SDS Management AI: Extraction, Version Control, Purchasing Gates and RFP

Paints, solvents, cutting oils, adhesives, cleaning agents, plating chemicals: Japanese plants in Thailand use hundreds of them, and suppliers send the safety data sheets (SDS) piecemeal by email, paper, portal and USB, in Thai, English and Japanese, with no clear latest version. SDS management AI, which reads SDS automatically and turns them into a register, is what plants then consider.

This article makes one argument. Whether SDS management AI succeeds depends not on reading accuracy, but on whether you decided the following three things first. (1) Version identification: the rule for deciding which SDS is the valid version when the same chemical has several SDS that differ by supplier, language and revision date. (2) Where to draw the line on items that people verify: items that directly affect human safety if wrong, such as hazard classification, components with CAS numbers and concentrations, first-aid measures and accidental release measures, must not be finalized with the AI’s extracted values as they are. (3) Linkage with purchasing and goods receipt: a gate that blocks both ordering and receiving of any chemical that lacks a valid SDS or an adoption review. AI reads SDS fast. But it does not decide “which SDS is the correct one” or “whether this chemical may be used.”

Below, so that EHS (environment, health and safety) staff, quality assurance, purchasing and plant managers at plants that “use” chemicals, as well as environment and safety departments at head offices in Japan, can assemble the material they need for internal decisions, we go through regulatory developments, the differences between approaches, where SDS management breaks down, the items left to AI and the items people verify, update management, regulatory screening, purchasing linkage, product options, multilingual cautions, a model calculation of costs and benefits, RFP items, and PoC and acceptance testing, in that order. The model plant’s item counts, unit rates, hours, amounts and payback periods in this article are all placeholder figures set for this article. They are not industry averages, survey results or vendor prices. Also, the parts that touch on law are an introduction to publicly available information and are not legal advice.

Why review SDS management with AI now

An SDS is not a document you collect once and are done with. Each time standards and laws move, suppliers revise their SDS, and the register on the user side goes out of date. Here are the main developments of the past few years, seen from the position of a plant that receives SDS.

The UN GHS is at Revision 11

The UN GHS, the international framework for the classification and labelling of chemicals, has been revised repeatedly. On UNECE’s official GHS page (archived as of September 8, 2025), Revision 11 (2025) is listed as the latest, alongside Revision 8 (2019), Revision 9 (2021) and Revision 10 (2023). The amendments behind Revision 11 were adopted by the UN Committee of Experts on the GHS at its 12th session on December 6, 2024. The main amendments listed are clarification of the classification criteria for aerosols and chemicals under pressure, new guidance on skin sensitization classification using non-animal test methods, classification of substances and mixtures that contribute to global warming, rationalization of precautionary statements, and added guidance on simple asphyxiants. The same page also said that the electronic version of Revision 11 was scheduled to be published in September 2025.

When the GHS is revised, national laws do not adopt that revision right away. The fact that each country has adopted a different revision is what makes things difficult for the user side.

Thailand’s GHS notification follows Revision 3

In Thailand, under the Hazardous Substance Act B.E.2535 (1992), the Ministry of Industry notification “Hazard Classification and Communication System of Hazardous Substances B.E.2555 (2012)” put the GHS into the regulatory system. It was published in the Royal Gazette on March 12, 2012 and took effect the following day. It covers hazardous substances under the jurisdiction of the Department of Industrial Works (DIW), requires manufacturers and importers to carry out classification, labelling and SDS preparation, and set the deadline for full implementation at within “1 year for substances and 5 years for mixtures” from the effective date. It specifies that an SDS consists of 16 items.

This notification follows Revision 3 (2009) of the UN GHS, which is far removed from the latest Revision 11. Whether there is a successor notification that updates it to a newer GHS revision could not be confirmed from primary sources in the research for this article. It is entirely possible that SDS from local distributors in Thailand and SDS from manufacturers in Japan or Europe follow different GHS revisions or show different classifications.

For factories, the notification on safety measures for chemical management

On the Factory Act side, the Ministry of Industry notification “Safety Measures for Chemical Management in Factories B.E.2565 (2022)” sets obligations for Type 3 factories concerning chemical registers, reporting, labels and SDS. It was published in the Royal Gazette on April 25, 2022, and its entry into force was set at 180 days after the day following publication. Labels (Article 8) and SDS (Article 9) were set to take effect 1 year after the day following publication for single substances and 2 years after for mixtures. Counting from the publication date (April 25, 2022), even the latest of these, the 2 years for mixtures, had run out by April 2024, and as of October 2026 all of the grace periods have passed. The content is covered in detail in the regulatory screening section.

Japan’s SDS-related amendments under the Industrial Safety and Health Act have two effective dates

For environment and safety departments at head offices in Japan, and for plants that ship products to Japan, the SDS-related amendments to Japan’s Industrial Safety and Health Act are also relevant. Note that the SDS-related changes made by the May 2022 ministerial ordinance amendment and related measures take effect on two different dates.

According to the Q&A of Japan’s Ministry of Health, Labour and Welfare (MHLW), if the 5-year review confirms that no change is needed, there is no obligation to update the SDS or to notify the recipient. However, it is considered desirable to keep a record of the date of the review. In other words, an old revision date alone does not prove that an SDS “has not been reviewed.” This is another place where the weakness of a register that judges old and new by revision date alone shows up.

Format revisions in Europe and the US

In the EU, Regulation (EU) 2020/878, which revised REACH Annex II (the SDS format), has applied since January 1, 2021, with a transitional measure that allowed SDS in the pre-amendment format to continue to be provided until December 31, 2022. In the US, OSHA’s Hazard Communication Standard (HCS 2024) took effect on July 19, 2024, with phased compliance deadlines for reviewing SDS by substance and mixture. Plants that import chemicals from European or US manufacturers may receive SDS revised to match these format changes and deadlines.

A structure in which the register goes out of date with every revision

Putting these developments together, the structure is simple. Each time the GHS is revised, national laws are amended, or a manufacturer changes a formulation or reassesses a product, the supplier revises its SDS. The revised version may arrive as an email attachment, may simply be placed on a portal, or may be packed on paper with the next delivery. The Excel register and shared folder at the user-side plant are not updated unless someone notices that it has arrived and registers it again. With hundreds of items, from a few to several dozen suppliers, and 3 languages, a manual register tends to fall behind somewhere.

AI reading speeds up this registration work. However, if the cause of the delay is not that “reading is slow” but that “nobody notices it has arrived” or “nobody has decided which version is valid,” faster reading will not make the register correct. That is why this article argues that version identification and purchasing linkage should be decided before reading accuracy.

Note that in OSHA’s published list of the top 10 most frequently cited standards (citations from US federal inspections, preliminary data for fiscal year 2026, October 1, 2025 to August 31, 2026), Hazard Communication (general industry) ranked second. These are US figures and do not apply directly to plants in Thailand, but they do indicate that communication through SDS and labels is an area that tends to be cited in inspections.

What SDS management AI is: OCR, LLM extraction, SDS management systems and chemical management systems

The phrase “managing SDS with AI” mixes four mechanisms of different character. Unless you align on which one you mean, neither RFPs nor quotations can be compared.

In practice these are combined. For example, SDS on paper or as scanned images are turned into text by OCR, items are extracted by an LLM, the result is registered as a version in an SDS management system, and it is linked to the chemical master in a chemical management system or ERP and used for the purchasing and goods receipt gates. We covered OCR approaches and how to run an evaluation in “AI-OCR Comparison: RFP and PoC Guide for Thailand.” This article narrows the focus from forms in general to SDS, and centers on “versions,” “human verification” and “purchasing linkage.”

Mechanisms for tracking chemical lots through the process are covered in “Chemical Lot Management 2026 – Traceability Breaks Where Batches Merge.” Whereas lot tracking deals with “which lot flowed where,” the SDS management in this article deals with “which information about that chemical is treated as correct and communicated to the shop floor.”

Five places where safety data sheet management breaks down

Before choosing a product, check where your current SDS management is breaking down. The ways it commonly breaks down at Japanese-affiliated plants in Thailand fall into the following five.

(1) Scattered receiving channels

SDS are received by many different people: purchasing staff, EHS staff, quality assurance, shop-floor team leaders. A revised version that a supplier’s sales rep sends as an attachment to someone’s personal email will not reach the register unless that person notices and forwards it. The Q&A of Japan’s MHLW also allows SDS updates to be handled by making the latest version always viewable through a URL or QR code given at the time of provision. This means that some SDS from Japanese suppliers may not arrive by email when revised; only the version on the web may be updated. When taking inventory of receiving channels, record not only “where it arrives” but also “where to go to find the latest version.”

(2) Version identification

For the same paint, there may be an SDS issued by a local distributor in Thailand, a Japanese-language SDS that came via a Japanese trading company, and an English SDS from an overseas manufacturer, side by side. The product name is the same but the product codes differ slightly. All three have different revision dates. If you ask AI “which is the current correct SDS” in this situation, the AI has no basis for the judgment. People need to decide the conditions for a valid version in advance. The next section on update management sets out concrete rules.

(3) Multiple languages

At plants in Thailand, the SDS for the same chemical may come in 2 languages, Thai and English, or 3 including Japanese. If the revision dates differ by language, you cannot link them in the register as “different languages of the same version” until you confirm whether they are translations of the same content or separate revisions.

(4) Composition tables and concentrations

SDS item 3 (composition/information on ingredients) is written as a table and often spans several pages. If the rows for component name, CAS number and concentration are misaligned when read, the concentration of a different substance gets recorded. Concentrations are not written uniformly either. For SDS under Japan’s Industrial Safety and Health Act, stating a numerical value in percent by weight became the rule from April 1, 2024, but where concentration varies because of the nature of the product, a reasonable concentration range may be given, and where it is a trade secret, it may be stated in 10% steps after stating that fact. SDS in the pre-amendment format still contain range notation such as “10-20%.” Unless you decide how the register holds numbers and ranges, discrepancies appear in totals and regulatory screening.

(5) Mismatch with shop-floor postings

Article 9 of Thailand’s B.E.2565 notification (which applies to Type 3 factories under the Factory Act) requires that relevant workers can access the Thai-language SDS and that the key points needed for their work are communicated. Many plants place SDS summaries or postings on the shop floor. Even if the SDS in the register is replaced with a new version, if the shop-floor postings still carry the old version’s content, what the shop floor is receiving is old information. The DIW checklist also includes the check item “Is the SDS up to date?” The design needs to record the register update and the replacement of shop-floor postings as a single task.

The general approach to storage, versions and approval for documents is covered in “Document Management System for Factories — 3 Axes to Compare.” An SDS is also one of a plant’s documents, but it differs from internal documents such as work standards in that it is an external document whose version is decided by the supplier and which you cannot revise yourselves.

AI SDS data extraction: items left to AI and items people verify, in four tiers S0 to S3

The most important design decision when using AI to read SDS is where to draw the line between extracted values that are finalized in the register as they are and values that people verify.

Generative AI carries its own risks. The US National Institute of Standards and Technology (NIST) generative AI profile (NIST AI 600-1, July 2024) lists “confabulation” as one of the risks specific to generative AI and defines it as producing confidently stated but erroneous content. This is what is commonly called hallucination. The document points out that this includes outputs that diverge from the input, that it arises naturally from the design of generative models, that it is particularly problematic in highly context-dependent domains requiring specialist knowledge, and that monitoring becomes especially important when it is built into uses that involve consequential decisions. SDS item extraction is exactly the kind of use where “output that diverges from the input (the original SDS)” is the problem.

In addition, OWASP’s “Top 10 for LLM Applications 2025” lists insufficient validation and handling of LLM output (LLM05:2025 Improper Output Handling) and the risk that an LLM generates misinformation and people over-rely on it (LLM09:2025 Misinformation). A design that sends AI-extracted values straight into the register or shop-floor postings without verification falls under both.

So the items extracted from an SDS are divided into the following four tiers.

There are three key points in drawing the line.

First, the S2 items are not few. For mixtures with many rows in the composition table, human verification takes time. The model calculation assumes 0.5 hours for “AI extraction plus human verification” because it presupposes this verification. Even if the extraction accuracy or processing time a vendor presents is high, that is no reason to skip human verification of S2.

Second, build the S1 “difference check” as a system function. Whether values for physical properties, storage conditions, disposal and transport classification have changed from the previous version is checked not by AI judgment but by mechanical comparison with finalized data. Items with changes go to a person even though they belong to S1. The S0 product code and supplier name are also mechanically matched against the chemical master, and if they do not match, a person confirms the version identification.

Third, do not let AI judge S3 and do not let it summarize S3 either. If you have AI answer in prose “May this chemical be used under Thai regulations?” or “Does it fall under the customer’s prohibited substances?”, the answer gets treated like a basis for the judgment even when it is a plausible error. Regulatory screening is done by mechanically matching the finalized S2 components and CAS numbers against regulatory lists and customer prohibited-substance lists that you manage yourselves, and a person decides after looking at the result.

Depending on the unit in which AI-OCR accuracy is measured (characters, items or documents), the figures for the same product change greatly. How to measure is covered in detail in “AI OCR Accuracy: PoC and Acceptance Testing for Multilingual Documents.” For SDS, the practical approach is to measure separately the item-level accuracy rate limited to S2 items, and whether errors were stopped by the S1 difference check or human verification.

SDS version control and update management: version identification, update requests to suppliers, retiring old versions

The second thing to decide before reading with AI is how SDS versions are handled.

Decide the conditions for a valid version

Decide the conditions for a valid version by combining at least the following five elements.

• Product name
• Supplier (who you buy from)
• Product code (the manufacturer’s part number or the code in your own chemical master)
• Revision date (or version number)
• Language

Even with the same product name, a different supplier is held as a separate row, because the author of the SDS and the timing of revisions can differ from distributor to distributor. Different languages are linked as different languages of the same version. For example, if the English version’s revision date is March 2026 and the Thai version’s is May 2026, confirm with the supplier whether the Thai version is a translation of the English one or a separate revision, and record the result in the register.

Retire old versions

When a new version becomes valid, keep the previous version as “retired.” Do not delete it, because in customer audits or checks by the authorities you may need to explain which version was shown on the shop floor at a given time. At the same time as the retirement operation, record the following tasks as a single procedure.

• Replacing shop-floor postings and summaries (hold in the register where they are and how many there are)
• Switching the SDS viewed on shop-floor terminals and tablets
• Where hazard classification, components or protective equipment changed, notifying the relevant departments and, as needed, reviewing work procedures and protective equipment

If an in-house generative AI search (RAG) refers to SDS, old versions must not remain in its search results. The problem of answers drawn from old versions of documents, and how to prevent it, is covered in “Factory RAG Document Version Control: Implementation and Acceptance.”

Request updates from suppliers

If you only wait for revised versions to arrive, the register falls behind. Decide on operations such as the following.

• Periodic confirmation requests: Once a year or similar, send suppliers a list of the revision dates of your valid versions and ask them to confirm whether there is anything newer.
• Checking SDS with old revision dates: As with the review at least once every 5 years under Japan’s Industrial Safety and Health Act, an SDS can have an old revision date and still have had its content reviewed. Do not retire an SDS just because its revision date is old; record the date you confirmed with the supplier and the result.
• Obtaining SDS at new adoption: At the adoption review, check whether a Thai-language SDS is available (the requirement in Article 9 of Thailand’s B.E.2565 notification is covered in the next section).
• Handling SDS provided via URL or QR code: For SDS where the version on the web is updated, save the file and the date you obtained it on your side, and go back periodically to check.

If you already gather supplier communication on a portal, putting SDS submission and update requests on the same mechanism reduces the scattering of receiving channels. For an approach to linking purchase orders and delivery commitments, see “Implementing a Supplier Portal for Thai Factories: PO Acknowledgment and Delivery Commitments.”

GHS compliance in Thailand and regulatory screening: what to check in Thailand, Japan, the EU, the US and Vietnam

This section introduces the laws and standards in each country that relate to SDS management, to the extent we could confirm them in published primary sources. This is not legal advice. Please confirm how they apply to your company with specialists or the authorities. The English translations of Thai laws are unofficial translations published by JETRO; only the Thai original is legally binding.

Thailand: the B.E.2565 notification on the Factory Act side (Type 3 factories)

The Ministry of Industry notification “Safety Measures for Chemical Management in Factories B.E.2565 (2022)” is based on the Factory Act and applies to Type 3 factories. Summarizing from the Thai original, the main obligations relevant to SDS management are as follows.

For both Article 8 and Article 9, the notification provides that equivalent compliance under other laws is deemed to satisfy the article.

Note that the only requirement for a Thai-language SDS that this article could confirm is Article 9 of this B.E.2565 notification on the Factory Act side. We found no SDS language requirement in the English translation of the GHS notification (B.E.2555) on the Hazardous Substance Act side. The two notifications differ in their legal basis and scope, so take care not to confuse them.

The DIW’s “safety assessment checklist for chemical management” for the B.E.2565 notification contains check items such as “Does the SDS have all 16 items?”, “Is there an SDS for every hazardous chemical in the work area?”, “Is the SDS up to date?” and “Is the SDS kept on paper or electronically?” These are useful pointers when thinking about requirements for an SDS management system.

Thailand: notifications on the Hazardous Substance Act side

The Hazardous Substance Act B.E.2535 has since been amended three times (No. 2 B.E.2544, No. 3 B.E.2551, No. 4 B.E.2562). The Ministry of Industry notifications closely related to SDS management are as follows.

• GHS notification B.E.2555 (2012): As described above, it requires manufacturers and importers of hazardous substances under DIW jurisdiction to carry out classification, labelling and SDS preparation. Manufacturers, importers, exporters and possessors have an obligation to communicate hazards through labels and SDS.
• Declaration notification B.E.2547 (2004) and No. 2 B.E.2563 (2020): These set out declarations by manufacturers, importers, exporters and possessors of hazardous substances under DIW jurisdiction. Importers and exporters declare on the prescribed form before customs clearance, and declare quantities of specified substances and so on every half year. The No. 2 notification changed this to declarations where more than 100 kg of a substance is handled per half year.
• Storage notification B.E.2551 (2008): It requires safety measures such as storage locations, segregation for storage and preventive measures, to be carried out in accordance with the DIW storage manual or international standards accepted by the DIW.

Which notification applies to your company, and in which capacity (manufacturer, importer, possessor and so on), depends on the chemicals you handle and the form of the transactions. In an SDS management mechanism, the prerequisite for that judgment is holding the CAS numbers of components and the annual quantities handled as finalized data, in a state that can be used for screening and totaling.

Japan: the head office and Japan-bound perspective

Under Japan’s Industrial Safety and Health Act, an SDS has 12 items. From April 1, 2024, “assumed uses and precautions for use in those uses” was added, and through an amended administrative notice, the type of protective equipment for the assumed uses must now always be stated in the protective equipment field of “precautions for storage or handling.” It is said that an SDS prepared in line with JIS Z 7253 covers these items. The review at least once every 5 years (effective April 1, 2023) and percent-by-weight concentrations (effective April 1, 2024) are as summarized at the start. Promptly notifying the recipient of changes is treated as a best-effort obligation under Article 57-2, paragraph 2 of the Act.

The “GHS Guideline: Guide to Preparing Labels and Safety Data Sheets,” edited by the Japan Chemical Industry Association and published by the Japanese Standards Association, is in its February 2026 edition and corresponds to JIS Z 7253:2025, which was revised to adopt Revision 9 of the UN GHS published in 2021. SDS from Japanese suppliers may be revised in turn to versions in line with this new JIS.

Japan also has an SDS provision system under the PRTR Act administered by the Ministry of Economy, Trade and Industry. Because we could not confirm primary sources for it, this article does not go into detail.

On March 31, 2025, the MHLW published an item definition document for its “SDS data exchange format” and a usage manual for system developers, and it subsidizes part of the cost of introducing or modifying systems that support this format (the SDS digitization subsidy). Applications for fiscal 2026 are accepted from July 1 to November 30, and acceptance will stop even before the deadline if the budget limit is reached. The subsidy is one half of eligible costs, with a cap of 1 million yen per fiscal year. The conditions for eligible systems include handling electronic SDS in the standard format or having “a function to read paper or PDF SDS” and restoring them as legible SDS, being able to output in the standard format, and having a function that prompts input where an SDS does not comply with the Act. Eligible applicants are small and medium-sized enterprises in Japan; this article has not confirmed whether a Thai entity, or its head office in Japan, would be eligible. We mention it as information for cases where the head office in Japan reviews its SDS management mechanism.

EU: SDS format and CLP

In the EU, Regulation (EU) 2020/878, which revised REACH Annex II (the SDS format), has applied since January 1, 2021, and the transitional measure allowing SDS in the old format to be provided ended on December 31, 2022.

The CLP Regulation, which governs classification and labelling, was amended by Regulation (EU) 2024/2865. It was adopted on October 23, 2024, published in the Official Journal on November 20, and entered into force on the 20th day after publication. Its main content covers label formatting (such as minimum font size), deadlines for relabelling when classification changes, and labelling obligations in advertising and distance sales; where a new hazard class is added or a classification is changed to a more severe one, labels must now be updated within 6 months. The original application dates were July 1, 2026 and January 1, 2027, but Regulation (EU) 2025/2439, published in the Official Journal on December 3, 2025, postponed the application of some provisions concerning label formatting, relabelling, advertising and distance sales to January 1, 2028. Some provisions that apply on July 1, 2026 and January 1, 2027 remain. The CLP amendment is mainly about labels and classification, not an amendment to the SDS format. However, when classification changes, the hazard items of the SDS are also revised, so for chemicals imported from the EU, watch for revised versions arriving.

US: OSHA HCS 2024

OSHA’s HCS 2024 final rule was published on May 20, 2024 and took effect on July 19 of the same year. On January 15, 2026, a final rule extending the compliance deadlines by 4 months was issued.

For chemicals imported from US manufacturers, SDS revised to match these deadlines may arrive.

Vietnam: the new Law on Chemicals

In Vietnam, the Law on Chemicals (No. 69/2025/QH15) took effect on January 1, 2026, replacing the 2007 Law on Chemicals. The implementing Decrees 24/2026/NĐ-CP (list of chemicals), 25/2026/NĐ-CP (development of the chemical industry and chemical safety and security) and 26/2026/NĐ-CP (management of chemical activities and hazardous chemicals in products and goods), together with Ministry of Industry and Trade Circulars 01/2026/TT-BCT and 02/2026/TT-BCT, were all issued dated January 17, 2026 and took effect on the same day. This article has not confirmed the SDS format or language requirements under the new law. Companies with sites in both Thailand and Vietnam should confirm the Vietnamese requirements separately.

How to think about turning regulatory screening into a mechanism

Even within the scope introduced in this article, the requirements in each country differ in legal basis, who is covered, revision and deadlines. In an SDS management mechanism, rather than having AI produce the “conclusion” of regulatory screening (S3), you version-control the regulatory lists used for screening yourselves (Thai hazardous substance categories, substances subject to notification for Japan, customer prohibited-substance lists and so on), mechanically match finalized components and CAS numbers against them, and have a person check the hits and record the decision. Because the regulatory lists themselves are amended, also keep a record of which version of the list was used for screening.

Chemical management system linkage with purchasing, goods receipt and ERP: a gate that keeps out chemicals without an SDS

The third thing to decide before reading with AI is the linkage with purchasing and goods receipt. However accurate the register is, if a new chemical is ordered and received without an adoption review, its SDS will not be in the register.

Where to place the gates

Place gates at the following three points.

Linking the chemical master and SDS

In an ERP or purchasing system, a chemical is registered as one row in the item master. To connect it with the SDS management mechanism, give each row of the item master “whether it is a chemical,” “the adoption review status” and “the identifier of the valid SDS version,” so that they can be referenced at ordering and at goods receipt. Where the same item is bought from several suppliers, link a valid SDS version for each supplier.

At plants using a WMS (warehouse management system), one possible operation is to display the revision date of the valid SDS version on the receiving inspection screen and, if it differs from the revision date of the SDS packed with the delivery note, pass it to EHS staff on the spot. General requirements for procuring a production management system at chemical plants are covered in “Chemical Production Management System 2026: RFP Requirements for Formulas, Quality and Lots,” but at plants that “use” chemicals, the purchasing and goods receipt gates take effect sooner than production management.

Handling exceptions

Once you place gates, requests for exceptions will certainly come, such as “the SDS will not be ready in time for an urgent sample evaluation” or “the supplier has not issued a Thai version yet.” Document in advance the conditions for allowing an exception (approver, time limit, upper limits on place of use and quantity, the information used in the interim), and track the number of exceptions and the days until they are resolved as indicators. If exceptions become routine, the gate becomes a formality.

Safety data sheet management software: product and approach options

SDS management mechanisms fall broadly into four families. The features and figures below are all vendor claims stated on each company’s official pages or press releases, and have not been verified by third parties. This is not a recommendation of any specific product.

1. SDS management SaaS

Cloud services centered on SDS storage, version management, search and AI item extraction.

• SDS Manager (Norway): [Vendor claim] It says AI can extract data from SDS in more than 30 languages, that the extraction covers all 16 sections of the SDS, and that pictograms and protective equipment icons are also identified by optical recognition. No figures for extraction accuracy are given, and we found no description of a human verification step. The pricing page lists a free plan for SDS management (up to 200 SDS), pay-per-use for its SDS authoring service (199 US dollars per SDS), Standard (1,499 US dollars a year for 15 SDS), and EHS features (9.99 US dollars per employee per month, 1 to 200 employees). SDS authoring appears to be a service used mainly by those who make chemicals.
• Quentic (Germany): [Vendor claim] It says AI can automatically extract key information from SDS regardless of format or language, that version management tracks every change, and that automatic reminders and automatically generated inquiry messages to suppliers keep SDS up to date. Prices are not published.

2. EHS suites

SDS management and chemical management offered as part of a product family covering EHS work in general.

• VelocityEHS (US): [Vendor claim] It says it has an SDS library of millions of documents and adds more than 20,000 new and revised SDS every week. It also advertises an “AI Ingredient Indexing” feature and support for more than 15 languages. Prices are not published.
• I&D “ESH Clouds Chemicals Management” (Japan): [Vendor claim] In an announcement on December 19, 2024, it said the accuracy of the AI-OCR that reads SDS had “improved to 90%” and that SDS system registration effort could be reduced by up to 70%. No definition or measurement conditions were given, such as whether this 90% is character-level or item-level, so it gives no basis for comparison with other vendors’ figures. The published prices (excluding tax), assuming a configuration of a head office plus 3 sites, 6 administrators and 200 users, are from 3.815 million yen in the first year (an implementation fee of 840,000 yen plus an annual system usage fee from 2.975 million yen) and from 2.975 million yen from the second year onward.

Note that Dynagon in Japan has announced that its “Smart SDS” was certified for the second year in a row as an eligible system for the MHLW’s SDS digitization subsidy (July 1, 2026). [Vendor claim] It says SDS authoring is “completed by entering 5 items,” and this too appears to be a product for those who make SDS. When a user-side plant chooses an SDS management product, first check whether it is aimed at authoring or at management.

3. General-purpose LLM plus in-house build

Using a general-purpose generative AI API to extract items from SDS and register them in your own register or database. It is flexible, but you have to design and verify everything yourselves: the S0 to S3 line, the difference check, the human verification screens, version identification, logs and purchasing linkage.

4. AI-OCR plus existing document management

Adding SDS reading definitions and a register to the AI-OCR and document management system you already use. It avoids adding a new product, but you have to build in yourselves the handling of SDS-specific items (composition tables, hazard classification, completeness of the 16 items) and the version identification rules. For how to compare AI-OCR costs as a whole rather than by API unit price alone, see “AI OCR Pricing 2026: Compare the Six-Layer TCO.”

How to choose

Whichever family you choose, the “supported languages” and “extractable items” a vendor cites are unknown until you actually try them with Thai-language SDS. Confirm them in a PoC using SDS that have actually reached your company.

Cautions when handling SDS in Thai, English and Japanese

SDS management at plants in Thailand involves multilingual circumstances that head office examples do not.

Link different languages of the same version. For the same chemical you will have an English version and a Thai version, and sometimes a Japanese version. In the register, do not register each language as a separate item; link files in several languages to one valid version. Where revision dates differ by language, record the result of confirming with the supplier whether it is a translation or a separate revision, as described above. For AI extraction, an effective design is to match the S2 items (hazard classification, components, CAS numbers, concentrations) extracted from each language against each other and send any discrepancy to a person.

Decide how translated SDS are handled. Where a supplier has not issued a Thai version, a plant may use an SDS or summary translated into Thai by itself or a translation company. Article 9 of the B.E.2565 notification requires keeping a Thai-language SDS or Thai-language chemical safety instructions. Decide what status your own translations or summaries have only after confirming with specialists or the authorities, and in the register hold SDS issued by suppliers separately from translations you made yourselves. If you translate into Thai with generative AI, have staff who can read Thai check the S2 items for mistranslation.

GHS pictograms and signal words are clues that depend little on language. GHS pictograms and signal words are items whose correspondence is easy to check even across languages. If pictograms or signal words do not match between SDS in different languages, it may be a separate revision, or the documents may follow different GHS revisions or classifications. As noted above, Thailand’s GHS notification follows Revision 3 and Japan’s JIS Z 7253:2025 follows Revision 9, so the same chemical can be classified differently in SDS for different countries. Do not assume either one is wrong; confirm with the supplier.

Item structure of Japanese SDS. Under Japan’s Industrial Safety and Health Act there are 12 items, which are counted differently from the 16 items set by Thailand’s GHS notification. The MHLW Q&A says that an SDS prepared in line with JIS Z 7253 covers the items under the Act. Building the AI extraction definitions on the headings of the 16 items in the Thai notification, and keeping a mapping table of Japanese, English and Thai headings, makes cross-language matching easier.

Scan quality. When SDS that arrived on paper are scanned, Thai tone marks and small characters can be crushed. Some paper SDS also have handwritten notes on them. How to handle the scan resolution and skew correction that reading depends on should be checked in a PoC. For an approach to reading handwritten and multilingual forms, “Handwritten OCR in Thailand: PoC, RFP and Acceptance for Multilingual Forms” is also a useful reference.

Model calculation of costs and benefits (placeholder figures)

From here, we look at the costs and benefits of introducing an SDS management mechanism at a model plant used to show the structure of the calculation. The item counts, ratios, hours, unit rates and amounts below are all placeholders for this article. They are neither vendor prices nor survey figures. Please replace them with your own figures.

Model plant assumptions

One site of a Japanese-affiliated manufacturer in Thailand (automotive parts, with a painting process), using 600 chemical items, with an EHS staff labor rate of 400 THB per hour. The currency is Thai baht (THB). We assume that revised SDS are received each year for 20% of items and that newly adopted chemicals amount to 10% of items, both proportional to the number of items.

Items (1) to (4) are set as tasks that do not overlap one another, so the same time is not counted twice. In (1), the 120 revisions are 600 × 20% and the 60 new items are 600 × 10%. The saving in (3) is 15 minutes − 3 minutes = 12 minutes = 0.2 hours.

The “0.5 hours with AI extraction plus human verification” in (1) is a placeholder. How long human verification of S2 items takes varies with the number of components, the language and scan quality. Do not use the extraction accuracy or processing time a vendor presents as is; replace it with the time actually measured in a PoC using your own SDS.

How the payback period moves when the number of items changes

With N as the number of items, the annual hours saved can be expressed as follows.

• (1) 0.3N hours (0.2N revisions + 0.1N new, 1.0 hour saved per case)
• (2) 0.15N hours (0.1N new × 1.5 hours)
• (3) 0.2N hours (N cases × 0.2 hours)
• (4) 48 hours (independent of item count)

In total, annual hours saved = 0.65N + 48 hours, annual benefit = 400 × (0.65N + 48) THB, and initial cost = 270,000 + 300 × N THB. The annual running cost is assumed fixed at 120,000 THB. The results of changing only the number of items N are as follows.

At a 300-item plant, the annual benefit of 97,200 THB falls below the annual running cost of 120,000 THB, and the annual net benefit is negative. Labor-hour savings cannot cover even the yearly running cost, let alone the initial cost. Net benefit reaches 0 when 0.65N + 48 = 300 (120,000 ÷ 400), that is, when N ≈ 388. Under these placeholder assumptions, below about 390 items, labor-hour savings alone cannot cover even the annual running cost. On the other hand, at a 1,200-item plant, payback is about 3.0 years. You can see that the payback period moves greatly each time the number of items doubles.

What this calculation does not include

• The effect of preventing accidents and occupational injuries caused by old SDS versions or wrong hazard information (because the probability of occurrence cannot be estimated)
• The effect of preventing findings by the authorities and nonconformities in customer audits
• The effect of stopping unapproved chemicals from getting in
• Sharing across multiple sites (the effect of reusing the same supplier’s SDS between sites)
• What the freed-up EHS staff time is used for. Freed-up time does not become a cash saving unless it is redirected to other work

Looking at labor-hour savings alone, payback depends strongly on the number of items, and at plants with a few hundred items it takes several years or does not happen at all. In investment decisions, it is reasonable to evaluate “assurance of valid versions” and “a gate that keeps out unapproved items” as investments in risk management, separately from labor-hour savings. This article does not put a monetary value on those effects, but do not exclude them from the evaluation just because they cannot be expressed in money.

Plants with fewer items have another option: first put in place only the register, the version identification rules and the purchasing gate, and use AI only for bulk extraction at the time of register migration. In this case you reduce only the effort of building the initial register, without permanently keeping a mechanism that incurs annual running costs.

RFP items for SDS management AI

In the RFP, write about versions, human verification, purchasing linkage, measurement and testing before asking about the AI’s reading accuracy.

Metrics should look not only at the AI’s extraction accuracy but at whether SDS are being kept correct across the mechanism as a whole. For example, tracking “valid-version coverage rate (the share of all items in the chemical register that have a valid SDS version linked),” “days to reflect revisions (days from receiving a revised version to reflecting it in the register and shop-floor postings)” and “correction rate in human verification (the share of S2 items corrected during human verification)” makes it easier to tell whether a problem lies with the AI or with operations.

What to check in PoC, FAT and SAT: acceptance tests that deliberately try things

In the PoC and acceptance tests, check not only whether the AI can read clean SDS correctly, but whether, when given broken or confusing SDS, it selects the valid version correctly, triggers human verification, stops unapproved items and leaves logs. To do this, deliberately create and run the confusing cases that actually occur on the shop floor. Split testing into tests in the vendor’s environment (FAT) and tests in an environment connected to your own ERP and purchasing systems (SAT).

Six cases to try deliberately

• Old version mixed in: Register the old and new versions of the same chemical with the order of receipt dates swapped. Is the old version selected as the valid version?
• Thai and English versions of the same product: Register a Thai version and an English version with different revision dates. Are they double-registered as separate items? Are they linked as different languages of the same version? Are those that need confirmation sent to a person?
• Poor scans and skew: Run skewed scans, low-resolution scans and paper SDS with crushed Thai tone marks. Does it avoid treating unreadable parts as read, and send them to human verification?
• Broken composition tables: Run SDS whose composition tables span several pages and PDFs with misaligned rows and columns. Does it avoid mixing up the combinations of component name, CAS number and concentration?
• Range notation for concentrations: Mix range notation such as “10-20%,” numerical percent by weight, and values written in 10% steps as trade secrets. Does it avoid rounding a range to a single number or swapping the upper and lower limits?
• Orders for unapproved chemicals: Try to order and receive items whose adoption review is not finished or that have no valid SDS version. Are they stopped at the gate, and do exceptions pass only with approval and a time limit?

Five items to check

For each case, check the following five items.

Building a table that crosses the 6 cases with the 5 items, and recording pass or fail and evidence in each cell, makes it easier to compare FAT and SAT results. Not every case relates to every item. For example, “orders for unapproved chemicals” centers on order blocking and log records, while “broken composition tables” centers on item extraction and triggering human verification.

Division between FAT and SAT

In FAT, run all 6 cases once in the vendor’s environment and check extraction, triggering of human verification and log items. In SAT, with your ERP, purchasing and WMS connected, run the same cases using test items and check whether the ordering and goods receipt gates actually stop things, whether old versions are not shown on the shop-floor viewing screen, and whether exception approvals are recorded. At the PoC stage, select SDS that have actually reached your company so that they are spread across languages, suppliers and paper versus digital, measure the time taken for human verification in (1), and replace the 0.5 hours in the model calculation. Even after go-live, run some of the cases periodically to confirm that changes to settings or integrations have not broken the gates.

In workplaces that handle chemicals, SDS information is sometimes viewed together with workplace environment measurement results. For measurement frequency and continuous IoT monitoring, see “Workplace Environment Measurement | Thai Factories 2026.”

Frequently asked questions

What becomes easier when you use AI for SDS management?

The main things that become easier are registering and updating SDS in the register, checking the versions of all items, and finding the SDS needed during audits and inspections. AI drafts the work of pulling items such as product name, revision date and components out of SDS whose format differs from supplier to supplier. On the other hand, deciding which version is valid, finalizing hazard classification and components, and deciding whether a chemical may be used remain as things people decide. In this article’s model calculation (placeholders), we assumed 438 hours saved per year at a 600-item plant, but please measure the actual value in a PoC.

Can hazard classifications read by AI be used as they are?

Finalizing them as they are should be avoided. Hazard classification, components with CAS numbers and concentrations, first-aid measures, accidental release measures, protective equipment and occupational exposure limits directly affect human safety if wrong. A NIST document also points out that generative AI carries the risk of confidently producing wrong content. As with the S2 tier in this article, treat AI-extracted values as drafts, have a person check them against the original and finalize them, and record the verifier and the date and time.

How much does an SDS management system cost?

It varies greatly with the product family, the number of items, the number of sites and the billing unit. As examples of published prices, SDS Manager offers a free SDS management plan up to 200 SDS, and I&D’s ESH Clouds Chemicals Management lists from 3.815 million yen in the first year (excluding tax) for a configuration of a head office plus 3 sites. Both are vendors’ published information. When comparing, align quotations to include not only licenses but also register migration, human verification effort, purchasing linkage and the PoC.

Is a Thai-language SDS mandatory?

Article 9 of the Ministry of Industry notification B.E.2565 under the Factory Act requires Type 3 factories to keep a Thai-language SDS or Thai-language chemical safety instructions for chemicals. There is also a provision that equivalent compliance under other laws is deemed to satisfy the article. Please confirm with specialists or the authorities whether your plant is covered and which measures satisfy the requirement. The content of this article is not legal advice.

From how many items should SDS management be systematized?

There is no single answer. In this article’s model calculation (placeholders), the result was that below about 390 items, labor-hour savings alone cannot cover even the annual running cost. However, the calculation does not include the effect of preventing accidents or nonconformities in audits. For plants with fewer items, one option is to first put in place the register, the version identification rules and the purchasing gate, and use AI only for bulk extraction at the time of register migration.

Where should SDS update management start?

With an inventory of chemicals and SDS. List items, suppliers, languages, revision dates, storage locations and using departments, and identify duplicates, gaps and old versions. Then decide the conditions for a valid version, retire old versions and replace shop-floor postings. Introducing AI after this inventory and the version identification rules are settled makes its effect easier to measure.

Summary

If SDS management AI starts with comparing reading accuracy, you may end up reading fast while the correct SDS never stays in the register. First take inventory of chemicals and SDS, decide the valid version by product name, supplier, product code, revision date and language, retire old versions and replace shop-floor postings too. Divide extraction items into S0 to S3; for S2 items such as hazard classification, components and first-aid measures, do not finalize AI-extracted values as they are, and for the conclusions of regulatory screening and decisions on whether a chemical may be used (S3), do not let AI judge or summarize. Then place a gate so that chemicals lacking a valid SDS or an adoption review can be neither ordered nor received. Deciding in this order makes it clear what to buy and what you need to decide yourselves.

On the regulatory side, the UN GHS Revision 11, Thailand’s B.E.2565 notification, Japan’s Industrial Safety and Health Act amendments (in two stages, April 1, 2023 and April 1, 2024), the EU CLP amendment and the US HCS 2024 keep things moving, and SDS revisions will keep arriving. In the model calculation (placeholders), looking only at labor-hour savings, payback depended strongly on the number of items: about 8.2 years at 600 items, with no payback at 300 items. That is exactly why, in the PoC and acceptance tests, you should deliberately run old versions mixed in, different languages, poor scans, broken composition tables, range notation and orders for unapproved items, and evaluate the assurance of valid versions and a gate that keeps out unapproved items as an investment in risk management separate from labor-hour savings.

TOMAS TECH helps Japanese-affiliated plants in Thailand digitize documents using AI-OCR and business systems, tailored to their operations. You can consult us even before the product or approach has been decided, for example at the stage of taking inventory of chemicals and SDS or deciding the version identification rules. Please contact us through this form.

References

• UNECE:”>https://unece.org/transport/dangerous-goods/ghs-rev11-2025″>UNECE: Globally Harmonized System of Classification and Labelling of Chemicals (GHS Rev. 11, 2025) (archived as of September 8, 2025)
• JETRO: Notification of Ministry of Industry: Hazard Classification and Communication System of Hazardous Substances B.E. 2555 (2012) (unofficial English translation)
• JETRO: Revised Hazardous Substances Act and Regulations
• Chulalongkorn University SHECU: Ministry of Industry notification on Safety Measures for Chemical Management in Factories B.E.2565 (Royal Gazette copy, April 25, 2022, in Thai)
• Department of Industrial Works (DIW): Safety assessment checklist for chemical management under the B.E.2565 notification (in Thai)
• JETRO: Notification of Ministry of Industry: Declaration of the hazardous substances B.E. 2547 (2004) (unofficial English translation)
• JETRO: Same notification (No.2) B.E. 2563 (2020) (unofficial English translation)
• JETRO: Notification of Ministry of Industry Re: Storage of hazardous substance B.E. 2551 (2008) (unofficial English translation)
• Ministry of Health, Labour and Welfare (Japan): Q&A on chemical substance measures, labels and SDS (updated February 28, 2024, in Japanese)
• Japan Chemical Industry Association / Japanese Standards Association: GHS Guideline, Guide to Preparing Labels and Safety Data Sheets (February 2026, preview, in Japanese)
• Ministry of Health, Labour and Welfare (Japan): Publication of the standard format for SDS information exchange (in Japanese)
• Japan Organization of Occupational Health and Safety, Hyogo Occupational Health Support Center: FY2026 SDS digitization subsidy leaflet (in Japanese)
• EUR-Lex: Commission Regulation (EU) 2020/878 of 18 June 2020
• EUR-Lex: Regulation (EU) 2024/2865 of 23 October 2024
• EUR-Lex: Regulation (EU) 2025/2439 of 26 November 2025
• Bao Dien tu Chinh phu (Vietnam Government News): Implementing documents for the new Law on Chemicals issued (February 16, 2026, in Vietnamese)
• OSHA: Hazard Communication – Rulemaking
• Federal Register: Hazard Communication Standard; extension of compliance dates (January 15, 2026)
• OSHA: Top 10 Most Frequently Cited Standards (updated October 5, 2026)
• NIST: AI 600-1 Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (July 2024)
• OWASP GenAI Security Project: OWASP Top 10 for LLM Applications 2025
• SDS Manager: Extract data from Safety Data Sheet (SDS)
• SDS Manager: Pricing
• Quentic: Safety data sheet management
• VelocityEHS: SDS Management
• I&D: ESH Clouds Chemicals Management press release (December 19, 2024, in Japanese)
• Dynagon: Smart SDS certified as an eligible system for the MHLW SDS digitization subsidy for the second year in a row (July 1, 2026, in Japanese)