Quick answer: UNSPSC is the United Nations Standard Products and Services Code, a four-level classification system used to structure procurement data. Construction firms use it to see spend by category, supplier and project rather than by free-text description. Getting it right depends on three things: one owner for the decisions, written rules for the edge cases, and a review loop that measures whether the labels stay consistent over time.
On this page: UNSPSC explained in four levels · Who owns UNSPSC classification decisions · How to handle UNSPSC edge cases · Using AI for UNSPSC classification · Frequently asked questions
Most construction firms I speak to face a version of the same problem. They have a high-level idea of what they are buying. They want to make better decisions with that information. And when they go looking, the data is technically there, but it is not organised in a way that shows the consequence of any decision they might make.
Decision making improves when you can predict the outcome. If we invest, we assess how the company or the stock is likely to perform first. If we plan a route to the airport, we take possible delays into account. When we negotiate with a supplier, we prepare as well as we can.
Procurement in construction usually runs on experience instead. Finance has a partial view of where the money goes. The reason is simple: every project carries its own details, so it is impossible to know months or years ahead exactly what goods and services will be needed to deliver it.
That does not mean nothing is predictable. Individual projects are unique. Buying patterns are not. Once every line is classified in detail, trends appear immediately: spend per product over time, price movement per supplier, and the anomalies. Eighteen suppliers for windowsills. A project manager who consistently buys outside framework contracts. None of that is visible in a free-text description field.
Classifying spend properly introduces three new questions. Which categories to use. Who is responsible for making sure every line carries the right one. And how to decide when the answer is genuinely ambiguous. The rest of this article works through each.
UNSPSC explained in four levels
UNSPSC stands for United Nations Standard Products and Services Code. It was designed by the UN as a single procurement standard, and it has two properties that matter here.
The first is that it is universal. It is used globally and across industries, so some of your suppliers and clients are already using it. That makes external comparison possible in a way an internal taxonomy never allows.
The second is depth. UNSPSC runs to four levels, from a broad segment down to an individual commodity: construction materials, then structural elements, then aluminium I profiles. That range lets you ask both kinds of question from the same dataset. How has subcontractor spend moved over the last year, at the top. How do our concrete suppliers compare on price, volume, delivery and payment terms, at the bottom.
The size of the standard puts people off. There are over 150,000 commodities in it. In practice you write off most of it on day one. If you are an HVAC contractor, agriculture is not happening. The working set is usually a handful of segments and a long tail that appears a few times a year.
One decision worth making early: how deep you actually need to go. Depth costs money and review time. Classify to family or class level for reporting, and go to commodity level only in the categories you intend to negotiate.
Who owns UNSPSC classification decisions
One person should own it.
Splitting the ownership creates a specific failure. Any line with ambiguity in it gets resolved differently by different people, and the inconsistency is invisible until someone tries to run a report and the numbers do not reconcile. At that point you cannot tell whether spend moved or whether two buyers disagreed.
AI can do the classification work, and the section below covers how. But the ownership question does not disappear when you automate. It changes shape. Someone still has to decide what the business means by a category and approve the lines the system is unsure about.
The most reliable way to run this is to treat the AI as a junior team member. It has a name, it does the volume work, it produces output that a person reviews, and the feedback goes back in. That framing gets the process right more often than treating it as a tool that either works or does not.
There is a second reason to name an owner, and it applies to any firm heading into an ERP migration. If nobody owns the classification, the migration inherits the mess. If someone does, new purchase orders start being coded correctly at the point of entry, and by the time the new system goes live there are already a couple of years of clean history behind it.
How to handle UNSPSC edge cases
Two rules. Make the business goal the deciding factor. Then write the rule down and apply it consistently.
The most common blocker is that a team is moving to UNSPSC without anyone in-house who knows UNSPSC. That is not a skills gap worth hiring for, and it is not a reason to delay. It is a workshop.
Put someone with real buying knowledge in a room with someone who knows the codes. The buyer explains, in plain language, what matters to the business and what the categories are meant to support. The UNSPSC side surfaces the edge cases hiding inside that. Then the team decides on every edge case in one sitting, and each decision is written down.
Compare that with the alternative. When nobody knows the standard, the team researches edge cases indefinitely, working out which codes could possibly attach to a product. Best case, someone documents the outcome and it becomes a rule everyone follows. Worst case, every team member does their own research, which slows the process further and produces exactly the inconsistency the exercise was meant to remove.
The edge cases are usually boring and business-specific. Does a line reading "hours" mean subcontracted labour or plant hire. Does a damper get classified as a component or as part of an assembly. Whether the answer is right in the abstract matters far less than whether it is the same answer every time.
Two categories that get skipped and should not: line items that are not products at all, such as transport charges, fuel surcharges and late payment fees, and credit notes. Leaving them unclassified quietly distorts every category total that sits above them.
Using AI for UNSPSC classification
We have tried and tested most of the ways this goes wrong, so you do not need to.
Load the taxonomy properly first. Language models invent categories. A code that does not exist in the standard defeats the entire purpose, which is repeatability. The taxonomy has to be loaded as a constrained reference the model selects from, not as context it draws on.
Balance the three signals. A line description alone is not enough. "Gloves" can be medical gloves, cleaning gloves, work gloves or winter gloves. The supplier and the nature of the business resolve it: an electrical and plumbing wholesaler is not selling surgical gloves. Weighting supplier identity against description against your own procurement patterns is where most of the accuracy comes from.
Encode the company-specific rules. Every business has categories the standard does not anticipate cleanly, and the rules from the edge-case workshop belong in the pipeline rather than in a document nobody opens.
Score confidence and route the weak lines to a human. A general chatbot gives you an answer with no indication of how reliable it is. A line with nothing but a part number in the description should be flagged automatically, not guessed at. Corrections then feed back, so the same ambiguous line is not decided twice.
Pearstop runs this as a managed pipeline: taxonomy constrained, supplier context weighted, company rules applied, low-confidence lines routed to review, and corrections learned from.
The practical reason to use a pipeline rather than a chatbot is cost, and it is not the cost people expect. Teams who try this internally usually arrive having already spent a meaningful amount on tokens and several weeks of a buyer's time, with output they do not trust. The spend was not wasted on the classification. It was spent discovering the guardrails.
The other comparison worth making is against hiring. Classifying a large backlog by hand is a full-time analyst role, and it never ends, because new lines arrive every month. The question is not whether a person or a machine is more accurate on a single line. It is whether you want a buyer spending their week on data coding or on the negotiations the data is supposed to inform.
Frequently asked questions
What does UNSPSC stand for?
UNSPSC stands for United Nations Standard Products and Services Code. It is a classification system designed as a single standard for procurement worldwide. It covers goods and services across every industry, including construction, manufacturing and facilities management, and organises them into a four-level hierarchy running from a broad segment down to an individual commodity.
How many levels does UNSPSC have?
UNSPSC has four levels: segment, family, class and commodity. Segment is the broadest grouping, such as structures and building components. Commodity is the individual item, such as an aluminium I profile. Most organisations classify to family or class level for reporting and only go to commodity level where a specific category is being actively negotiated.
Is UNSPSC suitable for construction procurement?
Yes. UNSPSC expands into construction-specific branches covering structural elements, building materials, HVAC equipment and subcontracted services. Because the standard is used across industries and borders, suppliers and clients often already hold codes against their own catalogues. That makes it easier to compare spend externally than a taxonomy built in-house for one company.
Can AI classify purchase orders to UNSPSC codes accurately?
It can, but not with a general-purpose chatbot alone. Language models invent codes that do not exist in the taxonomy, which destroys the repeatability the exercise depends on. Accurate classification requires the taxonomy loaded as a constrained reference, supplier context weighted against the line description, company-specific rules, and a review step for low-confidence lines.
How long does it take to classify a backlog of purchase orders?
A pilot on roughly one year of data takes about four weeks, which allows time to build the structures, review initial findings and adjust the rules. A full historical backlog runs after the pilot, and the timeline depends on how much custom configuration the data needs rather than on the number of lines involved.
How does Pearstop handle UNSPSC classification?
Pearstop loads the UNSPSC taxonomy as a constrained reference so codes cannot be invented, weights supplier context against the line description, applies company-specific rules agreed with the procurement team, and routes low-confidence lines to human review. Corrections feed back into the pipeline so the same ambiguous line does not need deciding twice.
Free resources
Not sure which UNSPSC code to use?
Paste any product or service description and get the correct 8-digit code instantly — or explore the full taxonomy tree to understand the hierarchy.
Stephanie Wiechers
CEO & Co-founder, Pearstop
Stephanie leads Pearstop's go-to-market and strategic direction. She works directly with procurement and FM leaders across Europe to understand how data quality affects margins, contracts, and AI readiness.
LinkedIn →Further reading
Why Construction Procurement Data Breaks Every Time You Try to Benchmark It
Every project prices materials and subcontractors differently. Benchmarking construction spend across projects requires classification most cost systems were never built to do.
Read more →ConstructionConstruction Procurement Data: Why Your Material Costs Are Harder to Control Than You Think
Construction spend is structurally fragmented across projects. This article explains why material cost control requires spend classification and what steel procurement looks like as a case study.
Read more →