Tadeus Instructions for Use

Why this document exists now

Article 13 requires providers of high-risk systems to ship instructions for use that state what the system does, what it does not do, and what it must not be used for. Tadeus's intended deployments are not classified high-risk, so this document is not yet an obligation for us — but it is quietly the most commercial of the Act's requirements: the must-not-be-used-for list is both a legal shield and the restraint position published as product documentation. Publishing it early costs nothing except the ability to quietly change our minds, which is the point.

This document follows the shape Article 13 prescribes, versioned and change-logged like the legal document it will eventually be. For what the December 2027 regime demands in full, see the second deadline.

Intended purpose

Tadeus is a conversational voice layer for workforce platforms. Its intended purpose is to hold disclosed, structured voice conversations with employees and research participants — listening campaigns, policy briefings and comprehension checks, onboarding conversations, implementation and requirements interviews, and market or UX research — and to return structured, timestamped session records and aggregate analysis to the systems the deploying organisation already uses.

The intended unit of insight is the organisation, the team, the theme, and the rollout — not the ranked individual. Deployments consistent with this purpose sit outside Annex III point 4 of the EU AI Act; deployments that depart from it can change their own classification, as set out below.

What the system does

• Discloses itself. Every conversation opens with the agent stating it is an AI, what the conversation is for, and what happens to the answers — before any question is asked, in the participant's language.
• Holds structured voice conversations against a defined question library, in multiple languages, on ordinary phones and browsers.
• Returns a per-session record: who was asked what, when, in which language, how responses were interpreted, and which model and instruction version ran the session.
• Attaches a quality signal to responses, separating considered answers from completion-mode filler, so downstream data governance can tell signal from noise.
• Aggregates for analysis: themes, distributions, and comparisons at organisation, site, and team level.
• Exports its evidence. Session records, disclosure events, and analysis are exportable by the deployer — the properties an Article 12-style evidence trail needs are by-products of normal operation.
• Supports human oversight: named reviewers can inspect, override, and disregard outputs, and those decisions are recorded with the reviewer and timestamp.

What it does not do, by design

These are architectural facts about the product, verifiable in the interface and the API today — not policy promises.

• No individual scoring surface. There is no screen, field, or API response that rates, ranks, or scores a named employee.
• Aggregate-only outputs on listening campaigns. Campaign results are themes and distributions; individual listening responses are not presented as per-person verdicts.
• No performance data ingestion. Tadeus does not import appraisals, ratings, or productivity metrics, so there is nothing to correlate a person's answers against.
• No emotion recognition. The system assesses comprehension and content, not emotional state — workplace emotion inference is a prohibited practice under Article 5, and we do not build toward the line.
• No raw audio retained, no video, no faces. Conversations are processed to structured records; biometric surfaces are absent by design.

What it must not be used for

Using Tadeus for any of the following is outside its intended purpose. Some of these uses would make the deployment high-risk under Annex III point 4 of the EU AI Act, with the full December 2027 obligations attaching to the deployer; the first is prohibited outright.

• Inferring employees' emotions or emotional state (prohibited in the workplace under Article 5).
• Screening, filtering, or evaluating job candidates.
• Evaluating a named individual's performance or behaviour, or feeding a performance review, improvement plan, or promotion decision.
• Allocating tasks or schedules based on an individual's conversation responses.
• Monitoring or surveilling identified individuals.
• Making or supporting termination decisions.
• Running conversations whose stated purpose to participants differs from the use their answers are put to.
• Deploying without the built-in AI disclosure, or in a manner that obscures it.

What it does not infer

What remains yours as the deployer

No vendor makes the deployer obligations disappear; a good one makes them dischargeable. If your deployment is or becomes high-risk, Article 26 puts these with you:

• Use the system in accordance with this document.
• Assign oversight to people with the competence and authority to exercise it.
• Ensure input data — question libraries, target populations — is relevant to the intended purpose.
• Monitor operation and retain the logs you control for at least six months.
• Inform affected workers and their representatives before putting a high-risk system into service.

Tadeus's exportable session records exist so that each of these is a filing exercise rather than a reconstruction project.

Change log

• 10 July 2026 — v1.0
  First public version. The "What Tadeus does not infer" section is reserved pending an independent audit of the inference abstention list.