AI Assistant over Quality Documentation with RAG: Amsterdam 2026

The verdict in three sentences

When 4,000 procedures and work instructions are spread across a document management system, shared drives and binders on the shop floor, an operator loses about 20 minutes per search and sometimes follows an outdated version. An AI assistant built on RAG answers in plain language and cites the source document and revision, for an API cost of 200 to 900 EUR a month. For a 350-employee manufacturer it is a medium to complex application at Kolonell, justified in year one by time recovered and non-conformities avoided.

The problem: rich documentation nobody can find

Around Amsterdam and the Eindhoven corridor, electronics manufacturers work under ISO 9001 and often ISO 13485 or AS/EN 9100 for medical or aerospace customers. Quality documentation is extensive: procedures, work instructions, machine setup sheets, inspection plans, non-conformance reports, audit findings. It is accurate, but nobody finds it quickly.

A cleanroom operator unsure of a reflow parameter calls the team lead, who searches the DMS, who calls quality. Twenty minutes later the line has waited or, worse, the operator carried on from memory. Every customer audit raises a few findings like "instruction not available at the workstation".

RAG architecture: what to build

Retrieval-augmented generation retrains no model. Documents are split into passages, turned into vectors and stored in a search index. For each question, the system retrieves the 5 to 10 most relevant passages and passes them to the language model with a strict instruction: answer only from these extracts and cite the source. If nothing matches, the assistant says so rather than guessing.

The hard parts are industrial rather than technical: reading parameter tables and diagrams correctly, indexing approved revisions only, enforcing permissions (an operator does not see supplier audits), and logging every question for the quality team.

A simple application, 2,500,000 to 4,000,000 FCFA (3,810 to 6,100 EUR), can serve as a pilot on one production area and a few hundred documents. Maintenance follows the 35,000, 75,000 or 120,000 FCFA monthly grid, excluding API fees paid directly to the provider.

Compliance and governance

Under the EU AI Act, a documentation assistant of this kind generally falls into the limited-risk category: users must be told they are dealing with AI and humans keep control of decisions. Data stays in the EU, the API provider commits contractually not to train on it, and the quality team signs off a set of 100 to 200 reference questions before go-live. Every answer is logged, which also helps in audits and in discussions with the works council.

Mini case study

Eva, quality director at a 350-employee manufacturer near Amsterdam, counts 120 document searches a day over 220 days, 26,400 a year. Going from 20 to 2 minutes saves 18 minutes each, 7,920 hours a year. At 48 EUR per loaded hour, that is 380,160 EUR, before counting line stoppages avoided.

Annual running costs: 600 EUR a month in API fees, 7,200 EUR, plus Premium maintenance at 120,000 FCFA a month, 1,440,000 FCFA (about 2,200 EUR). With a complex build around 9,000,000 FCFA (about 13,700 EUR), even counting only a quarter of the theoretical gain, the investment is covered in under three months.

FAQ

Can the assistant make up an answer?

The risk is sharply reduced by answering only from retrieved extracts and citing the source. On a 150-question test set, the target is over 95 % correct, sourced answers before go-live.

Do we need to reorganise the whole DMS?

No, but approved revisions must be identifiable. A 2 to 4 week clean-up of duplicates and obsolete documents clearly improves answer quality.

What does the API cost day to day?

200 to 900 EUR a month for 2,000 to 8,000 questions, as a 2026 order of magnitude. Each answer costs a few cents.

Does data leave the EU?

No, the architecture uses providers with EU hosting and a no-training commitment. Restricted documents stay filtered by user permissions.

How long until a first pilot?

Allow 6 to 8 weeks for a pilot on one area, then 3 to 4 months to extend to the whole plant. Operators are autonomous after 15 minutes of onboarding.

Let's scope your project. Tell us about your DMS, document volume and the production areas involved, and we will scope a focused pilot and the extension plan. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.