Beyond the demo: why AI in the Digital Product Passport requires a verification architecture

AI generated

In recent months, we have witnessed a proliferation of prototypes where Artificial Intelligence "reads" and "queries" digital product passports. For a C-level executive, these demos are seductive: they seem to promise total automation of compliance and ESG reporting. However, there is a profound gap between a chat that correctly answers a question on a PDF file and an industrial system capable of governing the data of a complex supply chain.

The risk, if not managed, is what I call <a href="https://www.matthewgoldman.com/delegation-and-verification/">"delegation without verification"</a>: delegating critical processes to probabilistic models without a control infrastructure is, to all intents and purposes, gambling.

To successfully integrate AI, we must map the DPP process by distinguishing two critical phases: <b>generation</b> (the collection and standardization of data from company silos) and <b>analysis</b> (querying the passports for audit or decision-making purposes). The most common mistake today is to approach these phases with architectures based solely on "fast thinking": systems that produce quick answers through statistical association but lack a real deterministic anchor.

In the generation phase, the bottleneck is <a href="https://aws.amazon.com/it/what-is/retrieval-augmented-generation/">data retrieval (RAG)</a>. If the system fetches obsolete or incorrect information from ERP systems—for example, confusing the technical sheet of virgin polyester with that of recycled polyester—no model "intelligence" can correct the error downstream. In an industrial context, hallucination is not a trivial "chat error": it is a false Digital Product Passport. Delegating data creation to an AI without a verification infrastructure means turning compliance into gambling, exposing the company to risks of involuntary greenwashing and heavy legal penalties.

To move from demo to real value, technical integration (API) is not enough. What we at Mangrovia Technologies call the <b>Process Layer</b> is necessary.

This is not just about connecting the AI to the data via Anchorpass, but about building an orchestration infrastructure that defines the workflows and the boundaries within which the agent can operate. A reliable industrial system is not a single AI that answers everything, but a chain of specialized sub-agents:

• 1. One agent extracts data from legacy systems (ERP/CRM).
• 2. A second agent verifies the consistency of this data with respect to the DPP requirements.
• 3. A third agent signals anomalies to a human operator.

In this scheme, the AI imitates "slow thinking": it plans, verifies, and acts only within the limits established by company policies.

In this scenario, the central figure of the future will not be the one who writes the prompts, but the <a href="https://www.researchgate.net/figure/The-architecture-of-the-verification-system_fig10_370791128">Architect of verification systems</a>. This figure is responsible for designing the rules with which the AI interacts with the real world and, above all, for establishing how to measure the system's reliability.

AI governance must shift upstream. The true return on investment (ROI) for a company is not achieved by being subjected to the AI process, but by designing an architecture where domain experts (those who know the supply chain and regulations) collaborate with technicians to validate the agents' work.

The Digital Product Passport is designed to generate trust among manufacturers, consumers, and regulators. Introducing AI into this flow without a rigorous verification architecture means fundamentally undermining that very trust.

<b>Innovation is not about adopting AI regardless, but about knowing how to orchestrate it within a protected process, where every delegation is supported by deterministic verification wherever possible. Only in this way will the DPP transition from a regulatory obligation to a true strategic asset for the company.</b>