The EU AI Act classification framework for workforce AI

How to use this framework

If you own AI governance for a workforce, your deliverable this year is not a point of view on the AI Act. It is a stack of classification memos: one per system that screens applicants, talks to employees, summarises listening data, or feeds anything that looks like a people decision. Download the template above, then use the sections below — each one annotates the matching section of the memo — to fill it in for your first system.

Two dates drive the exercise, and they are fixed. From 2 August 2026, Article 50 requires any AI system that interacts directly with people to say so, with the fining powers live the same day. From 2 December 2027, the Annex III high-risk regime applies to standalone employment AI — the Digital Omnibus replaced the original August 2026 date with fixed calendar dates, and the remaining flexibility runs earlier, never later. We've written up both deadlines in detail: the Article 50 disclosure duty and what December 2027 actually requires.

The framework is deliberately product-agnostic. It works on a chatbot in a benefits portal, an engagement survey platform, a video-interview screener, or our own product — we run it on Tadeus in public at How Tadeus classifies itself.

First, describe the system honestly

Classification runs on use, not on product category — so the memo starts by pinning down what the system is actually deployed to do, not what its marketing says. Three fields in section 1 do most of the work.

Your role under the Act. Providers build or substantially modify AI systems; deployers use them under their own authority. The split matters commercially: embedding a vendor's system transfers the provider obligations (risk management, documentation, logging design, conformity assessment) to the vendor, while the deployer obligations — using the system per its instructions, assigning competent oversight, retaining logs, informing workers and their representatives — stay with you no matter what you buy.

Deployment contexts. List every live or planned use separately. The same listening tool can be out of scope in a comprehension-check deployment and squarely high-risk in a performance-signals deployment. Each context gets its own pass through sections 2–4; if that produces different answers, the memo should say so rather than average them.

The Article 50 test: does it interact with people?

Article 50 is the near deadline and the easy test. If the system is built to interact directly with natural persons — a chatbot, a voice agent, a conversational survey — the person must be told they are dealing with an AI, before or at the first interaction, in a clear and distinguishable form. This applies to all AI systems in scope, not just high-risk ones, and it applies from 2 August 2026.

Two things trip teams up. First, the "obvious from context" exception is judged from the perspective of the average person actually exposed — and a voice that greets an employee by name, in their own language, with human pacing, is engineered specifically not to be obvious. The more human the system sounds, the more clearly it must say it is not. Second, the obligation is satisfied by evidence, not configuration: when a works council or regulator asks, "we enabled the disclosure setting" is not an answer. A timestamped, per-interaction record of the disclosure is.

So the memo asks the vendor question worth asking in writing: for a named person on a named date, can you produce a record showing they were told, and in what language? The full argument is in our Article 50 piece.

The Annex III point 4 test: what do its outputs touch?

This is the classification question proper. Annex III point 4 makes AI high-risk when it is used for any of six employment functions. Answer each limb per deployment context — with evidence of where the output actually flows, not an org-chart opinion.

The pivotal field is the aggregate-versus-individual one. A listening programme that is genuinely descriptive — aggregate outputs that never inform an outcome for a named person — likely sits outside the regime. The moment its output feeds who gets promoted, who gets managed out, or how work is distributed, the deployment is in. The system does not decide its own classification; its deployment does.

The derogation, and what claiming it costs

Article 6 contains a derogation for systems that fall within an Annex III category but do not pose a significant risk of harm: a narrow procedural task, improving the result of a previously completed human activity, detecting decision patterns without replacing or influencing human assessment, or a purely preparatory task. It is not available at all where the system profiles natural persons.

The temptation is to file everything under it — and the final Digital Omnibus text quietly closed the comfortable version of that move. Providers must document the assessment and register the system in the EU database even where they consider it exempt. Self-classification is now a documented, reviewable claim, not a private opinion. Claiming the derogation is sometimes right; claiming it casually converts a classification question into a filed position a regulator can mark wrong.

So the memo makes the cost explicit: name the limb, clear the profiling bar, and record where the assessment is filed and when the registration will be made. If a vendor tells you their employment-adjacent system is exempt, ask to see that filing.

Use-drift: the answer changes when the use changes

Most misclassification won't come from bad legal analysis. It will come from drift: a listening tool deployed for communication in January feeding a performance dashboard by June, with nobody re-reading the paperwork. The classification follows the drift whether or not the memo does.

Section 5 is therefore the most operational part of the memo: name the specific changes that would flip the answers above, and assign someone to watch for them. Typical triggers for a workforce tool: outputs routed into performance reviews; per-individual filtering added to a dashboard; integration into task allocation; export of individual-level results to line managers; a new deployment context that never went through this memo.

Conclude, sign, and calendar the review

The conclusion is a single line plus its consequences. Whatever the classification, write down the obligations that follow and their dates — that list is what turns the memo into a budget line.

Then sign it and calendar the re-run: when the Commission's final classification guidelines land (expected by the end of 2026), and quarterly regardless. A memo with a named owner and a review date is governance; a memo without them is a blog post. For the full translation of the Articles 9–15 obligations into what a conversational system actually has to produce, see the December 2027 piece.

Versioning

This framework and the memo template are maintained documents, not campaign content. Each version is checked against the Commission guidance current at the time — v1.0 against the draft classification guidelines of 19 May 2026 — and re-checked when the finals land. The change log below records every revision. If you'd rather not check back, the notification form at the end of this page emails you when the document changes, and nothing else.

Change log

• 10 July 2026 — v1.0
  First public version.