Designed to the rules for high-risk AI.
- Frame
- Annex III, point 4
- Human oversight
- Before release and in use
- High-risk rules apply
- 2 December 2027
Atlas does not argue its way out of the high-risk class. Its use in hiring and work decisions falls within Annex III, point 4, and Atlas designs to those rules before they apply.
The AI Act sets out prohibited practices, rules for high-risk AI systems and transparency duties for certain others. Each carries its own duties.
The Contribution Map supports hiring decisions, which Annex III, point 4(a) names directly. The Team Map supports decisions about roles and tasks, which point 4(b) covers. Your Atlas goes to the person it describes. Under Article 6(3), an Annex III system that profiles people is always high-risk.
The work concerns people's careers, so Atlas designs to the rules written for that work.
Outside the product: biometric data, emotion recognition and social scoring. Atlas works from people's own answers and the organisation's documents. The practices the Act prohibits at work have applied since 2 February 2025.
Atlas informs the decision. People make it. That order holds for all Atlas analyses: the Contribution Map, the Team Map and Your Atlas.
- 01 · InputEvidenceApproved role, answers and the organisation's documents
- 02 · MachineAtlas analysesFixed model and prompt versions. Checks raise flags.
- 03 · HumanA named expert reviewsReads the whole analysis. Can change it or stop it.
- 04 · HumanThe client decidesRecords the decision and its reason.
- 05 · LearnFollow-upWhat happened next is recorded for monitoring, not for model training.
A real check
A named Atlas expert reads the whole analysis before release, with the authority to change it or stop it.
A recorded check
The design records who reviewed, when, and which version of the analysis and of the AI set-up.
Machines flag
Automated checks raise flags for the reviewer. A result changes only when a person changes it.
The client decides
An analysis is one input to an interview or a decision. The decision and its reason belong to the client.
For each analysis, the design keeps a record of what went in, what ran and who approved it, so that each finding can be traced back to its inputs and its review.
When a person asks why an analysis says something about her, this chain is the answer.
- Role requirementSet by the client
- Trait measuredOne of 34 constructs
- ResultWith its evidence
- Sentence in the analysisWhat the reader sees
The Act holds a provider to the intended purpose it states. Atlas states one intended purpose for each analysis and builds the product around it.
One purpose, stated everywhere
A Contribution Map describes how one person fits one role. It supports an interview and ranks no one.
Areas in words, with evidence
Each area is described in words, with the evidence underneath. There is no overall score and no page that sorts candidates.
Only against an approved role
A Contribution Map runs only against a role profile the client has approved.
Thin evidence says so
Where evidence is thin, Atlas writes “not enough data”. Missing evidence is not shown as a weakness.
Tendencies, not verdicts
Findings describe tendencies, such as “not naturally drawn to” or “likely to cost energy”. Clinical terms are excluded.
Controlled change
Prompts and model versions are fixed per release. A change goes live only after it passes Atlas's tests.
Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744. Each line is a provider duty and Atlas's design for it.
Risk management
A risk register, including reasonably foreseeable misuse, reviewed at each release.
Data and bias
Results tested for disadvantage by gender, age and nationality before launch, and again as data grows. Any special-category data used for this follows the Article 4a safeguards.
Technical documentation
Technical documentation on purpose, design, data, testing and known limits.
Record-keeping
Automatic logs for each analysis. Logs under Atlas's control are kept for at least six months.
Instructions for use
What an analysis is for, how to read it and where its limits are.
Human oversight
A named Atlas reviewer before release, and the client's own overseer in use.
Accuracy, robustness, security
Evaluations for each release, fixed model versions, and an independent security review of code and live platform (May 2026, zero open findings; a point-in-time result, not ISO 27001 or SOC 2).
Quality management
A quality management system for design, change control and release, proportionate to Atlas's size.
Conformity and registration
Conformity assessment by internal control (Annex VI), EU declaration of conformity, CE marking and registration in the EU database, before the system is placed on the market.
Monitoring and incidents
A post-market monitoring plan, and reporting of serious incidents within the Act's deadlines.
- Uses Atlas in line with the instructions for use
- Assigns oversight to a person with the competence, training and authority to act
- Informs workers' representatives and affected workers before use at work
- Tells each person a decision concerns that a high-risk AI system is used
- Keeps the logs it controls for at least six months
- Explains to an affected person, on request, the role the AI played and the main elements of the decision
- Instructions for use and a short reading guide
- A notice for the people an analysis concerns
- A template for informing employee representatives
- An export of the log for the client's own files
- Documentation for procurement and the DPO
Atlas is in beta. It launches when each part of this page works as described.
The date for Annex III systems moved from 2 August 2026 to 2 December 2027 under Regulation (EU) 2026/1744, in force since 27 July 2026. Conformity assessment, CE marking and EU registration are provider duties before a system is placed on the market.
How does Atlas People handle the EU AI Act?
Atlas People's use in hiring and work decisions falls within Annex III, point 4 of the EU AI Act, and Atlas designs to the high-risk rules before they apply on 2 December 2027. A named expert reviews each analysis before release, the client makes every decision, and Atlas states no compliance status before its conformity assessment.
Is Atlas certified under the AI Act?
No. For this use the Act has no external certificate. The provider assesses conformity itself under internal control, signs an EU declaration of conformity, applies the CE marking and registers the system in the EU database. Atlas states its status only once that assessment is done.
Does the AI make the decision?
No. A named Atlas expert reviews each analysis before release, and the client makes the decision.
Does Atlas use emotion recognition or biometric data?
No. Atlas works from people's own answers and the organisation's documents.
Can a person see why an analysis says what it does?
Yes. Each person receives their own results, and the client receives what it needs to explain the role Atlas played in a decision, as Article 86 gives affected people the right to ask.
Is client data used to train AI models?
No. Client data is not used to train models and is not shared across organisations.