Compliance is becoming an AI problem

Regulatory complexity has changed in nature

The difficulty is not knowing the regulation: it is crossing it with the moving reality of the portfolio. An additive's maximum dose depends on the food category; carry-over through a compound ingredient must be counted; a claim depends on nutritional thresholds recalculated at every reformulation. Each check is simple in isolation; it is their multiplication — hundreds of recipes times dozens of rules times permanent change — that overwhelms teams.

Compliance by design

The corrective approach — check at the end of development, fix, loop — piles up iterations and lets non-conformities slip through. The by-design approach reverses the flow: rules are encoded in the product repository and evaluated at formulation time. A dose overrun is flagged at data entry; a claim endangered by a reformulation shows up before validation; the label regenerates from the recipe instead of being maintained separately.

What AI assistants do

Beyond encoded rules, AI brings three new capabilities. Interpretation: reading a supplier technical sheet, extracting the relevant regulatory data and structuring it. Case preparation: assembling a dossier — claim substantiation, export labelling, audit response — by gathering the evidence and drafting justifications from the repository. Applied watch: when a text evolves, immediately identifying the affected recipes rather than circulating a memo everyone must decline on their own.

Human validation remains the keystone

AI prepares, the expert decides. A regulatory interpretation commits the company: it must be validated by a professional who owns the responsibility. A good system makes that validation efficient — every AI proposal arrives sourced, traced and reversible — and logs each decision, building along the way the company's regulatory memory, defensible in audits.

Food Scientist's Corner

A concrete example: reformulating a sauce by swapping a tomato concentrate for a double concentrate of another origin. The cascade: the sorbate brought in by the compound ingredient (carry-over to recount), the primary-ingredient origin declaration, the QUID if tomato is highlighted, and the declared nutritional values. Four regulatory checks for a seemingly trivial change — exactly the kind of cascade a by-design system runs on its own.

Future outlook

Regulatory affairs is evolving from final inspection to rules engineering: encode, supervise, validate — while the machine applies. Teams win back what paperwork stole: expert time for the genuinely difficult cases.

That is crumble-ai's architecture: EU 1333/2008 and Codex GSFA reference data embedded in formulation, compliance checked at entry, labels generated from the recipe and every decision traced. Compliance stops being a toll at the end of the project: it becomes a property of the system.