Every finished product containing the material, at every concentration. Every specification referencing it. The classification and labelling of anything whose hazard profile shifts. Safety data sheets in each language and market. Dossiers and notifications that stated the previous composition. Allergen and claims statements. Costings. Open purchase orders and any batch in progress against the old specification.

Written out, nobody would attempt that from memory. In practice it's attempted from memory constantly, because the alternative is a fortnight of cross-referencing.

Only if the relationships are recorded. If the formula knows its substances, the specification knows its formula version, the dossier knows the specification it was written against and the batch knows the specification it was made to, then the impact list falls out of the data.

If those four things live in four systems joined by a spreadsheet, no amount of AI helps, because the relationships were never written down. That's the architectural point behind AI-native versus AI bolted on.

A list, grouped by what has to happen: documents to reissue, notifications to update, batches to hold, customers to inform. Each with the reason attached and a link to the record, so the work can be assigned rather than re-investigated.

And it should run in both directions. Forwards from a proposed change, to decide whether to make it. Backwards from a regulatory change, to find every product it affects. The second is what turns a restriction announcement from a fortnight of anxiety into an afternoon of work. The document end of that is covered in the SDS authoring buyer's guide.