At Valenc, the answer is specific. AI reads documents and drafts text. It does not classify.
When a supplier SDS lands in the system, extraction models parse it into structured data: substance identities, CAS numbers, concentrations, physical and toxicological properties. That is reading work, and it's the kind of work modern AI is genuinely good at. A human doing the same job re-keys fields from a PDF for an hour and still fat-fingers a decimal somewhere around document forty.
Classification is a different kind of problem. Whether a mixture lands in Acute Tox 4 or Skin Corrosion 1B is not a judgment call. It is the output of published rules: additivity formulas, concentration cut-offs, bridging principles, all defined jurisdiction by jurisdiction in black and white. The correct way to implement published rules is to implement them.
Deterministic, by design
So our classification engine is deterministic. Same inputs, same answer, every single time, with the specific rule that fired recorded next to the result. There is no temperature setting, no confidence interval, no "the model usually gets this right."
The difference shows up the day someone asks you to defend a classification. A probabilistic classifier can be right 99% of the time and still be impossible to stand behind, because you can't show why it produced the answer it did. A rules engine shows its work: this cut-off, this formula, this regulation, this revision. That's the difference between an answer and an audit trail.
Where the AI earns its keep
None of this means the AI is window dressing. It does the parts of the job that eat a product steward's week: reading inconsistent supplier documents, drafting all sixteen sections in every language you ship to, and reviewing its own output to flag exactly which fields need a human's eyes before publish. The expensive human judgment gets spent on judgment.
It also means that when a rule changes, the fix is structural. Update the rule, re-run the engine, and every affected product reflects the new requirement, identically and provably. You don't retrain anything and hope.
So when a vendor tells you their platform is AI-powered, the useful follow-up is simple: which decisions does the AI make, and which would you be comfortable explaining to an inspector? Our answer fits in one sentence, and we think that's a feature.