Internal AI assistant over manufacturing procedures in Berlin 2026: cost and security

The verdict in three sentences

When a site with 350 employees manages 2,000 procedures spread across a document management system, shared drives and binders, operators lose time and sometimes apply an outdated version. An internal RAG-based AI assistant, answering in natural language and citing the source procedure, costs roughly 25,000 to 50,000 EUR excl. tax then 300 to 1,200 EUR a month in usage. The two success conditions are per-department access rights and EU hosting, for delivery in 3 months.

How a RAG assistant works on your procedures

RAG stands for retrieval-augmented generation. In practice, the system splits your procedures (work instructions, safety sheets, inspection plans, maintenance instructions) into indexed passages. When an operator asks a question, the tool retrieves the relevant passages, checks the user is allowed to see them, then the model writes a short answer citing the document and its revision.

The assistant does not replace the procedure: it leads to it. Every answer must show the link to the official document, its revision number and effective date. If no reliable source exists, the right answer is 'I cannot find an approved procedure on this point'.

2026 budget and architecture

Monthly cost depends mostly on question volume. For 150 active users asking 10 questions per working day, that is about 33,000 questions a month, at 1 to 3 euro cents each depending on the model: the 300 to 1,200 EUR monthly range covers the model, index hosting and logging.

Data security is non-negotiable. Procedures hold industrial know-how, sometimes information covered by customer confidentiality agreements. Require EU hosting, no reuse of data to train models, access filtering applied before generation rather than after, and a log the CISO can audit.

Mini case study

Marc, plant director of a components factory (350 employees, three shifts), estimates his 120 operators and technicians lose an average of 15 minutes a day searching for a procedure or waiting for a team leader to provide it.

Calculation: 120 people x 15 minutes x 220 days = 6,600 hours a year. At 38 EUR loaded hourly cost, that is 250,800 EUR. Prudently counting only 40 % of that time as truly recovered, the gain reaches about 100,000 EUR per year, not counting fewer non-conformities caused by outdated versions.

On cost: a 40,000 EUR excl. tax project, AI usage at 800 EUR a month, 49,600 EUR in year one. Payback arrives around month six. Illustrative scenario, not a real client.

FAQ

Our procedures are scanned PDFs, is that a problem?

Scanned PDFs go through a text recognition step, adding 2,000 to 5,000 EUR to the project depending on volume. Complex diagrams and tables sometimes require partial re-typing of the most consulted documents.

How do we stop an operator from seeing a confidential document?

Filtering happens at retrieval time, based on corporate directory groups. An unauthorised document is never sent to the model, so it cannot appear in an answer.

What happens when a procedure is updated?

The index resyncs automatically, usually nightly or within the hour for critical documents. The old revision is removed from the index, which reduces the risk of applying an outdated version.

Can we host everything on our own servers?

Yes, with an open-source model deployed on premise. Plan for a GPU server at 15,000 to 40,000 EUR and answer quality slightly below the best hosted models.

Do we need to consult the works council?

In Germany, yes: the works council has co-determination rights over any technical system able to monitor employee activity, as does the CSE in France. Consulting upfront avoids blocking deployment for several weeks.

Let's scope your project. Tell us your document sources, number of users and hosting constraints: we will price an assistant between 25,000 and 50,000 EUR excl. tax with a 3-month plan. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.