Customers, regulators and the EU now ask for the carbon footprint of individual products, not only of the company. Doing that by hand means matching thousands of BOM lines to emission factors and chasing suppliers for data. AI maps materials and purchases to LCA datasets, fills supplier data gaps with documented estimates, calculates cradle-to-gate footprints per product, and assembles the results for customer requests, CBAM reporting and Digital Product Passports (DPP). Typical benefit: a footprint per SKU in days instead of an expert consultancy project per product.
Problem & value
- Today: sustainability or quality staff, or external consultants, build an LCA model per product family in spreadsheets or LCA tools (SimaPro, openLCA, GaBi/Sphera). Each BOM line is matched manually to a dataset; supplier-specific data is rare; results go stale after the next design or supplier change. Large customers (automotive, consumer goods, retail) send PCF requests to suppliers in many formats.
- AI value: (1) automated, explainable BOM-to-factor matching (multilingual descriptions, part numbers, HS/CN codes); (2) gap filling with spend-based or proxy factors, clearly flagged with a data-quality rating; (3) scaling from a few pilot products to the whole portfolio; (4) re-calculation on each BOM or supplier change; (5) one data model feeding PCF exchange, CBAM and DPP.
- Why now: ESPR (Regulation (EU) 2024/1781) creates the Digital Product Passport framework; the Battery Regulation (EU) 2023/1542 requires a battery passport and carbon footprint declarations for relevant batteries (first DPP deadline: February 2027 for EV, light means of transport and industrial batteries above 2 kWh); CBAM entered its definitive period on 1 Jan 2026 and requires embedded emissions of imported goods [1][2][3]. Delegated acts for other product groups were still being prepared in 2026; dates vary by source, so verify them on the Commission site [1].
- Distinct from DATA-012: that use case reports at company level (CSRD); here the unit of analysis is the product, the data are BOMs and LCA databases, and the method is emission-factor matching.
Variants
- Manufacturing / automotive supply chain: PCF exchange with OEMs via the PACT standard or Catena-X; supplier PCF ingestion into the BOM model [4][5].
- Batteries: carbon footprint declaration and battery passport data (see PRED-026 for state-of-health data in the passport).
- CBAM importers and exporters (steel, aluminium, cement, fertiliser, hydrogen, electricity): embedded emissions per consignment, supplier installation data, default values where actual data is missing.
- SMB supplier: a small supplier answers a customer request with a simplified, template-driven PCF, using industry-average factors and a guided wizard.
- Food and retail: recipe- and ingredient-based footprints from agricultural LCA data (e.g. Agribalyse), thousands of own-brand SKUs, label and shelf communication (Carbon Maps is a French example).
How it works
- BOM / purchase / process data (PLM, ERP, MES, supplier portals)
- normalisation (units, languages, part-number
- material/process description)
- matching (embedding retrieval over LCA/emission-factor catalogues
- reranker/LLM picks and justifies)
- gap filling (supplier PCF if available
- proxy dataset
- spend-based factor; each flagged)
- calculation engine (deterministic LCA maths: cradle-to-gate, allocation, energy mix, transport)
- data-quality rating + uncertainty
- human review of low-confidence/high-impact lines (hotspot-first)
- verification-ready report; export to PACT/Catena-X, CBAM, DPP data carrier
- The calculation is deterministic. The LLM or embedding model only proposes mappings, extracts data from supplier documents and drafts explanations. It must never invent emission factors or numbers; factors come from licensed or open databases with versioned IDs.
- Matching: hybrid search (text + code classes such as HS/CN, ECLASS, ETIM, see DATA-010) over a vector index of dataset descriptions; an LLM or cross-encoder selects and records its rationale.
- Human-in-the-loop: an LCA practitioner approves mappings above an impact threshold and signs off the model; external verification for claims, CBAM and battery declarations.
Data preparation
The core data are bills of materials (PLM or ERP), purchase records, supplier master data, energy and transport figures per site, and supplier PCF questionnaires. In an SMB these live in the ERP, in Excel BOMs, in mail attachments and in supplier PDFs. Typical gaps: BOMs without material or weight, free-text part descriptions in several languages, duplicate part numbers, missing energy data per product, and supplier declarations only as scans.
Preparation steps: (1) export BOMs and purchase data from the ERP (PREP-005); (2) turn Excel BOMs into tidy tables with units and weights (PREP-011); (3) OCR supplier declarations (PREP-006, DOC-014); (4) clean part descriptions and add glossary and synonym lists (PREP-015); (5) assign HS/CN or ECLASS codes for matching (DATA-010); (6) name a data owner per site for energy data (PREP-004). Tools: Excel/Power Query or OpenRefine for cleaning, Paperless-ngx or Azure Document Intelligence for scans, the ERP export for the rest.
Budget a large share of the project, roughly 50-60 percent of effort, for this work: the calculation is quick, but weak BOM and supplier data decide the quality and the data-quality rating. Start with the top products by volume and turnover (hotspot-first), not the whole portfolio. See also GUIDE-003 and DATA-003.
Implementation paths
Which path for whom
- SMB (A): mature EU PCF platforms (Makersite, Plan A, One Click LCA) exist at SMB prices and customer PCF requests follow a standard format; a small firm has no LCA team to build its own.
- Enterprise (B): large product portfolios and own PLM/ERP/Catena-X integration justify a pipeline on cloud APIs around licensed LCA databases, with the deterministic calculation core kept in-house.
- Developer (C): open emission-factor data and open-weight embedding models run on HW2 and let you build matching and PACT export without vendor lock-in.
Path A — Buy (recommended for SMBs)
- PCF/LCA automation platforms: Makersite (DE/UK), Carbon Maps (FR), Sphera, One Click LCA (FI), Ecochain (NL), Climatiq (DE, emission-factor API), Plan A and Sweep (EU), SAP Sustainability Footprint Management, Siemens SiGREEN, plus Catena-X-ready PCF services. Many also offer DPP modules.
- Check: coverage of your product type, ISO 14067 and PACT/Catena-X conformance, explainable mappings, database licences, EU hosting, verification support, CBAM module.
- Choose A when: you need results for customer requests within months and have no LCA team.
Path B — Build on cloud APIs
- Models: an EU-hosted frontier LLM or multimodal API (Mistral, or Claude/Gemini/GPT in EU regions) with structured outputs for classification, document extraction and justification; an embedding model for retrieval.
- Data: Climatiq API or licensed ecoinvent/EXIOBASE/Agribalyse; the calculation engine remains your own deterministic code or openLCA (via its IPC/Python API).
- Orchestration: n8n / Python pipeline, PLM and ERP connectors, PACT-compatible API endpoints.
- Choose B when: you have unusual products or want to own the model, and have a data engineer plus an LCA expert.
Path C — Self-host (open source)
- LCA engines: openLCA, Brightway, Activity Browser; open data such as Agribalyse, ProBas, and open parts of EXIOBASE.
- PACT: open reference implementation and the PACT data model/API specification for exchange [4]; Eclipse Tractus-X for Catena-X connectors [5].
- AI: open-weight embedding model (e.g. BGE-M3, multilingual-e5) plus an open LLM (Qwen3, Mistral, Llama class) via vLLM or Ollama; vector store pgvector/Qdrant.
- Hardware: HW2 is enough for matching and extraction; commercial datasets such as ecoinvent require a licence even when self-hosted.
Path D — Custom training
Worth it only for a large catalogue with consistent history: fine-tune an embedding or classification model on your own approved BOM-to-dataset mappings (thousands of reviewed pairs), or train a surrogate model that predicts the footprint of new variants from parametric design data. Keep it as an accelerator behind human sign-off.
State of the art (2026)
- Methods: ISO 14067, ISO 14040/14044, GHG Protocol Product Standard, EU PEF rules [6].
- Data exchange standardised: the PACT Pathfinder data model/API (WBCSD) and Catena-X PCF rulebook enable machine-readable supplier PCFs; automotive and chemicals are furthest along [4][5].
- Digital Product Passport: the EU registry and data-carrier approach (QR code, unique identifier) are being implemented; product-specific data requirements come in delegated acts, so schemas will change [1][7].
- Limits and failure modes: wrong dataset match (a similar-sounding material with a very different footprint), geography and energy-mix mismatches, double counting across tiers, spend-based factors that hide design improvements, outdated LCA database versions, and hallucinated factors if the LLM is allowed to produce numbers. Secondary data can overstate or understate; supplier-specific primary data remains the biggest accuracy lever.
Shared resources & combinations
- Same stack (embedding-text + vector-db + llm-general): DATA-010 product data enrichment (ECLASS/ETIM), DATA-015 spend classification (spend-based fallback), DATA-003 master data management, DATA-007 catalogue metadata.
- Upstream/downstream: DATA-012 aggregates product results into company-level reporting (CSRD, Scope 3 category 1/11); DATA-020 for regulatory reporting patterns; PRED-026 battery state-of-health in the battery passport; VIS-049 for deforestation-free supply chains; DATA-016 for industrial energy and process data.
Effort, cost & KPIs
- Path A: weeks for first products, quarters for the portfolio; per-product or per-seat licences plus LCA database fees.
- Path B: 3-6 months for a production pipeline; cost drivers are data cleaning, database licences, expert review and verification.
- KPIs: share of BOM mass/impact covered by supplier-specific data, share of lines mapped without correction, review time per product, number of products with a valid PCF, time to answer a customer PCF request, data-quality rating, verification findings, DPP data completeness.
People & change
Involve management first for budget and the decision which customer or regulatory demands (CBAM, DPP, customer PCF requests) set the scope. The department lead (sustainability, quality or production) owns the process; data owners for ERP, PLM and energy data must release exports early. IT provides interfaces; suppliers and customers are involved through questionnaires and PCF exchange, so announce data requests in good time. Legal/compliance checks claims, CBAM and verification duties.
Works council: normally not required, because the system processes product and supplier data, not employee behaviour (§87(1) no. 6 BetrVG). Inform it anyway under §90(1) no. 3 if work processes change noticeably. Personal data is not expected; a DPO check is only needed if supplier contact data are processed.
AI literacy: support staff with a short training (GUIDE-006) on what the LLM may do (propose mappings) and what it must not do (invent factors). Pilot with one product line and two to three practitioners, compare against a manual calculation, and publish the review rules. Typical concern: "the number is a black box". Address it with traceable mappings, versioned factors and human sign-off above an impact threshold. See GUIDE-009 and GUIDE-008.
Risks, regulation & compliance
- EU AI Act: minimal risk; no Annex III use. AI literacy duty (Art. 4). No Art. 5 issue.
- Greenwashing: the Empowering Consumers Directive (EU) 2024/825 restricts generic environmental claims and offsetting-based claims; Green Claims proposals were uncertain in 2026. Every public claim needs substantiation and, where required, third-party verification.
- Regulatory data: CBAM and Battery Regulation declarations have legal consequences and verification requirements; AI-estimated values can only be used where the rules permit (default values, documented methods).
- GDPR: mostly no personal data; supplier contact names and employee data in questionnaires are exceptions.
- Mitigations: deterministic calculation, versioned factors, audit trail of AI suggestions versus approved mappings, hotspot-based human review, independent verification.
Getting started
- Pick 5-10 representative products and the customer or regulatory request that triggers the work (customer PCF request, CBAM, battery passport).
- Clean the BOM and purchase data; agree units, system boundary (cradle-to-gate) and standard (ISO 14067 / PACT).
- Run a pilot with one Path A tool and a manual expert model in parallel; compare mappings and totals.
- Define review thresholds (impact share, confidence) and start requesting primary data from the top suppliers.
- Prepare the export (PACT/Catena-X, CBAM, DPP) and track DPP delegated acts for your product group.
Sources
- Regulation (EU) 2024/1781 (ESPR, Digital Product Passport framework) — https://eur-lex.europa.eu/eli/reg/2024/1781/oj
- Regulation (EU) 2023/1542 (Battery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1542/oj
- European Commission, Carbon Border Adjustment Mechanism — https://taxation-customs.ec.europa.eu/carbon-border-adjustment-mechanism_en
- WBCSD PACT, Pathfinder Framework and data exchange specification — https://www.carbon-transparency.org
- Catena-X Automotive Network / Eclipse Tractus-X — https://catena-x.net and https://eclipse-tractusx.github.io
- EU Product Environmental Footprint method — https://green-business.ec.europa.eu/environmental-footprint-methods_en
- Council of the EU working document on DPP rollout (WK 14231/2025) — https://data.consilium.europa.eu/doc/document/WK-14231-2025-INIT/en/pdf
- openLCA — https://www.openlca.org
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj