How Tadeus classifies itself under the EU AI Act

The classification, at a glance

Tadeus is a voice layer for workforce platforms: one AI agent holds real voice conversations with employees and returns structured, timestamped records to the systems the organisation already uses. Classification runs on use, so we classify each supported deployment separately.

The reasoning, row by row

Article 50: every deployment is in scope, and that's by design

Tadeus talks to people; there is no reading of Article 50 under which a voice agent holding a natural conversation is "obvious from context." So we don't argue the exception. Every conversation opens with the agent stating it is an AI, what the conversation is for, and what happens to the answers, and that disclosure is captured in the same structured, timestamped session record as everything else, per person, per date, per language. Disclosure increases candour rather than suppressing it, which makes this a rare obligation that pays for itself: the evidence is here.

Annex III: intended use keeps outputs aggregate and away from individual outcomes

Three architectural facts do the work. Tadeus has no individual scoring surface: there is no screen that ranks or rates a named employee. Listening campaigns return aggregate-only outputs: themes, distributions, and quality-weighted signal, not per-person verdicts. And Tadeus does not ingest performance data: there is nothing in the system to correlate a person's answers against their appraisal. A system deployed this way informs decisions about the organisation, not about named individuals, which is what keeps the intended deployments outside Annex III point 4.

The corollary is stated plainly in the callout above: those are properties of the intended deployment, not a legal force field. Session records are exportable by design (that is what makes the evidence obligations dischargeable), and an organisation that exports individual records into a performance process has changed the classification of its own deployment. We say so in our Instructions for Use, in onboarding, and here.

Article 5: the prohibitions, screened first

The classifier's v3 update screens the Article 5 prohibited practices before anything else, so this classification now does too. The prohibition that matters for a workplace voice system is Article 5(1)(f): inferring emotions in the workplace, banned since 2 February 2025. Tadeus does not do it, by design rather than by policy: the system assesses comprehension and content, not emotional state, there is no emotion-inference surface in the product or the API, and our Instructions for Use list emotion inference as a must-not-be-used-for. The other prohibitions (biometric categorisation by sensitive attributes, offending prediction, social scoring, untargeted face scraping) describe functions Tadeus does not have. No prohibited practice is engaged.

Article 50(2): the agent's own synthetic speech

The agent's voice is generated audio, so we state a position on the Article 50(2) marking duty rather than leaving it unaddressed. The operative duty for a live, disclosed conversation is Article 50(1): every session opens with the agent saying it is an AI, and that disclosure is evidenced per session. Tadeus retains no raw audio by design, so no synthetic audio artefact persists to circulate unmarked; the structured session record identifies every agent utterance as machine-generated, which is the machine-readable provenance the duty is after. If we ever ship a feature that exports or publishes generated audio, that feature carries 50(2) marking before it ships, and this section changes first. The Commission's technical guidance on marking formats is still settling; we will align the record format with it when it lands.

The other seven Annex III categories: screened, none engaged

The July 2026 update to our classifier broadened the screen from the employment category to all eight Annex III domains, so this classification was re-run under the full framework. None of the other seven applies to a supported deployment. Two deserve a sentence rather than a footnote. Biometrics (point 1) is the category a voice system invites questions about: Tadeus does not identify people from their voice, does not categorise them by inferred traits, and does not read emotional state, which keeps it clear of point 1(c) as well as the Article 5 workplace prohibition. Education (point 3) covers admissions and assessment within education and vocational training institutions: workplace policy briefings and comprehension checks are not that, but an education provider using Tadeus to evaluate learning outcomes should classify that deployment on its own facts, and the classifier now covers the category.

Derogation: not claimed, deliberately

We do not claim an Article 6 derogation (an exemption that lets a system otherwise caught by Annex III skip the high-risk regime) for any supported use case. We could construct an argument, but we don't need one while intended deployments sit outside Annex III, and claiming one casually would convert a clear position into a filed claim to defend. If the Commission's final guidelines move the boundary and any supported use case lands in an Annex III category, our position is to either meet the full high-risk obligations for it or claim the derogation properly: documented assessment, EU database registration included. What we will not do is describe ourselves as "AI Act certified": no vendor honestly can while the harmonised standards are still being finalised.

Registration posture

As of this version: no Tadeus use case is classified high-risk under any Annex III category, no Article 6 derogation is claimed, and accordingly no EU database registration is currently required or filed. This position is re-run against every framework update (including the Commission's final classification guidelines, expected by the end of 2026), and any change lands here and in the change log before it lands anywhere else.

The drift triggers we tell customers to watch

Classification drift is the main real-world risk for a listening tool, so these are the specific changes we tell deployers to treat as a re-classification event, not a configuration change:

• Exporting individual-level session records into a performance review, PIP, or promotion process.
• Adding per-individual filtering to any dashboard consumed by line managers.
• Pulling session exports across campaigns and matching them on name or email to reconstruct a person's sentiment, engagement, or transcript history yourself. Tadeus renders no screen that does this, but the identifying fields and the per-session signal sit in the same export, so this takes a spreadsheet and a few minutes, not a deliberate build. It is the most likely way drift actually happens, and it counts as reclassification whether or not it was intentional.
• Feeding conversation outputs into task-allocation or scheduling logic.
• Using comprehension-check results to trigger consequences for a named person.
• Deploying any conversation whose stated purpose to employees differs from the use its outputs are put to.

Any of these puts the deployment in Annex III point 4, and the deployer obligations under Article 26 (competent human oversight, log retention, informing workers and their representatives before use) attach to it. If you are planning one of these uses, run the classification memo for it first and talk to us: high-risk deployments need the full evidence trail, and we would rather help you build it than discover the drift later.

Review cadence and ownership

This classification is reviewed quarterly and at every regulatory change, and is owned by our Founder. The version and last-reviewed date at the top of this page are updated on every review, even when nothing changes. A review that leaves the classification untouched is still a review, and the change log records it.

Change log

• 15 July 2026 · v1.2
  Re-run under classifier v3.0. Added the Article 5 prohibited-practices screen (none engaged; workplace emotion inference is excluded by design) and a stated position on Article 50(2) marking for the agent's own synthetic speech: 50(1) disclosure is the operative duty for a live conversation, no raw audio is retained, and session records identify agent speech as machine-generated.
• 15 July 2026 · v1.1
  Re-run under classifier v2.1, which screens all eight Annex III categories instead of employment alone. The classification is unchanged: intended deployments remain outside Annex III and Article 50 stays in scope by design. Added the screening result for the other seven categories, with biometrics and education addressed explicitly, and broadened the registration posture from point 4 to any Annex III category.
• 10 July 2026 · v1.0
  First public classification. Reviewed quarterly.