Uniclass in 2026: five shifts worth tracking

26 July 2026 · Carlos Greblo

Classification used to be the part of a project that someone did once, badly, near the end. That is no longer viable. Uniclass changes four times a year, the standard that gives it its purpose is mid-rewrite, and the tooling around it has quietly moved from spreadsheets to APIs and models.

Here are five shifts we think matter, and what each one changes in practice.

Shift 01

Uniclass is a moving dataset, not a fixed reference

NBS publishes Uniclass on a quarterly cycle, and 2026 has kept that pace. The January release touched eight tables. The April release touched six (Activities, Form of information, Project management, Products, Spaces/locations and Systems), with a modest 12 new classifications and 40 amended codes.

Small numbers, but they compound. If your project classified against a 2024 export, you are three to eight releases behind, and the gap shows up as codes that no longer exist, titles that no longer match what your team searches for, and two disciplines classifying the same product two different ways because they downloaded the tables in different quarters.

What to do about it

Treat the Uniclass release like any other dependency. Pin the version you classified against, record it, and schedule a re-check when you upgrade. A classification with no version stamp is not traceable, it is just a string.

Shift 02

ISO 19650 is dropping "BIM" for "information management"

A Draft International Standard for ISO 19650 Parts 1, 2 and 3 went out for public consultation on 10 March 2026, with the final revision expected in 2027. Two proposed changes stand out. The language shifts away from "BIM" toward "information management" throughout, and the delivery phase (Part 2) and operational phase (Part 3) merge into a single continuous nine-step process across the whole asset lifecycle.

Read that second change carefully, because it is the one with teeth. If delivery and operations are one process, then the classification applied during design is the same classification the asset owner inherits at handover. There is no convenient boundary at practical completion where someone re-tags everything.

What to do about it

Classify at the point where information is created, which usually means at the requirement, not at the model element weeks later. The requirement is the earliest artefact that knows what the thing is supposed to be.

Shift 03

Machine-readable access is now the default expectation

Uniclass ships as an Excel and CSV download, but it also has a REST API that lets applications navigate and retrieve classifications across every table directly, and NBS has been actively pushing tooling to lift adoption.

The practical consequence is that private copies of the tables are becoming a liability. Every locally maintained spreadsheet of "our Uniclass codes" is a fork that drifts a little further each quarter and eventually has to be reconciled by hand.

What to do about it

Consume the tables programmatically, from one source, and let your tools do the lookups. If a person is typing Pr_25_71_50 from memory, the process has already failed.

Shift 04

Infrastructure and asset owners are steering the vocabulary

Uniclass started as a UK building-sector scheme and is increasingly neither. The January 2026 update added surface water management codes across the Products, Systems and Spaces tables, driven by the Environment Agency's use of Uniclass for asset classification. That is an operator, not a designer, shaping the scheme around the assets it has to maintain for decades.

The same pattern shows up geographically. In Australia, NATSPEC has endorsed Uniclass in its comparison of international classification systems, and Transport for NSW publishes its own application guide for using Uniclass on transport projects.

What to do about it

If you work in infrastructure, water, or transport, stop assuming Uniclass has nothing for you. And if the codes you need genuinely do not exist, submit them, because that is visibly how the recent additions got there.

Shift 05

Automated classification crossed the credibility line

For years, automated Uniclass assignment was a demo. It is now a research result with numbers attached: a 2025 study in Buildings classified BIM objects by extracting semantic, spatial and dimensional features from IFC data and reported accuracies of 1.00 and 0.99 on the Elements/Functions and Systems coding tasks respectively. Separately, embedding-based search has made it practical to rank candidate codes against free text rather than requiring an exact keyword hit, which is what most people actually needed from a Uniclass search box.

The caveat matters as much as the result. Those accuracies come from structured IFC input on a curated object set. Free text written by an engineer at 4pm on a Friday is a harder problem, and the honest output of a classifier on that input is a ranked list with confidence attached, not a single answer.

What to do about it

Adopt suggestion workflows, not replacement workflows. Auto-apply the high-confidence tail, where the model is effectively doing data entry, and put human attention on the genuinely ambiguous middle. The value is not that the machine is always right, it is that it removes the 80% of the work that was never a judgement call.

The through-line

All five shifts point the same way. Classification is becoming continuous (updated quarterly), earlier (at the requirement, not the model), programmatic (API, not spreadsheet), broader (infrastructure and operators, not just UK buildings), and assisted (ranked suggestions, not manual lookup).

That is roughly the shape of how ReqTwin handles it. Every Uniclass Product (Pr) and System (Ss) code, 11,174 of them, is vector-embedded, so any requirement or bill-of-materials item can be ranked against the full catalogue by meaning rather than keyword. Each row gets a Product and a System suggestion with a confidence percentage, the two tracked independently because an item can legitimately have one and not the other. A confidence threshold auto-applies the obvious matches in bulk, and the rest stay in front of a person, project-wide, in one table.

Classification stops being an end-of-project chore when it takes a click instead of an afternoon.

Sources