Every spend report, savings target and category strategy depends on one unglamorous step: deciding which category each line of spend belongs to. A purchase order line, an invoice and a card transaction record what was paid and to whom. None of them reliably records what was bought in a form a category manager can use. Spend classification fills that gap, and its quality caps the value of everything built on top of it.
What spend classification is
Spend classification maps each transaction to a node in a category taxonomy, so analysts can total spend by category as well as by supplier and business unit. Tools bundle it with data preparation: Gartner's Peer Insights market page lists data cleansing, normalization and enrichment as mandatory features of spend analytics solutions, alongside reporting and visualization (Gartner Peer Insights).
Classification also depends on supplier normalization, which rolls variant spellings of one company and its subsidiaries up to a single parent. Without that step, category totals split across duplicate supplier records.
Why the source data resists it
Finance systems record accounting treatment, not purchasing intent. A general ledger code says an expense was "professional fees" or "repairs and maintenance", which is too coarse to source against. Supplier names are a weak signal too, because an industrial distributor or an online marketplace can sell many unrelated categories under one vendor ID.
Public sector data shows the same limit. In a 2012 review of the US Department of Defense's inventory of contracted services, GAO reported that the Federal Procurement Data System-Next Generation did not identify more than one type of service per contract action (GAO, 2012). Where a contract covered several services, users entered the code representing most of its value (Nextgov, 2012). Any design that stores one code per record loses detail the same way.
Standard and custom taxonomies
Organizations choose between a public standard and a taxonomy of their own. The United Nations Standard Products and Services Code (UNSPSC) is a four-level hierarchy of segment, family, class and commodity, in which each level adds two digits to form an eight-digit code (UN Global Marketplace; O*NET Resource Center).
Public buyers in the EU use the Common Procurement Vocabulary (CPV), set out in Regulation (EC) No 213/2008. Its main vocabulary uses an eight-digit code plus a check digit, with divisions, groups, classes and categories identified by successive digits (EUR-Lex, 2008; Malta Department of Contracts).
Standard codes make benchmarking easier and let suppliers send pre-coded catalog data. A custom taxonomy can instead mirror how the category team is organized, and Gartner's market page lists custom taxonomies as one form of spend data enrichment. Either way, depth matters. A tree that stops at "electrical components" cannot separate cable from switchgear.
How the main approaches differ
Rules-based classification assigns a category from the supplier, the GL code or keywords in the line description. It is transparent and easy to correct, but it fails on multi-category suppliers and on any supplier the rules have not seen.
Machine-learning classifiers predict a category from the text of the line itself. JAGGAER announced a neural-network classifier in 2019 that takes a product description or product number from an invoice and predicts an eCl@ss code, with UNSPSC as an alternative (JAGGAER, 2019; CPO Rising, 2019). Models cope better with unfamiliar suppliers than rules do, but their mistakes are harder to trace.
Hybrid designs pair the two. Rules handle stable, high-volume suppliers and a model takes the rest, with low-confidence lines sent to an analyst for review.
What to check when buying
Accuracy figures need a definition. A vendor can measure accuracy by line count or by spend value, and the two can diverge: a few correctly classified large invoices can cover most of the spend while many small lines stay wrong. Coverage, the share of spend that carries any category, is a separate measure and says nothing about whether the category is right.
The test method matters as much as the score. Purchase data is repetitive, with the same descriptions recurring across months and sites. If a vendor scores a model on lines whose text also appeared in its training data, the result partly measures memory rather than classification. A September 2026 arXiv preprint tested this on 1.26 million labelled purchase lines from two US state governments. With random splits, about 61% of test rows had matching item text in the training data. On California orders, the best classical baseline scored 56.0% on repeated text and 31.9% on text it had not seen (arXiv, 2026). It has not been peer reviewed. The practical request is a trial on the buyer's own data, including suppliers and descriptions the model has not seen.
Buyers should also ask whether an analyst's correction survives the next data refresh and feeds back into the rules or model. If not, the same errors come back on every reload. When a category is split or merged, the tool should reclassify historical spend so trend lines stay comparable.
What commonly goes wrong
One failure is treating classification as a one-off project. New suppliers, new cost centers and reorganized categories erode a clean dataset unless someone owns it after go-live. Another is classifying at supplier level only, which files every purchase from a broad distributor under a single category.
Taxonomies can also outrun the data that feeds them. A commodity-level tree is of little use when the line description reads "misc services" and nothing else, because no rule or model can recover a category the requisitioner never recorded.
Sources
- Gartner. "Spend Analytics Solutions Reviews and Ratings" (Peer Insights market page, mandatory features updated September 2025). Accessed October 2026. https://www.gartner.com/reviews/market/spend-analytics-solutions
- U.S. Government Accountability Office. "Defense Acquisitions: Further Actions Needed to Improve Accountability for DOD's Inventory of Contracted Services" (GAO-12-357). 2012. https://www.gao.gov/products/GAO-12-357
- Nextgov. "GAO sees progress in DOD contract inventory." 2012. https://nextgov.com/acquisition/2012/04/gao-sees-progress-in-dod-contract-inventory/206392
- UN Global Marketplace Help Centre. UNSPSC code help article. Accessed October 2026. https://help.ungm.org/hc/en-us/articles/360012816160
- O*NET Resource Center. "UNSPSC Reference." Accessed October 2026. https://www.onetcenter.org/dictionary/24.3/text/unspsc_reference.html
- European Union. "Commission Regulation (EC) No 213/2008 on the Common Procurement Vocabulary (CPV)." 2008. https://eur-lex.europa.eu/eli/reg/2008/213/oj/eng
- Malta Department of Contracts. "Common Procurement Vocabulary (CPV)." Accessed October 2026. https://contracts.gov.mt/en/resources/cpvs/
- JAGGAER. "Automatic Spend Classification Helps Overcome the Challenge of Unreliable Spend Data." 2019. Vendor press release. https://www.jaggaer.com/press-release/automatic-spend-classification-challenge-unreliable-spend-data/
- CPO Rising. "Technology Round-Up, December 11, 2019." 2019. https://cporising.com/2019/12/11/technology-round-up-december-11-2019/
- arXiv. "Spend Classification Without Leakage: An Evaluation Harness and What It Changed in a Deployed System." Harshit Gupta. 2026. Preprint, not peer reviewed. https://arxiv.org/abs/2609.25502