Digital Product Passport: When AI Fills the Data Gaps

Digital Product Passport: Why AI-Extracted Data Needs a Second Look

The digital product passport is about to turn a fashion brand's supply chain data into a public, scannable record. Under the EU Ecodesign for Sustainable Products Regulation, textiles are among the first product groups expected to carry a passport linked from a code or tag on the garment, holding details on materials, origin, and environmental attributes. The central EU registry was due to go live in July 2026, and technical work on what textile passports must contain is well advanced. Brands now face hundreds of supplier certificates, spreadsheets, and PDFs in a dozen languages, and many are turning to AI to extract the data. That works well, right up until an extracted number becomes a compliance claim.

What the Passport Needs vs Where the Data Lives Today

For most brands, the required data exists somewhere. The problem is that it is scattered, inconsistent, and rarely tied to a specific product.

What Is the Digital Product Passport for Textiles?

The Ecodesign for Sustainable Products Regulation entered into force in 2024 and creates a framework in which detailed requirements for each product group are set through separate delegated acts. Textiles and footwear sit among the priority groups. Each passport is a digital record sitting behind a scannable tag on the item, whether a QR code, an NFC chip, or an RFID label, giving consumers, repairers, recyclers, and market surveillance authorities access to relevant information about the product.

The regulation required a central registry by July 2026. The textile-specific delegated act, which will fix the exact data points and deadlines, is expected around 2027, with obligations applying after a transition period, so the first fully compliant collections are likely later in the decade. Timelines in EU product regulation do shift, and brands should track official Commission announcements rather than vendor summaries.

Who the Digital Product Passport Applies To

The digital product passport follows the product into the EU, and it makes no difference where the brand keeps its head office or where the garment was sewn. That reaches deep into global supply chains. Spinners, mills, dye houses, and garment factories across South Asia, Southeast Asia, Turkey, North Africa, and China will be asked for data many have never been asked to provide in structured form. A separate rule under the same regulation restricting the destruction of unsold clothing and footwear by large companies has also begun to apply. And it arrives alongside other EU regimes, including EU AI Act compliance for companies using AI inside these very processes.

Why Brands Are Turning to AI for Passport Data

The volume is the first reason. A mid-sized brand can have thousands of active styles, each touching several suppliers across multiple tiers, each supplier holding certificates, test reports, and declarations in different formats and languages. Manually keying that into structured fields is slow, expensive, and error-prone in its own way.

AI is genuinely good at this kind of reading. Models can pull fibre percentages from a specification sheet, identify certificate numbers and validity dates, translate supplier documents, and match records across systems. Agent-style tools can go further, retrieving certificates from certification databases and linking them to purchase orders. Structured, accurate product data is also becoming valuable well beyond compliance, since machines increasingly read it to recommend and buy products, as we described in our piece on agentic commerce.

Six Ways AI Extraction Goes Wrong on Supplier Data

These failures all share one trait: the extracted value looks exactly like a correct one.

1. Blend percentages that do not reconcile. Rounded or partially read compositions add up to slightly more or less than one hundred percent, or silently drop a minor fibre.
2. Certificate scope mismatch. A certificate covers a facility or a material, not necessarily the specific order or style it gets attached to.
3. Origin collapsed to the final stage. The country of final assembly gets recorded as the origin of every production step.
4. Expired or superseded certificates. Validity dates are misread or ignored, and an old certificate supports a new claim.
5. Translation of technical terms. Fibre names and processing terms vary across languages and trade usage, and near-synonyms get conflated.
6. Gaps filled with plausible values. Where a document is silent, a model can infer a typical value rather than flagging the field as missing.

Pros and Cons of AI-Built Passport Data

AI makes passport readiness achievable. It also creates a new route for errors to reach the public.

• Pro: makes the volume feasible. Thousands of documents can be processed in the time a team would spend on dozens.
• Pro: reads many formats and languages. Supplier documents no longer need to be standardised before anyone can use them.
• Pro: surfaces missing data early. Gaps across supplier tiers become visible years before obligations apply.
• Con: plausible fills become public claims. An inferred value published in a passport is a statement the brand has to stand behind.
• Con: greenwashing exposure. EU rules on environmental claims are tightening, and an overstated recycled content figure is exactly what enforcement targets.
• Con: supplier relationships suffer. Errors introduced by extraction can be wrongly blamed on suppliers who reported correctly.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how digital product passport data risk plays out in practice rather than presented as verified case studies.

A brand's extraction tool reads organic cotton content from a supplier's marketing brochure rather than from the transaction certificate. The certificate covers only part of the order volume. The passport would have stated a higher certified share than the brand could evidence.

A denim programme records Türkiye as the origin because final sewing happens there. The fabric is woven and dyed in a different country. The passport, as drafted, would misstate origin for every stage before assembly.

An agent-based tool links a recycled polyester certificate to the wrong purchase order because two style codes differ by one character. Nothing looks wrong in the dashboard, and the mismatch is caught only because a sourcing manager recognises that the supplier does not make that fabric.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

A Verification Layer Before Data Goes Public

Treat every extraction result as a draft, never as the record. Require each passport field to link back to a specific source document and page, so any value can be traced in seconds. Check certificates for three things before accepting them: that the scope covers the material and facility in question, that the validity dates cover the production period, and that the certificate matches the specific order. Put human sign-off on the fields that carry the most risk, namely composition, recycled or certified content, origin by stage, and substances of concern. Flag every value where extraction confidence is low or where two sources conflict, rather than letting the system pick one.

Start now, even though obligations arrive later. Building a reliable data trail across several supplier tiers usually takes multiple seasons, and the brands that begin early will be correcting gaps while the stakes are still low.

Why Talkory Wins

Extraction errors are hard to see because a wrong value looks exactly like a right one. Talkory gives the same certificate or specification question to GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass. Where all six read the same composition, scope, and validity dates, the extraction is probably sound. Where the readings differ, that field goes to a person before it goes into a public passport. Across thousands of fields, disagreement is the fastest way to find the handful that genuinely need attention.

Final Verdict

The digital product passport will make textile supply chain claims public, structured, and checkable by regulators, recyclers, and customers alike. AI is the only realistic way many brands will assemble that data at scale, and it reads messy supplier documents remarkably well. It is also good at producing plausible values where those documents are silent. Build the source trail now, verify certificate scope and origin by stage, keep people accountable for the fields that make environmental claims, and use cross-model disagreement to decide where to look first.