Spend classification & procurement spend analytics

Companies buy from thousands of suppliers booked in several ERPs with free-text line descriptions, so nobody can answer "how much do we spend on temporary labour, and with whom?". ML and LLMs normalise supplier names and map every AP, PO and card line to a spend taxonomy (UNSPSC, eCl@ss or an in-house tree). The resulting spend cube exposes savings potential, maverick buying and supplier consolidation options, and delivers the activity data for spend-based Scope 3 estimates.

Problem & value

• Today: spend data sits in many ERPs, P2P tools and card feeds. Supplier names appear in dozens of spellings, GL accounts are a poor proxy for what was bought, and line texts are short, multilingual and abbreviated. Category managers rely on consultants or yearly Excel projects that are stale on delivery.
• AI value: classification refreshed continuously (monthly or weekly) instead of annually, with human corrections feeding back into the model. Typical outcomes are visibility of tail spend, supplier consolidation, contract and catalogue compliance, price benchmarking across entities, and category-level baselines for negotiation. Vendors publish customer cases with savings and coverage figures (e.g. Sievo customer stories [1]); treat them as vendor claims and measure your own baseline.
• Sustainability angle: spend-based Scope 3 category 1 (purchased goods and services) multiplies spend per category by an emission factor per currency unit; it needs exactly this category-level spend data [4]. It is the coarsest GHG Protocol method, suited to screening and hotspot finding, not to supplier-specific reporting.
• Scope boundary: this is historical transaction classification. Sell-side product master data is DATA-010; executing sourcing is AGENT-005; fraud detection is ANOM-009.

Variants

• Spend classification service: one-off or periodic cleansing of an ERP extract, delivered as a classified file (consultancy or SaaS).
• Continuous procurement analytics (spend cube): automated ingestion from multiple ERPs, dashboards for category managers, savings tracking and alerts on off-contract spend.
• Public sector: classification against CPV codes for TED/e-procurement reporting and transparency obligations; see also ANOM-020.
• SMB: a lightweight version inside accounting or expense tools, or an LLM-assisted spreadsheet workflow (DATA-006) classifying a few thousand lines a year.
• Scope 3 variant: the same classified cube mapped to emission-factor categories (see DATA-012, DATA-021).
• Legal spend: outside-counsel invoice review is its own use case (LANG-039).

How it works

1. extract (ERP AP/PO, P-card, expenses, contracts)
2. harmonise (currency, language, entity, dates, deduplicate)
3. supplier normalisation (fuzzy + embedding match, VAT ID / DUNS / LEI, parent hierarchy)
4. line classification (rules
5. ML/embedding classifier
6. LLM for hard lines)
7. human review of low-confidence and high-value lines
8. spend cube + analytics (savings, maverick, consolidation, price variance, Scope 3)
9. feedback loop: corrections become rules and training data

• Supplier normalisation: blocking plus string similarity, then embeddings and external identifiers (VAT ID, DUNS, LEI, Handelsregister) to link entities to parents. Same family of techniques as DATA-003.
• Classification: layered. Deterministic rules (GL account, vendor-to-category, contract or catalogue match) resolve the easy majority. A fine-tuned encoder or embedding plus nearest neighbour classifier handles the bulk. An LLM with the taxonomy as retrieval context handles ambiguous or multilingual lines and explains its choice. Hierarchical taxonomies (UNSPSC has segment, family, class, commodity levels [2]) allow partial credit: confident at family level, review at commodity level.
• Human-in-the-loop: category experts review low-confidence lines and the top suppliers by value; approved mappings are stored as rules so each correction is made once.

Data preparation

Needed are AP invoice lines, purchase orders, card and expense transactions, the supplier master and, where available, contracts and catalogues. They typically sit in the ERP (SAP, Dynamics, DATEV, Business Central), in the accounting system, in PDFs in the DMS or mail, and in Excel lists kept by buyers. Typical gaps: free-text line descriptions, one supplier under many spellings and IDs, missing or inconsistent GL-to-category mapping, several ERPs or entities with different currencies and languages, and invoices that exist only as scans.

Steps: (1) export lines for 12 to 24 months through ERP exports or connectors (PREP-005); (2) harmonise currency, language, dates and entity; (3) deduplicate and normalise suppliers using VAT ID or DUNS (DATA-003); (4) define the target taxonomy (UNSPSC, eCl@ss or a company scheme, PREP-019); (5) harvest existing labels such as GL accounts and approved categories as training data (PREP-020); (6) set aside a reviewed test sample (PREP-018). Scanned invoices need OCR first (DOC-014 and PREP-006). Expect roughly 40 percent of effort on extraction and supplier cleanup, 20 percent on taxonomy and labels, the rest on review and analytics. GUIDE-003 offers a readiness checklist.

Implementation paths

Which path for whom

• SMB: Path B, because spend-analytics products target large enterprises and are priced accordingly, while an SMB can classify an ERP export with a cloud LLM, embeddings and a workflow tool such as n8n.
• Enterprise: Path A, because mature spend-analytics suites handle many ERPs, taxonomies and supplier hierarchies and are a commodity function.
• Developer: Path B, because cloud APIs plus open taxonomies (UNSPSC, eCl@ss) give a fast, flexible build for a product or service.
• Public sector: Path C, because procurement rules and sovereignty favour a self-hosted stack on a sovereign cloud, and a small open model suffices for classification.

Path A — Buy

• Dedicated spend analytics: Sievo (Finland, EU), SpendHQ, Jaggaer, Coupa (spend analysis / community intelligence), SAP Ariba Spend Analysis, GEP SMART, Zycus, Tradeshift. Specialist classification services: Spend Matters-covered vendors and consultancies; Ivalua (France, EU) and Jaggaer (with European roots) cover spend visibility within source-to-pay suites.
• Check: EU hosting, taxonomy support (UNSPSC, eCl@ss, CPV, custom), ERP connectors, multilingual accuracy, ability to export corrections and mappings (avoid lock-in), Scope 3 emission-factor mapping.
• Choose A when: more than one ERP, many suppliers, and no in-house data science team.

Path B — Build on cloud APIs

• Classification: embeddings (EU-hosted options such as Mistral Embed, Azure OpenAI or Vertex AI in EU regions) to retrieve candidate taxonomy nodes, then an LLM with structured output picks the code and gives a confidence and rationale. Batch APIs make re-classifying millions of lines affordable.
• Platform: warehouse (Snowflake, BigQuery, Databricks, Microsoft Fabric) with dbt for the spend cube; BI via Power BI, Looker or Metabase; orchestration with Airflow or Dagster.
• Choose B when: the data team exists and the taxonomy is proprietary.

Path C — Self-host (open source)

• Models: multilingual embeddings (BGE-M3, multilingual-e5, Qwen3-Embedding), a fine-tuned XLM-RoBERTa or German BERT classifier, and an open LLM (Qwen3, Mistral Small, Llama) via vLLM or Ollama for hard lines.
• Stack: DuckDB or PostgreSQL with pgvector, Splink for probabilistic supplier matching, dbt, Metabase or Superset, Label Studio for review.
• Hardware: HW2 is sufficient; classification of millions of short lines is mostly embedding and small-model inference.

Path D — Custom training

Worthwhile when you hold several hundred thousand lines already labelled by category managers or a previous consultancy project. Fine-tune an encoder classifier per taxonomy level; labelled history is the main asset. Retrain when the taxonomy changes.

State of the art (2026)

• LLM plus retrieval over the taxonomy is the dominant approach for the long tail, with fine-tuned small classifiers for cost-effective bulk work. Reported vendor accuracy depends on taxonomy depth; deep commodity-level codes are materially harder than top-level categories.
• Coverage vs. accuracy: short, ambiguous text ("service fee", "misc") cannot be classified from text alone; best systems combine supplier, GL account, cost centre, amount pattern and contract context.
• Taxonomies: UNSPSC is licensed and versioned by GS1 US [2]; eCl@ss is the German-origin standard for industrial goods [3]; CPV is mandatory in EU public procurement. Mapping between them remains partly manual.
• Failure modes: LLMs invent plausible but non-existent codes (constrain output to valid codes), drift when taxonomy versions change, mis-classify multilingual abbreviations, and confuse the parent company with the invoicing subsidiary.
• Scope 3 limits: spend-based factors (e.g. EXIOBASE-derived or national input-output datasets) are inflation- and price-sensitive; do not use them for supplier-level claims [4].
• Agentic analytics: natural-language access to the cube (DATA-001) and agents that act on tail-spend findings (AGENT-005, AGENT-029) are emerging.

Shared resources & combinations

• Data cleansing: DATA-003 (deduplication, MDM) shares supplier matching; DATA-007 (catalogue/metadata) and DATA-010 (ECLASS/ETIM classification) share taxonomy retrieval.
• Downstream: DATA-012 and DATA-021 (Scope 3 / carbon footprint), ANOM-009 (fraud on classified spend), ANOM-011 (supplier risk by category), AGENT-005 (sourcing events for the consolidation opportunities found), DATA-001 (conversational BI on the cube).
• Infrastructure: one embedding service plus warehouse serves several data use cases.

Effort, cost & KPIs

• Path A: typically 2-4 months to a first classified cube across several ERPs; SaaS licence tied to spend under management or data volume.
• Path B: 2-4 months with a small data team; LLM cost is modest because rules and classifiers resolve most lines before the LLM sees them.
• Path C: 3-6 months; mostly engineering and review tooling.
• KPIs: classification coverage (% of spend categorised), accuracy at category level on a held-out audit sample, supplier count reduction after normalisation, share of spend under contract, identified vs. realised savings, review hours per million lines, time to refresh.

People & change

Involve management for goals and budget (savings target, Scope 3 reporting) and the head of procurement as process owner, who owns the taxonomy and decides on category rules. Finance and the ERP data owners supply and validate exports; IT handles access and integration. Buyers are the end users: let them correct classifications, because their corrections feed the rules. Suppliers are rarely involved directly, but consolidation insights affect negotiations, so agree on confidentiality of price data.

Works council: usually no, because the system analyses spend and suppliers rather than employee behaviour. Check it if buyer-level reports (for example per buyer maverick-spend rankings) are planned, since then §87(1) no. 6 BetrVG could apply; keep reports at category and supplier level. §90(1) no. 3 still calls for early information if work processes change materially. DPO: low risk, as supplier contact names and card holder data are the only personal data; sign an AVV (Art. 28 GDPR) with any cloud vendor.

Communicate that the tool takes over tedious coding of lines, not buying decisions. Pilot with one category group (for example indirect spend) for 6 to 8 weeks, measure classification accuracy and review time, then extend. Typical concern: "the AI will misclassify". Answer with confidence thresholds, mandatory review for high-value lines and visible correction logs. Link: GUIDE-009 for the pilot and GUIDE-006 for short training of buyers.

Risks, regulation & compliance

• EU AI Act: minimal risk; no Annex III category. AI literacy (Art. 4) applies [5].
• GDPR: mostly company data, but sole traders, freelancers and card/expense lines can identify individuals (travel, employee-linked purchases). Use data processing agreements, minimise employee fields and avoid sending free-text with names to non-EU APIs.
• CSRD / ESRS E1: spend-based Scope 3 figures need documented method, factor source and uncertainty; the Omnibus simplification changes scope and timing, so check current status with the reporting team.
• Public procurement: CPV and TED reporting follow Directive 2014/24/EU rules.
• Business risks: wrong categories lead to wrong savings targets; classification bias towards large suppliers; vendor lock-in on proprietary mappings; confidentiality of contract prices.
• Mitigations: audit sample each cycle, constrain outputs to valid codes, keep mapping tables exportable and versioned, log classifier version per record.

Getting started

1. Pick the target taxonomy and the 5-10 priority categories; define the question the cube must answer.
2. Extract 12-24 months of AP, PO and card data from the main ERPs into one warehouse schema.
3. Normalise suppliers first (VAT ID and parent matching): it often produces the first surprises.
4. Classify a 2,000-line sample with rules, an LLM and a vendor pilot; score all against an expert-labelled audit set.
5. Set up the correction loop and a monthly refresh; add Scope 3 factors once categories are stable.

Sources

1. Sievo — spend analytics and customer stories — https://sievo.com/
2. UNSPSC — GS1 US (UN Standard Products and Services Code) — https://www.unspsc.org/
3. eCl@ss — product classification standard — https://eclass.eu/
4. GHG Protocol — Technical Guidance for Calculating Scope 3 Emissions (Category 1, spend-based method) — https://ghgprotocol.org/scope-3-technical-calculation-guidance
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Splink — probabilistic record linkage (UK Ministry of Justice) — https://github.com/moj-analytical-services/splink
7. EU CPV — Common Procurement Vocabulary (SIMAP) — https://simap.ted.europa.eu/web/simap/cpv