Safety data sheets with local AI: drafting, translating and reading SDS under REACH, CLP

What a local AI does in safety data sheet work

A local AI server helps with safety data sheets (SDS) in two directions: it drafts the free-text sections of your own SDS and translates them into the language of every market you serve, and it reads the SDS your suppliers send into your chemical register. Because the model runs on a server inside the company, the composition in section 3 and the formulation behind it never go to an external AI service; classification stays in rule-based SDS software, and a competent person releases every sheet.

This guide is for the managing director or the product safety manager at a maker of paints, adhesives, cleaning agents or other mixtures, and for a downstream user who receives hundreds of supplier SDS. The server comes in one section further down; the notes at the end are for the IT team. Everything legal here is general information on EU law as of October 2026; your regulatory adviser confirms what applies to your products.

What REACH and Annex II ask of a safety data sheet

Article 31 of REACH, Regulation (EC) No 1907/2006, is the core rule. A supplier of a substance or mixture that meets the criteria for classification as hazardous provides the recipient with a safety data sheet "compiled in accordance with Annex II". The sheet must reach the recipient free of charge, on paper or electronically, "no later than the date on which the substance or mixture is first supplied" (Article 31(8)). It is supplied "in an official language of the Member State(s)" where the product is placed on the market, unless those states provide otherwise (Article 31(5)). A company selling in ten Member States therefore keeps the same sheet in up to ten languages.

The sheet has 16 headings, from section 1, identification of the substance or mixture and the company, through section 2, hazards identification, and section 3, composition and information on ingredients, to section 16, other information. Article 31(9) requires an update "without delay" as soon as new information that may affect the risk management measures, or new hazard information, becomes available, and the revised sheet goes free of charge to every recipient supplied in the preceding 12 months. Under Article 36, manufacturers, importers, downstream users and distributors keep the information they need for their duties for at least 10 years after they last supplied or used the substance or mixture.

The content of each section is set by Annex II, in the version of Commission Regulation (EU) 2020/878, which has applied since 1 January 2021 after a transition that ended on 31 December 2022. Four of its points shape any AI project:

1. A competent person. Point 0.2.3 says the sheet "shall be prepared by a competent person" who takes into account the needs and knowledge of the users, and that such persons "have received appropriate training, including refresher training". The AI prepares drafts; the competent person reviews and releases them.
2. Plain language. Point 0.2.4 asks for text that is "simple, clear and precise, avoiding jargon, acronyms and abbreviations".
3. Composition. For mixtures, section 3.2 gives the concentrations as exact percentages or ranges of percentages, in descending order by mass or volume. Section 1.1 carries the unique formula identifier (UFI) where the mixture has one.
4. Visible changes. Section 16 asks a revised sheet to give "a clear indication of where changes have been made to the previous version".

Why sections 2 and 3 keep the work inside the company

A safety data sheet is a published document, but its drafts are not. To write sections 2, 3, 8 and 11, the author works from the full formulation: every raw material, its exact share and the supplier data behind it. That is the trade secret of a coatings or adhesives maker, and an SDS project that sends formulations to a cloud AI service puts the recipe into another company's systems. KI-Syndikat, a German catalogue of AI use cases for small and mid-sized companies, notes for its SDS use case that formulations are usually trade secrets and calls an on-premise installation or private hosting the safest option for highly confidential formulations.

With a cloud service the question becomes one of contracts: what the provider may do with inputs, where it processes them and which subcontractors it uses, as our guide to trade secrets, NDAs and AI tools explains. On a local server the model, the formulation database and the drafts stay on hardware the company controls.

The CLP changes behind the next SDS update wave

SDS work comes in waves set off by changes in classification rules. Two are under way in October 2026.

New hazard classes. Commission Delegated Regulation (EU) 2023/707 added four hazard classes to CLP, Regulation (EC) No 1272/2008: endocrine disruption for human health, endocrine disruption for the environment, PBT and vPvB properties, and PMT and vPvM properties, with the new statements EUH380 to EUH451. New substances had to be classified from 1 May 2025; substances placed on the market before that date "are not required to be classified" under the new classes until 1 November 2026. Mixtures have to be classified under the new classes from 1 May 2026, and mixtures placed on the market before that date by 1 May 2028. By our reading, every reclassification that changes a hazard is new hazard information under Article 31(9) and so starts an SDS update, in every language.

The CLP revision. Regulation (EU) 2024/2865, published on 20 November 2024, revised CLP itself. Its provisions apply in stages, a first set from 1 July 2026 and others from 1 January 2027. Regulation (EU) 2025/2439 of 26 November 2025 moved a further part to 1 January 2028; its recitals name the obligations on label formatting, advertisements, distance sales offers and relabelling. Which provision applies to which of your products, and which transitional rule covers products already on the market, is a question for your regulatory adviser.

For a company with a few hundred products, each date means checking every formulation, updating the affected sections, translating the changes and sending the new sheets to every recent customer. KI-Syndikat estimates 4 to 8 hours of manual work for a new SDS and 2 to 4 hours for an update after a regulatory change, figures that come from the catalogue and not from a measurement of ours. Most of that time goes into text, versions and languages, where a language model helps.

Drafting: what the AI writes and what stays rule-based

SDS authoring software is mostly rule-based. KI-Syndikat describes it as a substance database linked to the formulations, with CLP classification calculations and phrase catalogues, and gives language models a supporting role: drafting free-text sections such as section 8 and checking texts for completeness and consistency. That division is the right one, because CLP classification is arithmetic and the hazard statements have fixed wording. Germany's federal institute for occupational safety and health (BAuA) explains that the H statements set out in CLP Annex III, Part 1, "are listed in all of the languages of the EU", and that the EUH statements are in Parts 2 and 3. A model that paraphrases an H statement creates an error.

Section headings from REACH Article 31(6) and Annex II as amended by Regulation (EU) 2020/878 (EUR-Lex, read 25 October 2026); division of work after KI-Syndikat's SDS use case (read 25 October 2026) and our reading.

In practice the assistant reads the formulation record, the classification result, the raw-material SDS and your earlier sheets for similar products, and proposes the text of sections 4 to 8 and the change summary in section 16, in a fixed structure that the SDS software imports into a review queue. It does not invent physical or toxicological values: sections 9 to 12 take numbers only from a document it names, and an empty field stays empty. The competent person accepts, edits or rejects each draft and releases the sheet in the SDS software.

Translating SDS into the languages of your markets

Translation is often the largest share of the work, because each change multiplies by the number of languages. The fixed phrases come from the phrase catalogue of the SDS software; the model translates the free text around them.

EuroLLM-9B-Instruct covers the languages. Its card lists 35 languages, all 24 official languages of the EU among them, and it is published under Apache 2.0, a licence that allows commercial use. Its four safetensors files total 18.3 GB, which is 17.05 GiB, and so the model fits one RTX PRO 4500 Server Edition with about 8.6 GiB left for working memory by our users-per-card rule (0.9 × the driver-visible memory, less 3 GiB per card and less the weights). The card's training table gives a sequence length of 4,096 tokens, so a pipeline translates section by section, which suits an SDS anyway. The card also warns that the base model "has not been aligned to human preferences", one more reason for the review step.

As a planning example with values to replace by your own, an SDS of 4,000 words translated into 23 further languages is 92,000 words, and an update wave over 200 products is 18.4 million words. Much of an updated sheet is unchanged text, so a translation memory, a database of approved sentence pairs, reuses the approved translation, and only new or changed sentences go to the model. We found no published translation throughput for EuroLLM on the cards we supply; a day's sample of your sheets on a test card gives the figure. Our guide to on-premise machine translation compares other translation models and their licences, and tokens per word in European languages explains why Polish or Hungarian take longer than German.

We supply the RTX PRO 4500 Server Edition and build AI servers to order for this work; send us your number of products, languages and incoming SDS per year, and our team replies with a configuration.

Reading incoming supplier SDS into the chemical register

A downstream user receives SDS in every layout: scans, PDFs with tables, sheets in another language. Point 0.2.2 of Annex II says the sheet "shall enable employers to determine whether any hazardous chemical agents are present" at the workplace, so the information has to reach the chemical register and the risk assessment. A vision-language model, one that reads page images as well as text, does the intake. It reads each sheet, returns the product name, supplier, version and revision date, the H statements, the substances in section 3 with CAS numbers and ranges, and whether exposure scenarios are annexed for identified uses (Article 31(7)), and writes them as a record the register can check.

Qwen3.8-27B is a model for this step: Qwen's card calls it "a native vision-language model" that reads documents, under Apache 2.0. By the method of our guide to vision-language models for document AI, one RTX PRO 6000 Server Edition reads about 0.24 pages a second with the model in FP8, our estimate scaled from a published OCR measurement. A backlog of 1,000 supplier SDS of 15 pages each, example values, is 15,000 pages, about 17 hours of one card by our estimate. On RTX PRO 4500 Server Edition cards the FP8 version needs a pair of cards; no page rate has been published for that layout, so a test with your sheets decides.

The useful part comes after extraction: the model compares the new sheet with the version on file and flags changes in classification, composition or exposure limits for the person who maintains the register, who checks every flag. Our page on supplier declarations with AI for RoHS, REACH and SCIP covers the related declarations, and drafting the workplace risk assessment from the register is the subject of workplace risk assessments drafted with AI.

Which server for SDS work, and what it needs in the building

SDS work needs a modest model, and none of it needs the largest GPUs. It pays to run it on the company AI server that also carries the assistant for other departments, so the server is sized for all of them together. Three configurations cover the usual cases; the capacity column follows our users-per-card rule with an FP8 cache.

Usable memory by our rule: 25.7 GiB per RTX PRO 4500 Server Edition, 83.04 GiB per RTX PRO 6000 Server Edition; Qwen3.8-27B FP8 28.75 GiB (Hugging Face files page, read 25 October 2026), 1.14 GiB FP8 cache per 32K conversation; power by our estimate (cards, processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

Four RTX PRO 4500 Server Edition cards. The NVIDIA RTX PRO 4500 Blackwell Server Edition is a passive, single-slot data-centre card with 32 GB of GDDR7 at 800 GB/s and 165 W. Four of them in a 2U server carry the reading model as a pair, the translation model on its own card and a search model on the fourth, at about 1.4 kW, within one 230 V 16 A feed (2.94 kW at 80 per cent). This is the configuration we would size first for a company whose AI server does SDS and document work. Our page on four RTX PRO 4500 Server Edition cards in 2U lists the chassis, and Qwen3.8-27B on the RTX PRO 4500 Server Edition works out the layouts.

Two RTX PRO 6000 Server Edition cards. Our entry server, with two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives, holds Qwen3.8-27B on one 96 GB card and leaves the second for translation and an assistant for other departments.

Four RTX PRO 6000 Server Edition cards. The inference node with 768 GB serves several departments at once, for example SDS, formulation work on your lab data and a general assistant.

All three belong in a server room or a cooled rack, not in a laboratory or an office. Research workloads such as simulation are a different case, covered in our GPU server for chemistry and materials research. Without a server room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. A company without its own IT team will find the division of tasks in our guide to an AI server without an IT department.

Notes for the IT team

1. Integration. The SDS software stays the system of record. The AI reads the formulation database with read-only rights and writes drafts into a review queue; vLLM's structured outputs make it answer in a JSON schema, so imports can be validated.
2. Fixed phrases. H and EUH statements come from the phrase catalogue in each language, and a check after translation compares them with the catalogue.
3. Models and versions. Load models from a pinned revision in safetensors format. Run vLLM 0.31.0 or later, keep the engine on an internal network behind a gateway, and keep the NVIDIA driver at a release that NVIDIA's bulletin of 30 September 2026 lists as fixed, as our guide to securing a GPU server explains.
4. Evaluation. Before going live, compare the extraction on 50 to 100 sheets you have already registered by hand, per layout and language, and record the error rate per field.

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 supply the RTX PRO 4500 Server Edition and RTX PRO 6000 Server Edition and build AI servers to order with two, four or eight 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 models, the retrieval on your earlier sheets and the connection to your SDS software; describe your SDS volume and your languages, and we propose a pilot on one product family.

FAQ