We treat the EU AI Act as a design constraint, not a checkbox. Two articles shape the product more than any other: Article 4 (AI literacy) and Article 14 (human oversight of AI-graded assessments).
Article 4(b) — no employer-side classification
Employers must promote AI literacy without classifying their staff. We mirror this in the product: org admins, managers and lecturers never see per-person mastery scores, AI literacy tiers or derived badges.
Aggregate-only on employer surfaces
Manager portal, Reporting Hub and admin dashboards show group-level signals only. Personal data stays with the member.
k≥3 suppression
Any aggregate covering fewer than 3 members is suppressed — preventing de-anonymisation through small-group exclusion.
Enforced at the database
The rule is enforced inside the RPC layer, not just the UI. Client requests for per-person classifications return nothing, regardless of frontend.
Member view preserved
Each member keeps the full picture of their own competence in their private workspace.
Article 14 — human oversight on AI-graded assessments
Where AI contributes to a competence call, the user must be able to understand and contest it. Our safeguards:
Server-side consent gate
AI-graded competence flows are off until the member explicitly consents. No quiet enrolment.
Contestability
Each AI-graded result includes the inputs and the rule that produced it, plus a one-click contest path.
Behavioural over inferential
Wherever possible we measure what people did, not what we infer they "are".
Retention floors
7-, 30- and 90-day retention checks frame learning as a curve, not a verdict.
An AI literacy "score" visible to managers. Hidden classification of staff for performance reviews. Surveillance dashboards. We've turned down feature requests for all three.
Article 5(1)(f) — why typing cadence is not emotion recognition
Keypra collects optional Behavioural Telemetry — pause counts, deletion ratios, session duration — to render a private "Crafting Fluency" chart for the member. The Article 5(1)(f) prohibition targets systems that infer emotions or affective states in the workplace. Crafting Fluency does no inference: it counts pauses and deletions and divides them by duration. There is no AI model, no affective label, and the result is shown only to the member who produced it. Full disclosure on the Behavioural Telemetry methodology page.
Annex III §4 — why Keypra is not workplace performance monitoring
Annex III §4 covers AI systems used to monitor and evaluate worker performance. Keypra is contractually prohibited from being used this way: the "Prohibited Uses" section of our Terms of Service and the Behavioural Telemetry annex of our Data Processing Agreement bind Customers and Organisations not to use the Service, or any data derived from it (including ALI scores, Crafting Fluency metrics, retention checks or behavioural telemetry), as the basis for hiring, promotion, performance evaluation, disciplinary action or termination. Combined with our database-level ban on per-person classifications on employer surfaces, this keeps Keypra outside the Annex III §4 scope.