts, each with a specific price and delivery date. If you approve the quote or some of its parts, I build a free prototype using part of your documents, and you make the final decision only once you've seen it work. After you order, deployment takes six to twelve weeks, depending mainly on how many systems need to be connected, and ends with handover, documentation and training.

An assistant for your whole team counts as a system your company runs on every day. I split the quote into parts, so you can start with a smaller step from around €500. The price depends mainly on how many sources need to be connected and how many roles with different access rights you need. Monthly running costs for hosting and AI model calls come to tens to low hundreds of euros, depending on the number of documents and questions. After the first call, you get a specific price and delivery date.

You pay 20% when you order and the rest after handover. I always fix bugs in the system for free (see how we work together for the other terms).

If you want your team to know how to work with AI beyond the assistant as well, you can follow up with AI training on your real tasks.

Use the form below to tell me where your documents are and what people in your company ask about most often. I'll reply within eight hours. Neither the consultation nor the prototype commits you to an order.