AI server for certification and notified bodies: technical file pre-checks kept in-house

What an AI server does for a conformity assessment body

For a notified body or a certification body, an AI server is a language model on the body's own hardware that reads a manufacturer's technical file, finds where each item of the requirements list is addressed, flags gaps and inconsistent document versions, and drafts report text from the assessor's notes. The assessor checks every finding and makes every decision. Because the model runs on the body's own server, the manufacturer's documents never leave its control, which is what the professional secrecy rules of EU product law expect of it.

This guide is for managing directors, heads of certification and technical managers at notified bodies, product certification bodies and management-system certification bodies, and for the IT team that runs their systems. It covers the confidentiality rules that shape the choice of tool, the tasks a model can take over in an assessment, how files of different manufacturers stay apart, and the server that carries the work. Manufacturers preparing their own files will find their side in our guides to an AI server for medical device manufacturers and to machine manuals under the Machinery Regulation. Financial audit, with its own confidentiality rules, is the subject of our guide to an AI server for audit and accounting firms.

This is general information on EU law as of October 2026, not legal advice. Your legal counsel, your notifying authority and, for accredited activities, your accreditation body confirm what applies to your organisation.

Professional secrecy in EU product law

The EU's common framework for marketing products, Decision 768/2008/EC, sets out reference provisions that sector legislation copies. Its Article R17 on requirements for notified bodies says in paragraph 10 that "the personnel of a conformity assessment body shall observe professional secrecy with regard to all information obtained in carrying out their tasks", and that proprietary rights are to be protected. Article R20 covers subcontracting: where a notified body subcontracts specific tasks, it ensures that the subcontractor meets the same requirements, informs the notifying authority, and takes "full responsibility for the tasks performed by subcontractors or subsidiaries wherever these are established".

The newer regulations keep this line and sharpen it. The Machinery Regulation, Regulation (EU) 2023/1230, repeats the secrecy duty in Article 30(10) for the tasks under its conformity assessment annexes and adds: "Proprietary rights, intellectual property rights and trade secrets shall be protected." Article 30(4) asks the body to ensure that the activities of its subsidiaries or subcontractors "do not affect the confidentiality, objectivity or impartiality of its conformity assessment activities". The AI Act, Regulation (EU) 2024/1689, requires in Article 31(7) documented procedures ensuring that "personnel, committees, subsidiaries, subcontractors and any associated body or personnel of external bodies" maintain the confidentiality of the information they receive, and in Article 31(2) "suitable cybersecurity requirements".

EUR-Lex texts of Decision 768/2008/EC and Regulation (EU) 2023/1230, and the Commission's AI Act Service Desk for Article 31 of Regulation (EU) 2024/1689, all read 25 October 2026.

By our reading, a cloud AI service that receives a manufacturer's technical file is an outside party that the body's confidentiality procedures and its contracts with manufacturers must cover, whether or not it counts as a subcontractor in the legal sense. A model on the body's own server removes that party from the chain. Certification bodies outside the notified-body system work under accreditation standards such as ISO/IEC 17021-1:2015, the requirements for bodies providing audit and certification of management systems, and ISO/IEC 17065:2012, the requirements for bodies certifying products, processes and services. Their contracts with clients usually carry confidentiality duties of their own, which the same reasoning reaches.

Technical file review and audit report drafting with AI

An assessment produces three kinds of desk work that suit a language model: matching a file against a list, checking a file for consistency, and turning notes into text. In each, the model prepares and the assessor decides. The method behind the first two is retrieval: the file is split into passages, an index finds the passages relevant to a question, and the model answers with the passage cited, as our guide to RAG on company data explains.

Our summary of typical desk tasks; the requirements list comes from the body's own checklist for the legislation concerned.

A requirements pre-check works through the body's checklist one item at a time. For a machine in Part A of Annex I of the Machinery Regulation, for which Article 25(2) requires EU type-examination followed by conformity to type based on internal production control, full quality assurance or unit verification, that checklist follows the essential health and safety requirements and the content of the technical documentation. For each item the model searches the file and returns the passages it found, with document name, revision and page. The assessor reads those passages instead of searching for them. Gaps are the items where nothing relevant was found; the assessor confirms each one, since a missing hit can also mean a scanned page without text.

Version checks catch a frequent kind of inconsistency: a risk assessment that cites a drawing in revision C while the drawing list shows revision E, or a test report for a component that the bill of materials no longer contains. A model reads the document lists and references across the file and lists every mismatch for the assessor. Scanned drawings, certificates and stamped pages need a vision-language model or text recognition first, which our guide to vision-language models for document AI sizes.

Report drafting is the most visible saving and the one to watch most closely. The model turns structured notes into the body's report wording, using approved phrases from the body's templates. It must not invent findings, so the input is the assessor's notes, and the draft is marked as a draft until the assessor has edited and signed it.

Why the workload grows

Two EU deadlines bring more files to notified bodies. The Commission's machinery page states that Regulation (EU) 2023/1230 "applies on a mandatory basis as of 20 January 2027", and its recitals 54 and 55 call for third-party conformity assessment of safety components with fully or partially self-evolving behaviour using machine learning, as our guide to GPU servers for robotics simulation quotes. For medical devices, the Commission's guidance MDCG 2025-6 treats AI in a device as high-risk under the AI Act when, among other conditions, the device is subject to assessment by a notified body, and under the AI Act as amended by Regulation (EU) 2026/1744 these rules apply from 2 August 2028. Our article on the AI Act for machine builders sets out the timeline.

More files with more software in them mean more pages per assessment for the same assessors. A model that prepares the desk work lets each assessor spend more of the day on judgement.

We supply AI servers built to order and set up the model, the document search and the access rights under our Private AI/ML service; send us the kinds of files you assess and their typical size, and our team replies with a configuration.

Keeping manufacturers' files apart

Professional secrecy applies between clients as much as towards the outside world. Manufacturer A's file must never appear in an answer to an assessor working on manufacturer B, even when both make similar products. We recommend four layers of separation:

1. One index per assessment. Each project gets its own document collection. A search runs only in the collection of the project the assessor is assigned to.
2. Rights from the assignment. Access follows the project assignment in the body's directory, so an assessor who leaves a project loses access to its index the same day.
3. No training on client files. The model is used as delivered. It learns nothing from the files it reads, so nothing from one client can surface in another client's answers.
4. Deletion with the file. Each index, the query log entries of the project and any cached text are deleted when the retention period of the assessment file ends, in the same step as the file itself.

Report templates and the body's own guidance documents form a shared collection that every project may search. Earlier reports for other manufacturers do not belong in it. Our guides to preparing company documents for an AI assistant and to trade secrets, NDAs and AI tools cover the preparation and the contract side.

Configuration options

Technical files are long, so the model needs a long context and room for several assessors at once. gpt-oss-120b, under the Apache 2.0 licence with 117 billion parameters of which 5.1 billion are active per token, accepts up to 131,072 positions according to its configuration file. Its weights take about 65 GB (60.8 GiB) on the Hugging Face files page. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card and less the weights, one RTX PRO 6000 Server Edition holds 19 conversations of 32K tokens with a 16-bit cache, and a tensor-parallel pair of two cards, one model spread over both, holds 93.

Users-per-card rule with the 16-bit cache of gpt-oss-120b (18 full-attention layers, 8 KV heads, head dimension 64 in its config.json): 1.125 GiB per 32K, 4.5 GiB per 128K; pair (166.08 − 60.8) / 1.125 = 93 and / 4.5 = 23. Power by our estimate (cards, two processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

A conversation here is one request being answered at the same moment, such as one item of a pre-check with its retrieved passages. A pre-check of a checklist with several hundred items is therefore a queue of several hundred requests, which the server works through many at a time and which can run as a batch job outside working hours. The 128K figure matters when an assessor asks a question about a whole document at once, such as a long test report.

Two cards suit a body with a few dozen assessors in one field. Four cards suit a body with several fields, or one that runs pre-checks for many files in parallel; the two pairs can also carry different models, for example one for text and one for scanned pages. Without a server room of your own, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU; that location and the service provider then go into the confidentiality documentation.

Notes for the IT team

1. Network. The AI server sits in the internal network without inbound access from the internet. Manufacturers upload files to the existing portal; a connector copies them into the project index.
2. Sign-in and rights. Sign-in through the company directory; each project collection is tied to directory roles that follow the assignment.
3. Logs. The query log contains passages from client files, so it carries the same confidentiality and retention as the files. Our guide to an LLM gateway with keys and logging shows one design.
4. Pinned versions. Record the model checksum, the serving engine version, the driver branch and the container images, so that a pre-check can be repeated with the same stack if a manufacturer disputes a finding.
5. Hardening. BMC, drivers and the container runtime as in our guide to securing a GPU server, which also helps answer the AI Act's call for suitable cybersecurity.
6. Delivery. On request the server arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed.

Power and room

By our estimate, a two-card server draws about 1.8 kW at the wall and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. Four cards at 600 W draw about 3.4 kW and need 32 A or three-phase. A measurement on the delivered system confirms the estimate. Our guide to preparing a server room for a GPU server lists what to check before delivery.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the model, the project indexes, access control and the query log; our Private AI/ML page states an NDA before technical detail, a data processing agreement on request and subprocessors named in the contract.

FAQ