ProcessChemicals

How to automate safety data sheet authoring without losing control of classification

AI can take much of the copying, cross-checking and translation out of safety data sheet work, if each part of the job goes to the right tool. Classification belongs in a deterministic rules engine, wording comes from approved phrase libraries, a language model handles intake, assembly and change detection, and a qualified person signs every revision. Here is that process step by step, with the EU and US rules it must encode.

Reviewed 8 min read

On this page
  1. Why safety data sheet production strains product stewardship teams
  2. The EU, US and UN rules an SDS pipeline has to encode
  3. From composition record to a signed-off SDS
  4. Five working steps for AI-assisted SDS and label production
  5. Rules engine, language model or expert: who does which SDS task
  6. Relabeling an adhesives family for the new CLP hazard classes
  7. Where AI-assisted SDS programs go wrong
  8. Questions and answers
  9. Sources

Why safety data sheet production strains product stewardship teams

One mixture can need a separate SDS and label for every market, each in the local language, each revised whenever an ingredient, a test result or a rule changes. Across a portfolio of hundreds of products, that becomes a permanent queue of near-identical documents that each carry legal weight.

Most of the effort is not judgment. It is finding the current supplier SDS for each raw material, re-keying concentrations, checking which statements follow from a classification and translating phrases translated many times before. That is where automation pays. Judgment, such as a toxicologist deciding that test data overrides a calculated classification, stays with people and is recorded as theirs.

The EU, US and UN rules an SDS pipeline has to encode

Check the consolidated texts first: several have been amended recently.

REACH Regulation (EC) No 1907/2006, Article 31 and Annex II (as replaced by Commission Regulation (EU) 2020/878)

European Union and EEA

Applies whenYou supply a substance or mixture that is classified as hazardous, is PBT or vPvB, or is on the Candidate List, to professional or industrial users1.

  • Use the sixteen-section Annex II format, in an official language of each Member State where the product is sold12.
  • Revise without delay when new hazard information, an authorization or a restriction affects the SDS, and send it to recipients supplied in the preceding twelve months1.
  • Show the UFI in Section 1 when a mixture has one and it is given on the SDS2.

CLP Regulation (EC) No 1272/2008, with Delegated Regulation (EU) 2023/707 and Regulation (EU) 2024/2865

European Union and EEA

Applies whenYou place substances or mixtures on the EU market and must classify, label and package them3.

  • Classify against Annex I, including the endocrine disruption, PBT/vPvB and PMT/vPvM classes added by Delegated Regulation (EU) 2023/7074.
  • Apply the targeted revision's label rules, such as minimum formatting and hazard information in online offers, on the dates left after Regulation (EU) 2025/2439 postponed several of them56.
  • Notify poison centers under Annex VIII and print the UFI on labels of hazardous mixtures within scope7.

OSHA Hazard Communication Standard, 29 CFR 1910.1200, as updated in 2024[^8]

United States

Applies whenYou manufacture, import or distribute hazardous chemicals that reach US workplaces8.

  • Classify under criteria aligned mainly with GHS Revision 7 and keep the sixteen-section SDS format8.
  • Where an exact concentration is a trade secret, disclose a prescribed concentration range instead8.
  • Meet the extended dates for manufacturers, importers and distributors: May 19, 2026 for substances and November 19, 2027 for mixtures9.

UN Globally Harmonized System of Classification and Labelling of Chemicals (GHS)

International, adopted country by country

Applies whenYou sell into countries that have written a particular GHS revision and chosen building blocks into national law10.

  • Record the revision and building blocks each market applies, because the same mixture can classify differently between countries10.

From composition record to a signed-off SDS

01Composition record02Rules-basedclassification03Phrase-based drafting04Controlled translation05Expert sign-off06Publish and distribute07Change monitor
  1. Composition record

    Versioned ingredients, concentrations, test data and supplier SDS references.

  2. Rules-based classification

    Reproducible calculation per jurisdiction, showing the ingredient behind each hazard.

  3. Phrase-based drafting

    Text assembled from approved phrase IDs, with gaps flagged rather than filled.

  4. Controlled translation

    Stored translations reused; only new text is machine-translated and reviewed.

  5. Expert sign-off

    A qualified person approves the classification and the diff against the last revision.

  6. Publish and distribute

    Revision sent to recipients and logged.

  7. Change monitor

    New data, supplier revisions and rule changes reopen the affected records.

Conceptual flow of an AI-assisted SDS pipeline; the change monitor feeds back into the record. It does not describe any specific software product.

Five working steps for AI-assisted SDS and label production

  1. Build a governed composition and test-data record

    Hold, per product, the CAS and EC identifiers of each ingredient, concentration ranges, impurities that affect classification, test results on the mixture, harmonized entries and specific concentration limits from CLP Annex VI, and the supplier SDS behind each value. A language model can read supplier PDFs into fields, but every extracted value is validated and conflicts between suppliers are flagged.

    Output
    One versioned record per product, with the source of every value
    Owner
    Product stewardship, with R&D and procurement

  2. Classify with a deterministic engine

    Mixture classification follows defined rules: additivity and summation methods, acute toxicity estimates, bridging principles, specific concentration limits and M-factors. These belong in an engine that gives the same answer every time and shows its working. A language model may explain the result, but never decides a hazard class.

    Output
    A classification per jurisdiction, with the calculation trail
    Owner
    Regulatory affairs; a qualified expert approves

  3. Draft sections and labels from approved phrases

    Hazard and precautionary statements come from the regulation’s coded lists; handling, storage and first-aid text comes from an approved library, such as the EuPhraC industry catalogue or your own. The model assembles phrases that fit the classification, fills values from the record and flags sections no approved phrase covers. New wording goes to an expert as a proposal, and translation reuses stored phrase translations first.

    Output
    Draft SDS and label text per language, with a phrase-ID trail
    Owner
    SDS authoring team

  4. Review, sign off and distribute

    Reviewers see what changed since the last revision, why, and which data version it rests on. The system records each approval. Proving that revisions reached recipients supplied in the preceding twelve months is far easier when shipping records link to product versions1.

    Output
    A signed revision and a distribution log
    Owner
    Qualified person for each jurisdiction

  5. Watch for change triggers

    Typical triggers are new test results, a supplier SDS revision, an adaptation to technical progress that changes a harmonized classification, a new Candidate List entry, a country adopting a newer GHS revision and a formulation change. Each event is mapped to the products it affects and opens a revision task stating the reason.

    Output
    A revision queue ordered by legal deadline
    Owner
    Product stewardship

Rules engine, language model or expert: who does which SDS task

Failed SDS automation usually gave a language model work that needed reproducibility.

TaskRules engineLanguage modelQualified expert
Mixture classificationCalculates from composition, limits and test dataExplains the result and flags missing inputs; never decidesApproves, and overrides calculation where test data or judgment requires
Supplier SDS intakeValidates extracted values against formats and rangesExtracts composition, classification and test data from PDFsResolves conflicts between suppliers or with harmonized entries
Handling, storage and disposal textEnforces phrases a classification requiresRetrieves approved phrases and drafts product-specific wordingApproves wording not already in the library
TranslationReuses stored translations by phrase IDTranslates the remainder and flags unapproved termsA native-language reviewer signs off new text
Regulatory change monitoringMaps an amendment or new listing to affected productsSummarizes the amendment and drafts the revision noteDecides whether to revise and how urgently

Relabeling an adhesives family for the new CLP hazard classes

Where AI-assisted SDS programs go wrong

A model paraphrases a hazard statement

Early signalDraft wording that matches neither the official text nor any phrase ID.

MitigationGenerate statements only from coded lists and block any sentence lacking a phrase ID or recorded expert approval.

Supplier data goes stale

Early signalRaw-material SDS dates older than the product revisions that rely on them.

MitigationTrack supplier revision dates and treat each supplier revision as a change trigger.

Confidential compositions leave the company

Early signalFull recipes pasted into public chat tools.

MitigationRun models in your own environment or under no-training terms, as on our enterprise AI page, and send ranges rather than exact concentrations where possible.

Accountability becomes unclear

Early signalNobody can say who approved a classification or on which data version.

MitigationName a qualified person per jurisdiction and store approvals against record versions.

Questions and answers

Does an AI-drafted safety data sheet change who is legally responsible for it?

No. Under REACH and the OSHA Hazard Communication Standard, the supplier placing the product on the market remains responsible for its safety data sheets, however they were produced. An AI system is part of that supplier’s own process, like authoring software, so you should be able to show who approved each classification and on which data.

How do poison center notifications and UFIs fit into an automated SDS workflow?

They draw on the same composition record. Hazardous mixtures within scope of CLP Annex VIII need a poison center notification and a unique formula identifier, generated from a company VAT number and a formulation number7. When the record changes, the workflow checks whether the notification needs updating or a new UFI is required, and keeps Section 1 of the SDS consistent with the label2.

Can we keep exact concentrations confidential on the SDS?

Partly. In the EU, Annex II allows concentration ranges in Section 3, and CLP Article 24 lets a supplier request an alternative chemical name for certain substances3. In the US, the 2024 HazCom update lets manufacturers withhold an exact concentration as a trade secret if they disclose a prescribed range8. Automation should apply these rules per jurisdiction, not one setting everywhere.

Should we replace our existing SDS authoring software to use AI?

Usually not. Established regulatory content systems often already hold the classification rules, phrase libraries and templates. AI adds most value around them: reading supplier documents in, assembling drafts, comparing revisions and mapping rule changes to affected products. Swapping a working rules engine for a language model trades reproducibility for convenience, the wrong exchange for legal documents.

Sources

  1. Regulation (EC) No 1907/2006 (REACH) — EUR-Lex · checked 10 October 2026
  2. Commission Regulation (EU) 2020/878 amending Annex II to REACH — EUR-Lex · checked 10 October 2026
  3. Regulation (EC) No 1272/2008 (CLP) — EUR-Lex · checked 10 October 2026
  4. Commission Delegated Regulation (EU) 2023/707 on hazard classes and criteria — EUR-Lex · checked 10 October 2026
  5. Regulation (EU) 2024/2865 amending the CLP Regulation — EUR-Lex · checked 10 October 2026
  6. Regulation (EU) 2025/2439 on dates of application and transitional provisions — EUR-Lex · checked 10 October 2026
  7. Poison Centres: harmonised information and the UFI — European Chemicals Agency · checked 10 October 2026
  8. Hazard Communication Standard, final rule (89 FR 44144) — US Occupational Safety and Health Administration, Federal Register · checked 10 October 2026
  9. Hazard Communication Standard: extension of compliance dates — US Occupational Safety and Health Administration, Federal Register · checked 10 October 2026
  10. About the GHS — United Nations Economic Commission for Europe · checked 10 October 2026

More in Chemicals

Back to Chemicals

Next step

Map one product family's SDS workflow before automating it

Send the composition records, current safety data sheets and target markets for one product family. We will show where classification, drafting and change monitoring can be automated and where your experts must stay in the loop.

Discuss an SDS workflow review