SStructera

What is procurement spend classification?

Last updated: 2026-05-20

Published 20 May 2026 · 7 min read

If you've ever opened a spreadsheet with 50,000 lines of supplier payments and tried to figure out where the money actually went, you already know procurement spend classification is a slog. This post walks through what it is, why it matters, and how to do it well in 2026 — with or without an AI tool.

What is procurement spend classification?

Procurement spend classification (also called spend categorisation, spend categorization in US spelling, or just “spend tagging”) is the practice of taking every line item in your procurement data and assigning it to a category in a structured hierarchical taxonomy. The taxonomy usually has 2-4 levels: a top tower like IT & Software, a family like Cloud infrastructure, and a leaf like Cloud compute services.

The classification is the foundation. Once it's done you can slice spend by category, run a spend cube, identify duplicate suppliers, surface tail spend, and benchmark against industry data. Without it you have a transaction log, not insight.

Why does it matter?

Three things come unstuck without proper classification:

  • Tail spend goes unmanaged. Without categories you can't see that 23% of your spend is going through 80% of your suppliers at uncompetitive rates, because you can't aggregate across vendors.
  • Reporting takes forever. Every board paper needs someone to manually pivot transactions by hand. With classified data, the pivot is a Power BI / Tableau report that takes seconds.
  • Negotiating leverage evaporates. When you can't show a supplier “you charge us GBP 50/user for the same service three of your competitors charge GBP 30 for”, you can't negotiate.

How spend classification used to be done

Until recently, the options were:

  1. Manual. An analyst tags every row in Excel. Works for a few thousand rows, breaks at scale, drifts across analysts.
  2. Rule-based. A long file of regex / supplier-name rules that map “AWS”, “Amazon Web Services”, “A.W.S.” etc. all to Cloud infrastructure. Tedious to build, fragile to maintain, doesn't generalise.
  3. Enterprise platforms. Large spend analytics suites do an excellent job for big corporates but require a multi-month implementation, a procurement analytics function, and enterprise budget.
  4. Consultants. Boutique firms do it as a service. Expensive, slow, and the moment your data changes shape you're back to square one.

What changed in 2025-2026

Large language models got cheap enough and good enough to read a normalised supplier description and pick a leaf category from a candidate list with high accuracy. That sounds simple but it shifted the economics: classification that used to require a rules engine and weeks of tuning can now be done in minutes by a model that costs a few cents per thousand rows.

The catch: LLM-only classification still hallucinates, drifts on edge cases, and gives different answers on different runs. The fix is the same fix that worked before LLMs: anchor the model to a fixed top-level taxonomy, validate on a sample before committing, and let humans override anything that looks wrong. Cache those overrides so the next run learns from them.

What good looks like

If you're evaluating classification (manual, rule-based, or AI-driven), the things to look for are:

  • Standards alignment. The leaf categories should map to a known standard like UNSPSC so your numbers are comparable with industry benchmarks and with future ERP migrations.
  • Sample preview. You should see classification quality on a real slice of your data before committing budget.
  • Human-in-the-loop. Every classification system gets some rows wrong. The system has to make corrections cheap (one click, with the correction sticking for that supplier forever).
  • Editable taxonomy. Your business has nuance. Every level of the taxonomy should be renameable / re-parentable / deletable without breaking everything downstream.
  • Clean export. The output should be an Excel (or CSV / Parquet) file with one row per transaction and the assigned Level 1 / Level 2 / Level 3, plus the original raw columns. No proprietary format lock-in.

How Structera does it

Disclosure: I work on Structera, so this section is biased. Our approach: fix the top of the taxonomy to 18 universal procurement towers anchored to the four global standards (UNSPSC,SIC, NAICS, NACE), generate Level 2 and Level 3 per-dataset based on what's actually in your data, use a frontier AI engine to classify a statistically representative sample for free, let you review and override, then unlock the full dataset for a pay-as-you-go fee. Overrides are stored per-supplier so the next run never re-asks the LLM about the same vendor. Pricing starts at GBP 25 minimum, tiered down to GBP 10 per 1,000 rows at scale.

If you want to try it on your own data, create a free account — sample classification is free, no card required.

Further reading

← All posts