Product data enrichment & classification (PIM, ECLASS, ETIM)

Manufacturers, wholesalers and online shops must deliver clean, standardised product data to dozens of channels, each with its own classification and attribute schema. Onboarding a single new article by hand — finding the ECLASS or ETIM class, filling 20–200 features, converting units, writing texts — takes minutes to hours. AI reads supplier files, datasheets and images, proposes the class and attribute values with confidence scores, and routes uncertain items to a data steward. The result is faster time-to-market and better search, filters and procurement matching.

Problem & value

• Today: product data arrives as inconsistent Excel files, PDFs, images and web pages. Data stewards map them manually to the PIM model, look up classes in the ECLASS or ETIM catalogue, harmonise units ("mm" vs "cm", "230V" vs "230 V AC"), and rework data per channel (own shop, marketplaces, wholesalers' BMEcat requirements, public tenders).
• AI value: (1) classification — mapping an article to a leaf class among thousands, (2) attribute extraction and normalisation — free text and datasheets into typed feature values, (3) enrichment — missing attributes, translations, descriptions, tags, (4) quality control — detecting contradictions between text, image and structured fields.
• Why standards matter: ECLASS and ETIM are established in German and European technical trade and industry; they are required by many B2B customers, platforms and tender systems. ECLASS also underpins digital product passports and Asset Administration Shells via semantic identifiers (IRDI) [1][2][3].
• Benefit evidence: vendors report large time savings, but independent, comparable figures are scarce. Measure on your own sample (stewarding minutes per article, share of auto-accepted items).

Variants

• Technical wholesale / manufacturer (ECLASS, ETIM): electrical, HVAC, tools, fasteners; very many technical attributes with units and value lists.
• Retail / e-commerce (GS1 GPC, Google product taxonomy, marketplace categories): fewer attributes, more descriptive text, images and variants (size, colour).
• Procurement-side: classify supplier catalogues on inbound for catalogue consolidation, see DATA-015.
• Public sector / tendering: classification supporting standardised item descriptions.
• Private use: not relevant (a small shop owner is covered under SMB).

How it works

1. supplier files / PDFs / images / web
2. ingest + parsing (tables, datasheets, OCR)
3. candidate class retrieval (embeddings + keyword search over class catalogue)
4. LLM/classifier picks class from shortlist, with rationale + confidence
5. attribute extraction per class schema (feature list, units, allowed values)
6. normalisation + rule validation (units, ranges, value lists, cross-field checks)
7. enrichment (translation, texts, tags, image-derived attributes)
8. steward review queue (low confidence, conflicts, new classes)
9. PIM write-back + channel export (BMEcat, ETIM XML, marketplace feed)

• Retrieve-then-classify: classification catalogues have tens of thousands of classes, so the class list is not placed in the prompt. A bi-encoder/embedding search (plus BM25 on class names and synonyms) builds a shortlist of ~10–30 candidate classes; an LLM or fine-tuned cross-encoder decides. This is the dominant pattern in 2026 and keeps cost and hallucinated class codes low.
• Schema-driven extraction: once the class is known, its feature definitions (name, datatype, unit, value list) become the JSON schema for structured output. Values not found in the source are returned as null — never guessed.
• Vision: a VLM reads packaging, labels, rating plates and drawings for missing attributes (e.g. voltage, IP rating, colour).
• Human-in-the-loop: data stewards approve below-threshold items, new class proposals and safety- or compliance-relevant attributes; every correction is logged and reused as examples.

Data preparation

Product data typically lives in the ERP item master, supplier Excel/CSV price lists, PDF data sheets and catalogues, image folders on file shares and sometimes a PIM. Typical gaps: duplicate items with different numbers, missing or free-text attributes, mixed units and languages, scanned data sheets, no ECLASS/ETIM class at all, and unclear ownership of master data. Supplier files arrive in a new layout every time.

Concrete steps: inventory sources and fields and decide the target classification and attribute set (GUIDE-003, PREP-002); export the item master (PREP-005); OCR data-sheet PDFs (PREP-006); turn supplier spreadsheets into tidy tables (PREP-011); write a glossary of attribute names, units and synonyms (PREP-012, PREP-015); build a gold-standard set of 200-500 correctly classified items with the product experts (PREP-016, PREP-018). Tools: Paperless-ngx or an OCR service, pandas/OpenRefine for cleaning, a deduplication step on item numbers and descriptions. Related: DATA-003, DATA-008. Rough share of effort: 50-60 percent; the gold-standard set doubles as the later quality test.

Implementation paths

Which path for whom

• SMB: Path B. The enrichment must plug into the company's own PIM/ERP and supplier formats, and no off-the-shelf product covers this cheaply; cloud LLM APIs with a thin pipeline fit.
• Enterprise: Path B. Large catalogues make seat prices of products high, while an API pipeline integrated with the central PIM scales per item and can move to an internal platform later.
• Developer: Path C. Open-weight models and embeddings on HW2 give controllable per-item cost for bulk classification.

Path A — Buy (off-the-shelf)

• PIM/MDM platforms with AI features: Akeneo (FR, EU; Product Cloud with AI enrichment), Pimcore (AT, EU; open-source core with AI integrations), Informatica Product 360 / MDM, Salsify, inRiver (SE, EU), Contentserv (DE/CH, EU), Stibo Systems (DK, EU), Syndigo, Plytix.
• Classification and data-quality specialists (many EU): ETIM- and ECLASS-focused data-quality tools from industry suppliers and consultancies, plus services of the standard bodies (ECLASS e.V., ETIM International/ETIM Deutschland) such as classification search and validation tools.
• Choose A when: you already run or plan a PIM, volumes are moderate, and the vendor supports your target standard in the required version (check ECLASS release and ETIM release support).

Path B — Build on cloud APIs (recommended default)

• Models: frontier multimodal LLMs with structured outputs (Anthropic Claude, Google Gemini, OpenAI GPT, Mistral — EU provider) for extraction and class decision; small/cheap models for bulk runs. Use EU regions or EU providers where contractual terms demand it.
• Retrieval: embeddings (e.g. multilingual-e5, BGE-M3, Cohere Embed, Mistral Embed) in pgvector, OpenSearch, Qdrant (DE/EU origin) or Weaviate (NL/EU origin).
• Reference data: load ECLASS (licensed, available in CSV/XML via the ECLASS download service) and ETIM (available through ETIM's open classification and the ETIM API/download) into a class catalogue with synonyms and feature definitions [1][2].
• Orchestration: n8n, Dagster/Airflow or the PIM's own workflow/webhook API, writing back through the PIM REST API.
• Choose B when: you need custom logic, several standards at once and tight PIM integration.

Path C — Self-host (open source)

• Models: Qwen3, Llama, Mistral/Mixtral-family or Gemma open-weight LLMs served with vLLM or Ollama; open VLMs (Qwen-VL family) for images; Docling for datasheet parsing; BGE-M3 or multilingual-e5 embeddings; GLiNER or fine-tuned BERT-class models for attribute NER.
• Stack: Pimcore or Akeneo Community Edition as PIM, Qdrant/pgvector, Label Studio for review.
• Hardware: HW2 is enough for catalogues of a few hundred thousand articles when run as batch; HW3 for larger VLMs. Operations effort is moderate; the licence for ECLASS content still applies.

Path D — Custom training

Worth it for large catalogues (hundreds of thousands of articles) with stable taxonomies: fine-tune an embedding model and a classifier on your own historic class assignments (thousands of labelled articles; steward corrections are ideal training data). Gains: lower latency, lower cost per item, higher accuracy on domain jargon. Research on LLM-based product categorisation and attribute-value extraction is active [4][5].

State of the art (2026)

• Retrieval + LLM re-ranking is the practical standard for large taxonomies; direct prompting with the whole taxonomy does not scale and hallucinates codes. Fine-tuned small models match large LLMs on top-level categories, but deep leaf classes remain hard.
• Attribute-value extraction with schema-constrained decoding is robust for explicit values; implicit or conflicting values (variants, bundles, "approx." statements) are the main error source.
• Multimodal enrichment: VLMs extract attributes from images and rating plates, but small print, reflections and unit symbols cause errors.
• Standards move on: ECLASS and ETIM publish regular new releases; class and feature changes require migration (mapping old to new class). LLMs support the mapping between versions and between standards (ETIM ↔ ECLASS), which is laborious manually. Pre-built mappings exist and should be preferred as ground truth.
• Digital product passport: the EU Ecodesign for Sustainable Products Regulation (ESPR) makes structured, machine-readable product data a legal topic for many product groups; clean PIM data is the foundation, see DATA-021 [6].
• Limits: invented attribute values, false precision in units, wrong leaf classes with plausible names, and silent errors on safety-relevant data. Always validate against value lists and numeric ranges, and measure per-attribute accuracy rather than overall.

Shared resources & combinations

• Same stack (llm-general + embedding-text + vector-db): DATA-003 (deduplication/MDM — match product duplicates before enrichment), DATA-007 (catalogue metadata), DATA-015 (spend classification, UNSPSC/ECLASS on the buyer side), KNOW-005 (semantic product search consumes the enriched attributes).
• Upstream/downstream: DATA-022 (supplier data onboarding and schema mapping), DOC-003 (datasheet and form parsing), LANG-001 (product descriptions), GEN-001 (product imagery), DATA-021 (product carbon footprint, digital product passport).
• One embedding index of class catalogues can serve both ECLASS and ETIM mapping and DATA-015.

Effort, cost & KPIs

• Path A: weeks for a pilot, plus PIM licence; AI add-ons are usually priced per module or usage.
• Path B: 2–4 months for classification + attribute extraction in one or two languages; LLM cost per article is typically small relative to steward time (calculate with your token counts).
• Path C/D: 4–8 months including evaluation sets and review tooling; GPU server and MLOps.
• Main cost driver: building the evaluation set and the review workflow, not model calls.
• KPIs: class accuracy at leaf level (on a golden set), per-attribute precision/recall, auto-acceptance rate, steward minutes per article, PIM completeness score, time-to-publish, channel rejection rate, return rate caused by wrong specifications.

People & change

Involve the product-data or purchasing lead as process owner from day one, plus the data owners of the item master and the end users who maintain the PIM. Management decides how much automatic write-back is allowed; IT integrates PIM/ERP and API access. Suppliers should be told early if they are asked for cleaner data formats or if their files are processed by AI. The DPO is rarely needed because product data contains no personal data, but check the AVV with the API provider (GUIDE-010).

Works council: normally no. The system processes product data and does not monitor employees; still inform the council in good time under §90(1) no. 3 BetrVG if work processes change noticeably, and check whether edit logs per user exist, which could trigger §87(1) no. 6.

Support staff AI literacy (Art. 4 AI Act) with a short session on typical errors such as invented attributes and wrong units (GUIDE-006). Pilot with one product group and compare to manual work (GUIDE-009). Typical concern: "my job is manual data entry". Answer by moving staff to review and exception handling, where confidence scores route uncertain items to a person. Keep humans in control: low-confidence and safety-relevant attributes are always approved manually.

Risks, regulation & compliance

• EU AI Act: minimal risk — product data processing is not an Annex III use; AI literacy (Art. 4) applies. Generated product texts shown to consumers are normally not Art. 50 content, but check marketing claims separately [7].
• GDPR: mostly no personal data. Exceptions: supplier contact data, images with people, review text; keep these out of prompts or use a DPA and EU processing.
• Licences: ECLASS content is licensed and ETIM has its own terms; check whether sending class content to third-party model APIs and embedding it is permitted by your licence.
• Product liability and safety: wrong values (voltage, load rating, material, allergens) can cause harm and legal exposure; label safety-critical attributes as "must be verified" and keep the manufacturer-supplied value as source of truth. Consumer-law claim rules (misleading information) apply to enriched texts.
• Data quality risks: hallucinated attributes, prompt injection via supplier files or scraped pages (treat all inputs as untrusted), drift after taxonomy updates.
• Mitigations: provenance per value (source file, page, model, version), null over guess, validation rules, golden-set regression tests, audit log.

Getting started

1. Pick one product group and build a golden set of 300–500 articles with verified class and attributes; document the standard versions used.
2. Load the class catalogue (ECLASS or ETIM) with synonyms and feature definitions into a search index.
3. Prototype retrieve-then-classify and schema-based extraction; measure leaf accuracy and per-attribute accuracy against the golden set.
4. Set confidence thresholds, build the steward review queue, and write results with provenance into the PIM.
5. Roll out group by group; re-run the golden set after every model, prompt or taxonomy update.

Sources

1. ECLASS e.V. — standard overview and releases — https://eclass.eu/
2. ETIM International — classification model and releases — https://www.etim-international.com/
3. Industrial Digital Twin Association — Asset Administration Shell — https://industrialdigitaltwin.org/
4. Brinkmann et al. — Using LLMs for product attribute value extraction (WDC, University of Mannheim) — https://arxiv.org/abs/2403.02130
5. Peeters, Bizer — Entity matching using large language models — https://arxiv.org/abs/2310.11244
6. Regulation (EU) 2024/1781 (Ecodesign for Sustainable Products Regulation, ESPR) — https://eur-lex.europa.eu/eli/reg/2024/1781/oj
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Akeneo — product experience platform — https://www.akeneo.com/
9. Pimcore — open-source data management platform — https://pimcore.com/
10. Docling project — https://github.com/docling-project/docling Accessed 2026-10; vendor feature claims are not independently verified and should be re-checked before purchase decisions.