The first instinct of every ERP implementation partner is to model regulation as custom fields on the product master. A flag for EU Cosmetics, a free-text field for the REACH registration number, a checkbox for CLP hazard classification, a notes column for INCI. The thinking is that regulation is metadata about the product, so it belongs in the product record.

This is wrong on the merits. EU Cosmetics Regulation 1223/2009 is not metadata about a cosmetic, it is the operating constraint that decides whether the cosmetic can be sold. REACH is not a field on a substance, it is a registration regime with tonnage bands, dossier obligations and authorisation lists that change quarterly. CLP defines the hazard pictograms, signal words and precautionary statements that must appear on the label, derived from the actual composition, not typed in by a clerk. INCI is the naming convention that decides what goes on the pack and in what order, calculated from the formula.

In every case the regulation is an executable rule against the live composition of the product, not a static fact about the SKU. A custom field cannot tell you that an ingredient was added to Annex II last week and that thirty of your SKUs are now non-compliant in the EU. A custom field cannot generate a CLP label when the concentration of a hazardous component crosses a threshold. A custom field cannot re-derive the INCI list when the formulator swaps a preservative. Regulation lives in the operational core or it does not work.

The generic ERP data model has one anchor: the SKU. Inventory, orders, costs, taxes and revenue all hang off the stock-keeping unit. A bill of materials decomposes the SKU into components. This is a faithful model of a screw, a sofa or a t-shirt. It is not a faithful model of a cosmetic formula or a chemical product.

A formula is a living document. It has versions, because the formulator reformulates when a supplier drops a grade, when an ingredient is restricted, when a customer asks for a vegan claim. It has validations, because the Margin of Safety, the preservative efficacy and the stability profile must be re-checked on every change. It has thresholds, because the regulation sets maximum concentrations per ingredient per market that must be enforced before the version is allowed to lock. It has a lock state, because once the Responsible Person signs off, the version is frozen and any further change opens a new version with a new safety assessment.

None of this exists in a generic ERP. The standard workaround is a parallel system, normally a Product Lifecycle Management tool or a folder of Word documents, with the ERP holding the latest approved bill of materials. The cost is real and recurring: the formulation team works in one system, the operations team works in another, and the regulatory team rebuilds the join every time something changes. A substance-based operating system treats the formula as a first-class object, with versioning, validation, thresholds and lock state native to the data model.

The current AI wave has produced a generation of ERP add-ons that drop a chat panel onto a legacy interface and call the result an AI ERP. For chemicals, cosmetics, and supplement businesses this is worse than no AI at all, because the answers are plausible and confidently wrong. A generic large language model cannot calculate a Margin of Safety, because the calculation needs the No Observed Adverse Effect Level for the ingredient, the exposure scenario for the product type and the current SCCS opinion. A generic LLM cannot screen against eighty plus markets, because the banned, restricted and labelling rules are not in its training data at the resolution needed. A generic LLM cannot cross-reference allergen interactions, because the interaction matrix is proprietary, jurisdictional and updated on a schedule the model has never seen.

Willow can do all three, because it is trained on chemical and regulatory data and wired into the operating system that holds the live formula. The MOS comes back with the calculation laid out: NOAEL, daily exposure, retention factor, dermal absorption, the resulting ratio and the SCCS opinion it cites. The multi-market screen returns a market-by-market verdict with the rule and the source for each flag. The allergen check shows the interaction, the regulatory basis and the suggested mitigation. Willow is the connective tissue across every tool in Worldover, with regulatory and chemical data as a first-class part of what it knows, and it shows its working on every answer.

The argument is not that AI is dangerous. The argument is that AI without domain data is dangerous in a regulated industry, because the gap between a fluent answer and a correct answer is invisible to the user. The fix is not less AI, it is AI built on the right data model.

None of the three arguments above are solvable with a better implementation partner or a longer configuration project. They are arguments about the shape of the data model and the shape of the AI layer. A generic ERP was designed for businesses where the SKU is the unit of work, regulation is a periodic audit, and AI is a productivity feature. A substance-based business has none of those properties. The operating system has to be built for the business, not bent around it.