Reformulating for cost when a raw material price moves, and needing to know which alternatives are functionally equivalent and compliant in the same markets. Reformulating for compliance when a substance is restricted somewhere you sell. Finding the trial you or a colleague already ran three years ago on a similar base. And checking a modified formula against restriction, allergen and classification data before it goes near a stability cabinet.

None of that is creative work. All of it is search and cross-reference against data you already hold, and all of it currently costs a technical specialist hours per formula.

Because the constraints that matter aren't in the data. Sensory performance, stability under real packaging, how a base behaves at scale in your vessels, what your customer will accept on an ingredient list: some of it is tacit, some of it only shows up at 500kg, and none of it is in a database.

What a model can do is narrow a search space of thousands to a shortlist of six that satisfy the stated constraints, and explain why each is on the list. The formulator then does the part that requires judgement and a bench.

Your trial history has to be structured rather than scattered across notebooks, and your raw material library needs compositions and functions rather than just names and prices. Most operations have the data and not the structure, which is why the first phase of any of this is usually getting the lab record straight.

That's the work described on R&D for cosmetics manufacturers, and it pays for itself before any AI is switched on.