The framework

P1

Decision Sovereignty

The organisation retains ultimate responsibility for the decisions taken by its AI agents, whatever their autonomy level.

Principle

An autonomous AI agent never moves decision accountability outside the organisation. Operational delegation (the agent acts) does not entail responsibility delegation (the organisation answers).

Operational implications

• Every agent is attached to an identified DDAO (Delegated Decision Agent Officer).
• Every decision perimeter is bounded by a written mandate (cf. ACF-12).
• Every critical threshold breach escalates to a human (cf. ACF-09).

P2

Doctrinal Traceability

Every automated decision must be reconstructible after the fact by pointing to the doctrine, rules and data that produced it.

Principle

An agentic decision without a reconstructible audit trail is legally indefensible. ACF® mandates that, at any time T, the organisation can answer the question: 'which doctrine version, which rules and which data produced this decision?'

Operational implications

• Time-stamped decision register (cf. ACF-05).
• Doctrine versioning (doctrine_version, doctrine_hash, doctrine_archive_url).
• Retention of structured inputs and outputs for the applicable legal periods.

P3

Ultimate Human Control

Every autonomous agent must include a human takeover mechanism, activatable at any time and calibrated to criticality.

Principle

An agent's autonomy is a conditional privilege, never a vested right. At any moment, an authorised operator must be able to freeze, redirect or revoke the agent. This is the operational expression of AI Act Article 14.

Operational implications

• Documented and tested kill switch (cf. ACF-07).
• Explicit escalation thresholds by criticality (cf. ACF-09).
• Regular human-takeover drills (at minimum quarterly for high criticality).

P4

Proportional Governance

Control intensity must be proportional to the agent's actual autonomy and to the criticality of the decisions it takes.

Principle

An N0 agent (assistance) does not warrant the same control apparatus as an N3 agent (autonomous execution). Over-investing controls on a low-autonomy agent wastes resources; under-investing on a high-autonomy agent creates regulatory and operational exposure.

Operational implications

• ACF-02 criticality matrix calibrated by dimension × business impact.
• Control selection from ddao-controls-mapping (level × risk → controls).
• Annual recalibration review.

N0

N0 — Assistance

The agent suggests; the human decides systematically. The human keeps control over every executable action.

Controls: 100% human control. No kill switch needed (the human is the kill switch). Logs of suggestions and the matching human decisions.

Examples: Writing copilot, non-personalised product recommendation, email drafting assistant.

N1

N1 — Supervised recommendation

The agent proposes a justified decision. The human validates or rejects. Every executable action goes through a human validation, but on an already-shaped option.

Controls: Human validation ≥ 95% of actions. Audit log of all proposals (accepted and rejected). Acceptance metrics tracked (drift signal).

Examples: Support ticket triage with proposed reply, pricing suggestion with manager sign-off, lead qualification with sales sign-off.

N2

N2 — Conditional execution

The agent decides and executes within a bounded perimeter. Above the thresholds defined in the mandate, it must escalate to a human.

Controls: Mandatory formal mandate (cf. ACF-12). Explicit numeric and qualitative thresholds. Active kill switch tested quarterly. Structured audit log.

Examples: Customer refund agent up to €200 without sign-off, stock reordering agent within a window, auto-correcting moderation agent.

N3

N3 — Autonomous execution

The agent decides and executes without prior human validation, within the mandate perimeter. Human control remains possible after the fact and the kill switch stays active.

Controls: Deep formal mandate, reinforced governance (monthly review minimum), multi-level kill switch tested monthly, decision-grained audit log, real-time drift metrics.

Examples: Bounded algorithmic trading agent, critical operational routing agent, L1 incident management agent.

D1

Strategy & governance

Alignment of agentic AI usage with the business strategy, dedicated governance, executive sponsorship.

Practices: AI committee constituted. Documented agentic strategy. Identified executive sponsor. Ring-fenced annual budget. Usage metrics tracked at exec level.

D2

Doctrine & framework

Adoption of a reference framework (ACF® or equivalent), publication of an internal doctrine, team training.

Practices: Formalised reference framework. Published and accessible internal doctrine. Training plan deployed. Methodological cards in operational use.

D3

Design & technical control

Secure agent design (mandates, kill switches, observability), mastery of the foundation models in use.

Practices: Formal mandate for every N2+ agent. Documented and tested kill switch. End-to-end observability. Inventory of models in use.

D4

Accountability & roles

Explicit role definitions (DDAO, DPO, CISO, Compliance Officer, Business Owner) with a clear RACI on agentic decisions.

Practices: DDAO appointed per N2+ agent. Published RACI. Documented sign-off process by criticality. Formal delegations.

D5

Regulatory compliance

Mastery of AI Act, GDPR, DORA, NIS2, ISO 42001 obligations according to sector and jurisdiction.

Practices: Up-to-date AI inventory. System-level qualification (provider/deployer, controller/processor). DPIA for high-risk systems. Article 49 register if high-risk. Digital Omnibus compliance roadmap.

D6

Audit & continuous improvement

Periodic internal audit of agents in production, incident review, doctrine updates.

Practices: Annual audit minimum. Quarterly incident review. Documented doctrine updates. Post-mortems published internally.