Agentic AI for Digital Product Passport Documentation: A Feasibility Study on Heterogeneous Manufacturing Documents

Abstract

Digital Product Passports (DPPs) are intended to make lifecycle product information accessible across the value chain and enable circular business models. Yet, the adoption remains constrained by the heterogeneity of existing industrial documentation and the cost for normalizing it into unified schemes. Agentic systems based on large language models (LLMs) offer an alternative path that operates directly on existing documents rather than restructuring them. This paper presents a feasibility study combining a PRISMA-guided systematic literature review (SLR) of 23 agentic document-reasoning systems with a small-scale empirical evaluation of one system on a purposive sample of 204 questions across the documents of 12 industrial and consumer products. The aim of the study is to generate qualitative insights about where current capabilities are sufficient and where they fall short. The system answered 85% of questions correctly, with 5% incomplete and 10% false. Two patterns emerge consistently. First, the reasoning capabilities of the frontier model appeared adequate for industrial document understanding. The practical bottleneck, for the single frontier model and orchestration pattern evaluated here, lay in retrieval, where general-purpose embeddings failed to bridge the terminological gap between user queries and manufacturer-specific vocabulary. Second, deterministic tools markedly improved the answer quality in question categories, where LLMs are structurally weakest. These observations suggest that the retrieval infrastructure adapted to manufacturer vocabulary and complementary deterministic tools may be more consequential design priorities for DPP deployments supported by agentic AI than further work on orchestration complexity.