Safety Data Sheet Management: Keep SDS Files Current With AI Help

Keep one register with a row for every product: the sheet's version, its revision date, the manufacturer, where the file lives and who last checked it. Let an AI assistant read each PDF into that table, flag sheets that look stale and find the right file when a customer asks. A named person must confirm every row against the actual document.

The division of labour matters more than the tool. AI is good at reading 200 PDFs into a table and noticing that 31 of them have a revision date older than five years. It is not a safe source of hazard, first-aid or disposal advice, and it can mix up two near-identical product names. So it builds the index, and a person owns the truth of each line.

Follow me on Instagram@sagnikteaches

Why an SDS folder goes stale without anyone noticing

A safety data sheet (SDS) is the document a manufacturer or supplier issues for a chemical product, covering what it is, its hazards, how to handle and store it, what to do after an exposure, and how to dispose of it. The internationally agreed format has 16 sections, from identification of the product and company through hazards identification, composition, first aid, firefighting, accidental release, handling and storage, exposure controls, physical and chemical properties, stability, toxicology, ecology, disposal, transport and regulatory information, to a final section for other information, which includes the date of the latest revision.

Connect on LinkedInSagnik Bhattacharya

A supplier does not write most of these. Manufacturers write them and send new versions when a formula changes, when classification rules change or when new hazard data appears. Some sheets are revised every year, some stay unchanged for a decade. Your folder is a snapshot of what arrived on the day you saved it, and nothing tells you when the manufacturer replaced it.

Subscribe on YouTube@sagnikteaches

The usual failure is quiet. A customer asks for the sheet for a cleaning solvent, someone searches a shared drive, finds solvent_SDS_final2.pdf, and sends it. The date inside is four years old and a new version has existed for eighteen months. Nobody did anything wrong in the moment; the process simply had no way to know. Rules on who must supply a sheet, in what language and how often it must be reviewed differ by market, so check your own regulator's guidance. If the rules where you sell expect a sheet to be updated soon after new hazard information appears, an unmonitored folder cannot keep up.

The fix is a register that treats each sheet as a record with a life, not as a file. Everything else in this tutorial hangs off that register.

The register: ten columns that make the library searchable

Use a spreadsheet to begin with. A shared sheet on a business storage plan is enough for a few hundred products; move to a database or a dedicated SDS tool only when you outgrow it. These columns do the work:

Two decisions here are easy to skip. First, record the revision date printed on the sheet, not the date the file was saved or the file's "modified" date, because those say nothing about when the manufacturer last changed anything. Second, give the register one owner. Someone has to be the person who says "this row has been checked against the PDF", and it is better to have one name than a shared sense that someone looks at it.

Building the register from an existing folder: a 140-product example

An illustrative case: a small industrial supplier sells cleaning chemicals, adhesives, lubricants and paints, bought from about a dozen manufacturers. The shared drive holds 168 PDFs across 140 current products, because some products have more than one sheet: an older version saved next to a newer one, or a separate sheet per pack size. Nobody has a full list.

The job has three passes. Pass one: a person drops all 168 files into one folder and renames none of them. Pass two: an assistant reads each file and produces a row. Pass three: a person checks the rows. The extraction prompt is narrow on purpose, since the less you ask the assistant to interpret, the less it can invent.

I am attaching one safety data sheet (PDF). Return one row with these fields only:
product name (section 1), manufacturer name (section 1), sheet version or revision number,
date of latest revision, language, and the page number where you found the revision date.

Rules:
- Copy values exactly as printed. Do not reformat dates; if two dates appear
  (issue date and revision date), return both and label them.
- If a field is missing, write NOT FOUND. Never guess.
- Do not summarise hazards or any other section.

A typical illustrative reply:

Product name: Degreaser XL 5 L
Manufacturer: [manufacturer name as printed]
Version: 4.2
Issue date: 12/03/2019
Revision date: 03/09/2024
Language: English
Revision date found on: page 14 (section 16)

Run the files through in batches of ten or twenty, in a business plan where uploads are not used for training by default, and paste the rows into the register. Then the person verifies. Open ten sheets chosen at random, covering different manufacturers, and compare the row with the PDF. If nine or ten agree, spot-check about one in five of the rest. If three or more disagree, tighten the prompt, or process by manufacturer, because one manufacturer's layout may confuse the reader. What breaks when an assistant reads PDFs and photos explains why scanned sheets are the usual culprits.

Look at the dates in the example above. A date printed as 12/03/2019 can be 12 March or 3 December depending on the manufacturer's habit. Do not let the assistant "fix" the format. Ask the person verifying to convert dates once, using the sheet's own conventions, or to ask the manufacturer when it is unclear.

What the first pass found, with illustrative numbers

In the 140-product example, the first pass over 168 files might give a table like this. The point is the shape of the answer, not the exact figures.

The rows sum to 140 products: 71 + 29 + 18 + 14 + 5 + 3, and the files to 168, because the 14 multi-sheet products hold 42 PDFs between them. Sorting those 14 is the quick win, because it removes the risk of the wrong version being sent at no cost, as long as a person reads both files first (the pack-size mistake below shows why). The 18 sheets over five years old are the real work: each is an email to a manufacturer, and the next section gives a template that makes it a ten-minute task for the lot.

Chasing manufacturers for current versions

Most manufacturers publish current sheets on their website or by request. Chase in one batch per manufacturer, not one email per product, and attach a list. An assistant can write the list from the register, grouped by manufacturer, and draft the message. You send it.

Subject: Request for current safety data sheets: [supplier name] account

Hello [first name],

We stock the products below from you and are updating our SDS records. For each one, please send the current sheet (or confirm that the version we hold is the latest), including its revision date.

1. Degreaser XL 5 L: we hold version 4.2, revision date 03/09/2024
2. Contact Adhesive 500 ml: we hold version 2.0, revision date 17/05/2019
3. Hydraulic Oil 46, 20 L: we hold version 7, revision date not shown

If a product has been reformulated or discontinued, please tell us so we can update customers.

Thank you,
[your name]

Log the date sent against each product in the register and set the status to "chasing". Give the manufacturer two weeks, then follow up once. If a sheet is still missing after a second request, escalate through your account manager at the manufacturer, and decide, as a business rather than as a spreadsheet, whether you are comfortable selling a product with no current sheet. That decision is not the assistant's.

When new files arrive, repeat the extraction prompt on them, then replace the old file, mark it archived (do not delete it, because you may need to show what was current on a past date), and update the "last verified" columns. If the new version changes the product's hazard wording, tell customers who bought it recently. The change might be small; judging whether it is small is a person's job.

Answering a customer's request in two minutes

Customers ask for sheets in three ways: "send me the SDS for X", "do you have the SDS in [language]", and "has this changed since the last one?" The register turns each into a lookup. The risk is the assistant answering from memory, so set the rule early: it retrieves and quotes; it does not advise.

A small supplier can use a shared Project or knowledge folder in a business chat assistant (a Claude Project, a ChatGPT Project, or a Gemini Gem, which Google is moving to skills: from 17 November 2026 on personal accounts and from March 2027 on Workspace business accounts), with the register and the PDFs as the only sources. Custom GPTs are scheduled to retire on 11 December 2026, so do not start new work in one. Give the assistant these instructions:

You help the sales desk find safety data sheets. Use only the register and PDFs in this project.
- To answer "send me the SDS", give the product code, the revision date, the file name and the link from the register.
- If two sheets exist for one product, say so and name the newer one.
- If the register status is "chasing", say the sheet is being verified and give the revision date we hold.
- Never describe hazards, first aid, storage or disposal in your own words. If asked, quote the relevant section exactly and give its page number, then say a trained person should confirm.
- If you cannot find the product, say so. Do not guess a similar one.

Try it with an illustrative customer email: "Hi, can you send the SDS for the 5-litre degreaser, and do you have a version in another language? Also is this the same as the one we got last year?" The assistant should return the English sheet's details, say the register holds no sheet in another language, and say it cannot tell what the customer received last year unless the order history is in the project. The last point is the useful one. A competent index admits what it cannot see.

The person at the sales desk then replies: attach the PDF, state the revision date, offer to ask the manufacturer for a translated version, and add a note if the version has changed since the customer's last order. The AI saves the search, not the judgement. If you later want something more customer-facing, the build, buy or skip decision for a supplier support chatbot sets out why a human review step usually survives.

The monthly pass that keeps it from going stale again

A register decays in the same way the folder did, unless something forces a look. Use a monthly or quarterly review of ten minutes.

1. Sort by revision date, oldest first. Anything that has crossed your age threshold (many suppliers pick three years) moves to "chasing".
2. Filter for "chasing" with no reply after 14 days. These are the follow-ups.
3. Filter for "last verified" older than 12 months. These need a person to open the PDF and confirm it still matches the row.
4. Check new products added to the price list this month. Each needs a sheet before the first sale, not after.
5. Check discontinued products: archive their sheets, but keep them findable for as long as customers may hold stock.

The assistant can build the list for each filter if you paste the register in. For a hundred rows, the spreadsheet's own filters do the same job faster and are the safer choice. Use AI when the reading is messy, such as new PDFs, odd dates and customer emails, and use filters when the data is already clean. The same line between clean and messy data is drawn in preparing your business data for AI.

A mistake that looked fine: the sheet for the wrong pack size

An illustrative failure. A supplier holds two sheets for a lubricant, one for the 1-litre bottle and one for the 20-litre drum. The manufacturer issued separate sheets because the product was labelled differently by pack size. During the first extraction, both rows came back with the same product name and nearly the same date, so the assistant treated them as duplicates and the person archiving "duplicates" kept only the sheet with the later date. A customer who bought drums was sent the bottle's sheet for months.

Nothing the assistant did was strictly wrong, since it extracted what was printed. The failure was a rule applied to the output: "same name, keep the newer date". The fix was a new register column, "pack or variant", filled in from section 1 of each sheet, and a rule that duplicates are resolved only after a person reads both files. Duplicates are not always duplicates. A better test is to ask whether the two files would be handed to the same customer for the same purchase.

Look for the same type of fault in your own register: two rows that look alike but are different, and one row that merges two distinct products. Sorting by product name and scanning for near-matches takes five minutes and finds most of them. A supplier selling coloured paints, for instance, may have one sheet per colour family or one per product line, and the register should follow whatever the manufacturer did.

What a person must keep, and where AI must stop

Keep the lines clear in writing, because they get blurred with busy staff and a tool that sounds sure of itself.

The comparison row deserves one more worked line. Given two versions of the same sheet, an assistant can list which sections differ, for instance "section 2 hazard statements added; section 7 storage temperature changed from 5 to 30 degrees C to 5 to 25 degrees C; section 14 unchanged". That is useful triage. It is not a judgement that the change matters to the customer, and it can miss a changed figure in a table, so the person reads the two sections themselves.

Wording for customers who ask how current your sheets are

Customers, especially trade buyers with their own audits, sometimes ask how you keep sheets current. A short, honest policy, written once, avoids improvising under pressure. An illustrative version:

We hold a current safety data sheet for every product we sell. Each sheet is recorded in a register with its version and revision date. We ask each manufacturer for their latest version at least once a year, and replace our copy when a new one is issued. If you need confirmation of the version you hold, quote the product code and we will send the revision date from our register.

Only publish a statement like this if it is true, and set the "at least once a year" to what you actually do. If your register shows 18 sheets over five years old, this is the sentence that makes it urgent to chase them. Do not say you comply with a particular regulation unless the regulator's guidance, read by someone qualified, says you do.

A brief word on records beyond the sheet. Some rule sets expect you to retain what you supplied to a customer on a given date. That is why old versions are archived rather than deleted, and why the register keeps an archived status.

What it costs and what to do first

The tools are cheap. A spreadsheet you already have, a business chat assistant seat, and a shared folder cover everything above. Specialist SDS software exists, and some of it includes authoring and label printing for businesses that write their own sheets; as a supplier of other people's products, you may not need it. Start small:

1. Day one: create the register and give it an owner.
2. Week one: process the sheets for your ten best-selling products and verify them all.
3. Week two: extend to everything, verify a sample, and send the first batch of manufacturer chasers.
4. Month one: set the monthly review and write the customer wording.

If your supplier base is mostly importers, the wider question is documents beyond sheets, which the importer product compliance checklist covers with the same flag-the-gap logic. And if your quality, safety or consultancy needs go further than a register, the right adviser is a qualified safety professional; AI helps you keep the paperwork in order, and nothing in this approach replaces a trained person's judgement.

Questions about running an SDS library with AI

Can AI tell me what hazards a product has?

It can quote what the sheet says, but it should never be the source of hazard, first-aid or disposal advice. Treat the supplier's own current sheet as the only authority. Have the assistant point to the section and page, then let a trained person read that section before anything goes to a customer.

How often should an SDS be replaced?

Replace it whenever the manufacturer issues a new version. Some rule sets also set a maximum time between reviews, and the period depends on where you sell, so take it from your regulator's guidance rather than from a figure on a website. Ask the manufacturer for their current version at least once a year and record the answer.

Is it safe to upload SDS files to a chat assistant?

Safety data sheets are usually shared freely with customers, so the sheets themselves are low risk. Do not upload customer lists, order histories or price files alongside them. Use a business plan, where content is not used for training by default, and keep the register itself in your own storage.

What if the manufacturer never replies about an old sheet?

Record every attempt in the register with dates, escalate through the manufacturer's account contact, and decide whether to keep selling that product. A sheet with no revision date or a very old one is a business risk: tell customers it is being verified, and do not describe it as current.

Further reads

• GPSR Compliance Checklist for Online Sellers: What the EU Rule Asks For — See what a product-safety rule asks of sellers who ship across borders.
• How to Clean Up Customer Records Before You Add AI — Tidy the records an assistant will read before you rely on it.
• Is It Safe to Put Customer Data Into ChatGPT? — Decide what may be pasted into a chat assistant.
• Check a Bill of Materials With AI: Template, Gaps and Revisions — Apply the same revision-tracking idea to a bill of materials.
• AI Compliance Checklist for Small Businesses: What Applies to You — Place SDS handling inside a wider compliance routine.
• MRO Inventory Management: AI-Set Minimums for Maintenance Spares — Link SDS records to the stock you hold for maintenance.
• Can AI Write a Beauty Salon Risk Assessment? What It Must Not Skip — How to draft a salon risk assessment with AI by treatment, and the on-site checks, records and sign-off that a competent person has to do themselves.
• AI for Technical Sales: Answer Spec Questions From Your Datasheets — Build an internal spec-answer routine with versioned datasheets, cited replies, abstentions and a 30-question accuracy test.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the 16-section safety data sheet format set out under the UN Globally Harmonised System and in regulators' published SDS guides (section 16 carries the date of preparation or last revision), Google's help page on the move from Gems to skills, OpenAI's custom GPT retirement FAQ, and Claude's help page on Projects. Rules on who must supply a sheet, in which language and how often it must be reviewed differ by market: check your regulator's guidance.