A general model asked to write a product information file will produce something that reads correctly and describes a product that doesn't exist. That isn't a hallucination problem to be tuned away, it's the predictable result of asking a model to write about data it was never given.

A grounded system works differently. It reads your formula, your supplier certificates, your test reports and the applicable regulation, and it writes only from those. The value of each statement can be traced back to the line it came from. Where the data is missing, the correct output is a gap flagged for someone to fill, not a plausible sentence.

So the question to put to a vendor isn't how accurate the model is. It's what the model is allowed to read, and what it does when the answer isn't there.

The Responsible Person, the safety assessor and the person who released the batch are all people, named in your file, and none of that changes because a draft was machine-written. What changes is what the file has to record.

A defensible trail holds five things: which source data the output was generated from, what the system produced, what a person changed before approving, who approved it, and when. Held on the record rather than in a separate log, that trail is stronger than the manual equivalent, where the honest answer to 'where did this number come from?' is often a spreadsheet somebody has since overwritten.

The document standards this has to satisfy are set out in the product information file guide and the SDS authoring buyer's guide.

This is where a connected system earns most of its keep and where a document store fails hardest. A substance moves onto a restriction list. In a folder-based operation, someone has to remember which products contain it, in which markets, and which dossiers referenced the old position.

If the substance data sits on the product record, that's a query rather than an act of memory: every affected product, every affected market, every document that needs revisiting, produced immediately. The AI's job is the detection and the list. Deciding what gets reworked, and when, stays a regulatory judgement.

Ask three questions and get the answers in writing. Where is the data processed, and under which jurisdiction. How long is it retained by the model provider. Is it used to train any model, including with a vendor's assurance that it's 'anonymised'.

For Worldover the answers are: processed under contract with no training use, retained only for the length of the request, and held within your tenant. Formulations are the most commercially sensitive asset in this industry, and a vague answer to any of the three is a reason to stop the evaluation.