Compliance is a data problem before it is a chemistry problem

Most compliance programmes do not fail on interpretation. Teams know what REACH Article 33 requires, what the SVHC candidate list contains, and roughly what the universal PFAS restriction proposal would mean for seals, coatings, wiring and thermal systems. The regulation is public. The obligation is legible.

The failure happens one level down, in the data. A declaration is due, and the question becomes: which part contains the substance, at what concentration, in which product, sold into which market — and can we show our working?

That question is answered by collecting data from sources that were never designed to answer it together.

It is also answered at a volume that defeats manual effort on its own. A mid-sized portfolio runs to tens of thousands of parts. A large one runs to hundreds of thousands, each with its own materials, substances, suppliers and supporting documents. Even where every fact could in principle be found by hand, no team of any realistic size could find them all, reconcile them, fill the blanks and keep them current inside a regulatory deadline — economically or at all. This is not a problem more headcount solves. It is arithmetic.

The hard part is not reading the regulation. It is collecting the evidence at that volume, deciding which version of a fact is true, and being able to defend that decision.

One fact, four sources

Take something as simple as the mass of a part.

It exists in the supplier’s declaration. It exists on the engineering drawing. It exists in logistics data because it was declared at customs. Somebody may have put a sample on a scale. All four values are stored. Some are right. Some are wrong. Nothing in the record says which is authoritative.

This is not a rounding problem. Concentration is mass over mass. Under REACH, an SVHC above 0.1% w/w in an article triggers communication down the supply chain and notification to SCIP. Whether you are above or below that threshold is decided by a number that four systems disagree about, and that most organisations resolve by hand, in a spreadsheet, each time it is asked.

The same pattern repeats across every fact compliance depends on: material class, additive, substance identity, part revision, supplier identity. One supplier appears under three spellings across two systems. A material is described in free text in one place and by specification number in another. Multiply that by a hundred thousand parts. The evidence exists. It is not connected.

The evidence is already in the building

The instinct when data is missing is to ask suppliers again. More supplier outreach campaigns, more follow-ups, more escalation. That creates activity. It does not reliably create certainty, and it stalls entirely when a supplier does not answer — while the manufacturer remains accountable for the product regardless.

Meanwhile a substantial share of the answer is already held internally, in material safety data sheets, full material disclosures, specifications, engineering drawings, test reports, certificates and the supplier documents already received and filed. It is unstructured, inconsistent, and spread across document stores that no workflow reads.

And the authoritative external references — the CAS Registry for substance identity, the ECHA candidate list, SCIP — are public and stable. They are not the bottleneck. The bottleneck is the connection between a substance identifier in a public register and a material description typed into an ERP field eleven years ago.

What Viridium AI actually does

Four steps, applied continuously rather than as a project.

For readers who want the architecture: the output is not a report. Extracted facts are written into a private knowledge graph, structured by an ontology that models products, parts, materials, substances and suppliers and the relationships between them, with a semantic layer that lets a material description in one system and a substance identity in a public register be recognised as the same thing. The graph is what makes the next question cheap. Reasoning models then interrogate it — tracing a new restriction to affected parts, products and markets — because the relationships are already resolved.

Confidence, stated rather than assumed

Not every fact deserves equal trust, so each one is graded on a one-to-five scale. One star is AI-inferred with no document behind it. Two, a human has accepted that estimate. Three, documentary evidence has been found and the value extracted. Four, the value has been enriched and validated against sources and constraints. Five, an expert has reviewed the whole chain and signed it off.

This is what makes the output defensible. A regulator, a customer or an internal auditor can be shown not just the answer but the source, the derivation and who approved it. It also tells your team where to spend limited expert effort: on the one-star facts sitting under a threshold decision, not on everything.

And corrections feed back. An expert judgement made once improves the next result rather than disappearing into a report.

This makes your systems of record worth more

None of this replaces ERP, PLM or your compliance recording system. Those systems hold the authoritative record, the product structure and the campaign history, and they should continue to.

The gap is that their workflows only run as fully as the data inside them allows. A compliance workflow cannot assess a threshold when the weight field is empty. A change-impact process cannot trace a restriction when a material is free text. Viridium enriches those records — resolved substance identities, validated weights, connected material composition, evidence attached — and writes the knowledge back. The systems you have already paid for start completing workflows they previously could not finish.

Across manufacturing deployments this has meant 27% fewer supplier requests and 36% faster analysis, because the answer was found in evidence already held rather than requested again.

Where to start

Pick one obligation with a deadline. Establish which facts it actually turns on, grade your confidence in each, and fix only the ones that decide the outcome. Do that once properly and the foundation is there for the next restriction, the next customer questionnaire, the next substitution.

Compliance data does not get better by being cleaned. It gets better by being used, with the evidence retained.