Operating manuals with AI under the Machinery Regulation: drafts, translations, hardware

Operating manuals with AI: what changes in January 2027

A machine builder can use AI for operating manuals now: a language model on the company's own hardware drafts the sections that change from one machine variant to the next and pre-translates them, and the technical writer edits and releases every section. The risk assessment, the warnings and the signature under the declaration of conformity stay with people. Regulation (EU) 2023/1230 on machinery applies from 20 January 2027, permits digital instructions on conditions and leaves the duty for their content with the manufacturer, whichever tool wrote the first draft.

The drafts come from data the company already has: the parts list, the CAD parameter lists and the existing base manual, with the terminology kept consistent across variants.

For a documentation team, a workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q cards or a server with two RTX PRO 6000 Blackwell Server Edition cards runs a drafting model and a translation model side by side. We wrote this for managing directors, heads of technical documentation and the design engineers who release a machine. Technical terms are explained where they first appear, the hardware is in one section with a table, and the notes for the IT team come at the end.

Machinery Regulation 2023/1230: what it asks of the instructions

This is general information on the EU text, not legal advice; your legal counsel or your notified body confirms what applies to your machine. The Commission's machinery page states that the Regulation "applies on a mandatory basis as of 20 January 2027"; machinery placed on the market before that date has to comply with the Machinery Directive 2006/42/EC. Article 10(7) of the Regulation sets the rules for instructions for use.

Regulation (EU) 2023/1230, Article 10(7), EUR-Lex, read 25 October 2026; summary in our words. Where machinery is meant for, or may reasonably be used by, non-professional users, the safety information also comes on paper.

The content of a manual follows harmonised and international standards. ISO lists ISO 20607:2019, "Safety of machinery – Instruction handbook – General drafting principles", which sets requirements for the machine manufacturer and builds on the information-for-use requirements in clause 6.4.5 of ISO 12100:2010. ISO marks the 2019 edition as to be revised, with ISO/FDIS 20607 expected to replace it, so check the edition your documentation follows when the new one appears.

For the documentation team, two points in the table change the work. A digital manual is a file that has to stay retrievable for a decade or longer, for every model and every version. And every language a market needs is a full copy of that file. A machine family with many options produces dozens of variant manuals in many languages, and the effort sits in keeping all of them consistent with the machine as built.

AI for technical writers: where a language model helps

A language model, or LLM, is software that reads and writes text. In a documentation team it works with retrieval, often called RAG: before it writes, it searches your own documents, such as the base manual, the parts list and earlier variant manuals, and it writes from what it found, with a reference to the source. Four tasks suit it.

1. Variant sections. From the bill of materials (BOM) in the ERP system and the parameter lists exported from CAD, the model drafts the sections that differ for a variant: technical data, the list of spare parts, the description of an option, the changed steps of a maintenance task. The writer starts from a draft that already has the right part numbers and values.
2. Consistency checks. The model compares a section with the parts list and the terminology list and reports what does not match: a motor rating in the technical data that differs from the BOM, a term that the style guide replaced, a cross-reference to a chapter that no longer exists.
3. Restructuring old manuals. Long legacy manuals in word-processor files can be split into topics, given headings and metadata, and moved into a component content management system, the database in which many documentation teams keep reusable text modules.
4. Pre-translation. The model translates new and changed segments against your translation memory and terminology, as described below.

The input data decides most of the quality. The German use-case catalogue KI-Syndikat, which describes this use case for machine builders, warns that free-form PDFs and unstructured spreadsheet lists need to be restructured before an AI can use them. The same page labels its own hour figures for a manual "Modellwerte, keine Erhebung", model values rather than a survey, and recommends measuring your own team's times first. We give no saving figure for that reason: a pilot with one machine family gives you your own.

The same index that feeds the drafts can later answer questions from service technicians and the hotline, which our guide to an AI service assistant for machine builders covers. Our article on an AI server for manufacturing shows how such an assistant sits next to inspection and simulation workloads in a plant.

What stays with people

The Regulation puts the design and the documentation on the manufacturer. Under Article 10(1), the manufacturer ensures that the product has been designed and constructed in accordance with the essential health and safety requirements of Annex III; under Article 10(2), it draws up the technical documentation of Annex IV, Part A, and carries out the conformity assessment. Recital 32 adds: "The manufacturer should also ensure that a risk assessment is carried out for the product", and that the risk assessment should also address future updates of software installed in the machinery. ISO 20607 sets out the principles for giving information on residual risks in the instruction handbook.

That gives a clear split between the model and the people.

• Risk assessment. The engineers carry it out and document it. A model can find earlier assessments of similar machines for the team, but it does not decide on a hazard or a protective measure.
• Safety and warning texts. They follow from the risk assessment. We recommend that warnings come from an approved library of safety messages that the writer links to a hazard, and that the model never formulates a warning freely.
• Release. Every generated section is marked as generated and released by a named person, in a step separate from the language review. The declaration of conformity is signed by the manufacturer.

An LLM that drafts text in the documentation office is a tool for the writers. AI that is part of the machine itself, for example in a safety function, is a separate subject under the Machinery Regulation and the AI Act, covered in our article on the AI Act for machine builders.

Translating machine manuals with AI on premise

Because the language is set by the Member State where the machine is made available, a machine builder that sells across the EU needs its manuals in many languages. Most documentation teams already work with a translation memory, a database of approved sentence pairs that translation tools reuse whenever the same sentence appears again. A local model fits into that workflow: exact matches come from the memory, the model pre-translates only new or changed segments with your terminology list, and a translator or a reviewer for each language approves the result.

The model has to know the languages. The card of EuroLLM-22B-Instruct-2512 lists 35 languages, among them all 24 official EU languages, under the Apache 2.0 licence, with a context window of 32,768 tokens; its files take 45.3 GB, 42.2 GiB. The card of gpt-oss-120b states no language list, so test it per language before you rely on it for translation. Our guide to on-premise machine translation compares dedicated translation models and their licences, and our article on tokens per word in EU languages shows why a German or Polish text needs more model capacity than the same text in English.

A worked example, with example values: a manual of 150 pages at 300 words a page has 45,000 words; into ten languages that is 450,000 words. At the planning figure of 2 output tokens per word from our translation guide, that is 900,000 tokens, about 31 tokens a second over an eight-hour day. EuroLLM-22B reads all 45.3 GB of its weights for each generated token, so one RTX PRO 6000 Server Edition at 1,597 GB/s reaches about 35 tokens a second for a single stream, by our estimate an upper bound. Segments translated in parallel share those weight reads, so the card produces several times that rate in total, and the example job fits a working day on one card by our estimate. A test with your own manuals replaces these figures.

Why parts lists and CAD data stay on your own hardware

The data that makes a draft useful is the data a machine builder protects: parts lists with supplier part numbers, CAD parameters, customer-specific options and the know-how in the base manual. Directive (EU) 2016/943 on trade secrets protects information only if it is secret, has commercial value because it is secret, and "has been subject to reasonable steps under the circumstances" by the person in control of it to keep it secret (Article 2(1)). Customer contracts and non-disclosure agreements can add their own limits on who may process design data.

A model on your own workstation or server keeps the documents, the prompts and the drafts on hardware under your access rules, with no external service unless you switch one on. Whether a public AI service is acceptable for a given data set is a question your legal counsel answers for your contracts; our article on trade secrets, NDAs and AI tools covers it in more detail. Manuals rarely contain personal data, but customer-specific manuals and service histories can, and then the GDPR applies as well.

We supply RTX PRO 6000 workstations and servers for this work and set up the models through our Private AI/ML service; send us the number of writers, your manual volume, the languages and where your parts lists and CAD data sit, and our team replies with a configuration.

Hardware for a documentation team

Two numbers size the hardware. The model has to fit into the memory of the graphics card, the GPU, and every request that is being worked on at the same moment needs extra memory, the cache. Our users-per-card rule takes 90 per cent of the 95.6 GiB that the driver reports for a 96 GB card, less 3 GiB per card, less the model's weights: 83.0 GiB per card.

Cards: RTX PRO 6000 Max-Q and RTX PRO 6000 Server Edition. Users-per-card rule as above; model files from the Hugging Face pages of gpt-oss-120b (about 65 GB) and EuroLLM-22B-Instruct-2512 (45.3 GB), read 25 October 2026; workstation power from our buy page for the two-card Max-Q workstation; server power by our estimate with the cards at 600 W, processors and 15 per cent for the rest of the server.

Drafting. gpt-oss-120b, released under Apache 2.0 with 117B parameters of which 5.1B are active per token, takes 60.8 GiB and leaves 22.2 GiB of cache on one card. Its config.json gives 18 full-attention layers with 8 key/value heads of dimension 64, which by our arithmetic makes 1.125 GiB of 16-bit cache per request of 32,768 tokens, so 19 such requests fit at once. A request of that length holds about 18,600 words of German at the mean of 1.76 tokens per word from our tokens-per-word article, enough for a manual chapter, the parts-list extract and the instructions. A team of writers rarely sends 19 requests in the same second, so one card carries a documentation department.

Translation. EuroLLM-22B at 42.2 GiB leaves 40.9 GiB of cache on the second card. Translation requests are short segments with their context, so the cache is rarely the limit; the generation rate in the worked example above is.

Workstation or server. The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; NVIDIA allows up to four in one workstation, and our page on the two-card Max-Q workstation puts it at about 1.1 kW at the wall, under 5 A on a 230 V circuit. It suits a team that works in one office and has no server room. In a rack server the Max-Q belongs only in a chassis whose maker lists it; for continuous operation in a rack we recommend the RTX PRO 6000 Server Edition. The two-card server has two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, draws about 1.8 kW and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent.

One server for several departments. Manuals are a modest load. A four-card server with 768 GB of memory carries the documentation models next to a service assistant, offer drafting or order entry; at about 3.4 kW it needs 32 A or three-phase, as our guide to GPU server power on 16 A, 32 A and three-phase feeds explains. For questions about the card itself, our article on RTX PRO 6000 specifications compares the three editions.

Notes for the IT team

1. Engine. vLLM serves both models through an OpenAI-compatible API, which the content management system or a small script calls. The gpt-oss-120b card states that the model "should only be used with the harmony format", so use an engine release that supports it.
2. Data sources. Give the platform read-only exports: the BOM from the ERP system, parameter lists from the PDM or CAD system, the terminology list and the translation memory. The model writes drafts into the authoring system, never back into ERP or PDM.
3. Traceability. Manuals stay online for at least 10 years, so store with every generated section the model version, the prompt template and the version of the source data. A question about a manual in 2035 can then be traced back.
4. Network. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, while routes such as /invocations and /metrics answer without the key (vLLM 0.32.0 security guide), so keep the engine on an internal network behind an LLM gateway. Run vLLM 0.31.0 or later, the NVIDIA Container Toolkit 1.17.8 or later and a driver release that NVIDIA lists as fixed in its bulletin of 30 September 2026.
5. Software on delivery. The workstation or server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime; our guide to RTX PRO Blackwell software support lists the engine versions for these cards.

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 workstations with two RTX PRO 6000 Max-Q and servers with two, four or eight RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered within 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 models, the retrieval over your documents and the connection to your authoring system. If you want to start with a test, describe one machine family and its manuals, and we propose a pilot and the hardware for it.

FAQ