EU AI ACT
& GDPR
COMPLIANCE
BLUEPRINT

Art. 6 mandates classifying the agent against Annex III categories (employment, credit, biometrics, law enforcement, etc.) before any design work. Art. 3 definitions determine whether the agent qualifies as an "AI system" and who is the "provider" vs "deployer." Art. 9 requires a risk management plan to be established from the outset — not retroactively. An agent used in hiring, credit, or critical infrastructure decisions is automatically high-risk, triggering the full Chapter III compliance regime.

Purpose limitation (Art. 5(1)(b)) must be defined before data collection begins — the agent's purpose determines the lawful basis under Art. 6. Art. 35 DPIA screening occurs here: does the agent involve systematic profiling, large-scale special category data, or public monitoring? If yes, a full DPIA must be completed before proceeding. Art. 25 means privacy architecture decisions (federated learning, differential privacy, data minimisation) are selected at this stage.

Art. 14 requires that the agent is designed so a natural person can understand outputs, override decisions, and stop operation — this must be in the architecture, not added later. Art. 11 means design documents, model selection rationale, and architecture diagrams form part of mandatory Annex IV technical documentation. Art. 15 requires cybersecurity-by-design: protection against prompt injection, model evasion, and adversarial inputs. Art. 5 compliance is checked at design: if the agent could enable subliminal manipulation or social scoring, the feature must be removed.

Art. 25 mandates that the minimum necessary data flows are built into the architecture — if the agent can complete its task without retaining conversation history, it must not retain it by default. Art. 22 determines whether the agent's decision pipeline is "solely automated" — if it produces legally significant effects, a mandatory human review step must be architected in. Art. 32 requires encryption, access controls, and isolation to be in the design specification, not security-bolted-on post-launch.

Art. 10 is the data quality mandate: training data must be relevant, sufficiently representative across demographic subgroups, and free from errors. Bias audits are mandatory for high-risk agents — underrepresentation of protected groups in training data creates discriminatory model outputs. Art. 10(5) allows processing special category data in training only to the extent strictly necessary for bias detection and correction. If the agent uses a GPAI foundation model, Art. 53 requires the GPAI provider to have documented training data sources and copyright compliance.

Every data subject whose personal data appears in the training set must have a lawful basis under Art. 6 for that processing. Consent (Art. 7) or research exceptions (Art. 89) are commonly cited. Art. 5(1)(c) data minimisation means only data necessary for the training objective is included — a name-prediction model should not train on health data. Art. 17 creates the machine unlearning obligation: if a person withdraws consent, their data must be removable from training pipelines. Art. 44 applies if training data is transferred to or processed in non-EU cloud infrastructure.

Art. 15 requires testing for robustness against adversarial inputs (prompt injection, jailbreaking, data poisoning attacks) — this happens during training and fine-tuning. Art. 16–17 QMS mandates documented procedures for model development, testing criteria, and sign-off processes. For systemic risk GPAI agents, Art. 55 requires structured red-teaming before deployment. Safety alignment (RLHF or constitutional AI approaches) is the mechanism that implements Art. 5 prohibited practice avoidance at the model level.

Art. 32 applies to the training infrastructure: compute clusters handling personal data must have access controls, encryption, and audit logging. Privacy-enhancing technologies mandated by Art. 25 are implemented here: differential privacy (adding calibrated noise to gradients), federated learning (keeping data on-device), or secure multi-party computation. Art. 28 requires DPAs with every cloud provider, GPU compute vendor, or annotation subprocessor in the training pipeline.

Art. 43 prescribes the conformity assessment route: self-assessment for most Annex III systems; mandatory notified body for biometric and critical infrastructure AI. The Annex IV technical documentation package (Art. 11) must be complete and on file before the Declaration of Conformity (Art. 49) is signed. Art. 47 mandates registration in the EU AI public database for high-risk systems. Art. 13 requires publishing clear Instructions for Use for deployers — including system capabilities, limitations, accuracy statistics by demographic group, and human oversight requirements.

The DPIA (Art. 35) must be completed and signed off before launch — if residual risk cannot be mitigated, Art. 36 prior consultation with the supervisory authority is mandatory (this can delay launch significantly). Art. 30 Records of Processing Activities must be updated to reflect the new AI processing activity. The DPO (Art. 37–39) must review and sign off on DPIA findings and ROPA entry before launch approval.

Art. 26 places active obligations on the deployer: assign human oversight responsibility, monitor performance against documented metrics, and ensure the agent is only used for its documented intended purpose. Art. 12 logging is live from day one: inputs, outputs, timestamps, operator IDs, and confidence scores are automatically logged. Art. 50 requires clear disclosure that users are interacting with an AI system. Art. 14 human oversight is operationally active: a human reviewer can pause the agent's decision output at any time. Art. 72 post-market monitoring plan goes live on deployment day.

Privacy notices (Art. 13–14) must be presented to users at the point of first interaction, disclosing AI processing and any automated decision-making. Art. 22 safeguards are live: if the agent makes legally significant decisions, a human review route must be available and communicated to users. Art. 32 security controls (WAF, rate limiting, anomaly detection, encryption) are active. Data retention schedules (Art. 5(1)(e)) are enforced: interaction logs are deleted after the defined retention period with automated purging.

Art. 26(6) requires deployers to inform natural persons when they are subject to an AI system's output — especially in high-risk contexts (credit, hiring, benefits). The agent's interface must display a clear "AI-powered" disclosure. Art. 14 mechanisms let users invoke human review: a clearly visible "Request human review" button or escalation pathway. Art. 13 transparency means affected persons can access the agent's documented accuracy, limitations, and intended use scope.

A dedicated rights management portal must be built into the agent ecosystem: Art. 15 SAR responses within 30 days including AI decision logic explanations; Art. 17 erasure triggers machine unlearning workflows; Art. 20 portability exports interaction data in structured JSON/CSV; Art. 21 objection immediately pauses AI profiling; Art. 22 contestation routes the specific decision to a human reviewer who must provide a reasoned response. All rights requests are logged in the ROPA and reported to the DPO monthly.

Art. 72 requires a formal post-market monitoring plan: KPIs for accuracy, fairness metrics by demographic subgroup, false positive/negative rates in high-risk decisions, and a reporting cadence to the market surveillance authority. Art. 73 mandates that serious incidents (death, serious harm, infrastructure disruption) are reported within 72 hours (life-threatening) or 15 days (serious). Discovered performance degradation or bias drift requires updating the Art. 9 risk register and implementing corrective actions. Regular adversarial testing (Art. 15) continues post-deployment to catch new attack vectors.

Art. 33 breach notification obligations are operationalised: a 72-hour incident response plan is maintained with pre-approved supervisory authority notification templates. Art. 32 security monitoring includes anomaly detection for unusual data access patterns by the AI system. If the agent's scope or data processing changes substantially, a new DPIA (Art. 35) is triggered. Art. 5(1)(d) accuracy principle means model drift resulting in inaccurate outputs about individuals must trigger correction — e.g., stale credit profiles must be updated before influencing decisions.

A "substantial modification" (changed architecture, new training data, new intended purpose, materially different performance) restarts the conformity assessment (Art. 43) and requires an updated Declaration of Conformity (Art. 49) and updated Annex IV technical documentation (Art. 11). Minor updates require an updated risk register entry. At retirement, Art. 12 requires operational logs to be retained for the minimum period specified in the conformity assessment (often 10 years for high-risk AI in regulated sectors). The EU database registration must be updated to reflect retirement.

At retirement, Art. 5(1)(e) storage limitation requires deletion of all personal data from the agent's databases, model weights (if memorisation is a risk), and operational logs beyond the defined retention period. Art. 17 erasure obligations must be flowed to all processors (cloud providers, sub-processors) with written confirmation of deletion. The ROPA (Art. 30) is updated to mark the processing activity as ceased. If the change involves a new processing purpose, a fresh DPIA (Art. 35) is mandatory. DPAs with processors must include data return/destruction clauses activated at contract termination.