From source documents to passports, labels and value
Upstream, AI turns documents and system data into candidate facts. In the middle, people own and approve the facts. Downstream, AI and templates generate every output from the same approved facts. Click any box.
Click a box above
Each box shows what it is, what AI does there, what people do, and how we would measure it.
Walk-through: one illustrative product
Tile adhesive TA-200 (illustrative) — a cementitious adhesive. All values below are invented for the example.
AI guardrails: fast where it is safe, human where it matters
Regulated data needs AI that is useful and accountable. The same pattern already runs in production for regulatory label content in crop protection: AI drafts, a person decides, everything is logged.
AI does
- Extract candidate facts from SDS, test reports and supplier documents, with a confidence score and a link to the source page
- Check consistency across documents (SDS vs label vs declaration) and flag conflicts
- Draft outputs and translations from approved facts
- Watch regulation changes and list affected facts, products and documents
- Answer staff questions from approved facts only, with sources
AI never does
- Invent a value that has no source — a missing performance value stays a gap
- Approve a fact or publish an output on its own
- Decide a hazard classification or a declared performance
- Overwrite a fact without a new version and a reason
Human in the loop
Low-confidence extractions go to the fact owner. Every output is released by a named person.
Audit trail
Each fact keeps source document, page, extraction run, reviewer and date — what an authority or a customer audit asks for.
Kill switch and limits
Any AI service can be switched off without stopping the process; spend caps and monitoring per service.
How we know it works
| Metric | Why it matters |
|---|---|
| Field-level extraction accuracy vs a reviewed sample | Decides how much review is needed |
| Share of facts auto-accepted vs sent to review | Shows where people still spend time |
| Review time per document | The real efficiency gain, measured against the baseline |
| Conflicts found between documents | Quality value that was invisible before |
Value case: from assumptions to an annual number
Move the sliders to your own assumptions. Defaults are illustrative only — the first step of the programme is to measure the real baseline for one product family.
Not counted: market access (revenue that stays sellable once passports are mandatory) and reuse across future regulations. Both are often the largest items — show them as risk avoided, not as savings.
How to take it to a steering committee
- Measure today for one product family: hours per document, errors, launch lead time.
- Run the pilot with the same measures — no new numbers without a baseline.
- Present three numbers: effort saved (hard), launch time saved (revenue), risk avoided (market access).
- Ask for funding by phase, each phase released on measured results.
Roadmap: start small, prove it, then scale
The passport deadline is not here yet — that is the opportunity to build the foundation calmly instead of a rushed passport project later.
0–90 days
Listen, map, prove one flow
- Listening tour: regulatory, product data, R&D, IT, a few countries
- Fact inventory for passport and packaging: ~15–20 facts, each with source and owner
- Baseline one product family
- Generate one output from governed facts; show the measured difference
6–12 months
Foundation and first AI services
- Fact register live for priority families
- AI capture from SDS and test reports with review
- Consistency checks across SDS, labels, declarations
- Mapping to data templates (EN ISO 23386/23387)
12–24 months
Passport-ready and scaling
- Passport records generated per family when the delegated act lands
- Packaging facts for the labelling rules
- Regulation-change monitoring
- Roll out family by family, measured each time
Regulatory clock (as far as published)
Industry context: nobody has to solve this alone
What the passport must carry
Declaration of performance and conformity, instructions and safety information, technical documentation and labelling, unique identifiers — reached through a data carrier on the product.
A data model already exists
CEN/TC 442 makes declarations machine-readable through data templates (EN ISO 23386 and 23387; 23387 revised 2025). The task is to map existing facts to them, not to invent a private format.
The industry pilot
CIRPASS-2 construction pilot (EU-funded, 2024–2027), led by Cobuilder with GS1 France, Soprema, Velux, Autodesk and others: QR codes to passports and a standard data model.
What batteries learned first
The Battery Pass consortium concluded the passport is a strategic opportunity, not only a compliance cost. Pilots: open standards work, engage suppliers early, value appears only when data is trusted.
Packaging is the same foundation
Material, weight and recycled-content facts are product-master facts too. The packaging rules become one more reader of the same register.
Done before, in another regulated industry
A common label data model agreed by competing crop-protection companies for a 27-country EU digital label, and AI agents in production for regulatory content with human review. The method transfers; the facts are new.
Questions to bring: Which product families and rules come first? Where do regulatory facts live today? Which associations or pilots does the company follow?