# EU AI Act Classifier for Workforce AI Systems

- Version: 1.0
- Last reviewed: 2026-07-10
- Owner: Founder
- Checked against: the European Commission's draft classification guidelines of 19 May 2026
- Status: a maintained compliance document, not legal advice — take the reasoning to your counsel
- Canonical: https://tadeus.net/trust/ai-act-classifier

The interactive classifier at /trust/ai-act-classifier walks a workforce AI system through the classification framework in eight questions and outputs a likely classification, the obligations and dates that follow, and a pre-filled memo. This document records the questions and the rules, so the tool's reasoning is inspectable.

## The questions

One pass per system. Questions 4–6 appear only for systems used in the employment context; question 7 appears only when an answer has placed the system in an Annex III category.

| # | Question | Statutory hook |
| --- | --- | --- |
| 1 | What is your relationship to this system? | Article 3 |
| 2 | Does it interact directly with people? | Article 50(1) |
| 3 | Who is it used on or about? | Annex III point 4 |
| 4 | Do its outputs inform recruitment — advertising, screening, filtering, or evaluating candidates? | Annex III 4(a) |
| 5 | Do its outputs inform decisions about named individuals — promotion, termination, task allocation, monitoring, or performance evaluation? | Annex III 4(b) |
| 6 | What do its outputs concern? | Annex III point 4 |
| 7 | Does an Article 6 derogation limb honestly describe the system's role? | Article 6(3) |
| 8 | Could outputs plausibly drift into individual performance or task decisions later? | Use-drift |

## The classification rules

1. A system used in the employment context whose outputs inform recruitment or selection (including planned use), or decisions about named individuals (including advisory outputs that reach decision-makers), or whose individual-level outputs are routinely visible to managers, is treated as within **Annex III point 4** → likely high-risk, with the December 2027 obligations.
2. Within Annex III, if an **Article 6 derogation** limb honestly fits and the system does not profile people, the output is "derogation claimed" — which still requires a documented assessment and EU database registration.
3. Outside Annex III, a system that interacts directly with people carries the **Article 50** transparency duty from 2 August 2026.
4. Everything else is out of scope as described — the output is then a recorded reasoning plus the drift triggers that would change it.

The rules lean conservative on purpose: ambiguous answers ("advisory outputs reach managers", "recruitment use is planned") classify as if the riskier reading holds, because the memo's job is to survive scrutiny, not to win an argument with it.

> **Not legal advice**
>
> The output is a likely classification and a pre-filled memo. Take this reasoning to your counsel — the line is printed inside every document the tool produces.

## What the tool produces

- A likely classification per system, with the reasoning line per answer.
- The obligations that follow, with the fixed dates (2 August 2026 for Article 50; 2 December 2027 for the Annex III regime).
- A downloadable, editable memo (Markdown or Word) pre-filled with your answers — one classification memo for a single system, a mini AI-system register for several.

The memo skeleton is the same [classification memo template](/trust/classification-framework) the framework annotates; answers land in the matching sections.

## Change log

- 2026-07-10 — v1.0: First public version, mirroring framework v1.0.