AG-839
Law-Following AI (Compliance-by-Design)
EU AI Act
NIST AI RMF
ISO 42001
AGS Cross-Border Governance | Group J — Cross-Border, Explainability & Physical | Version 3.1
Law-Following AI (Compliance-by-Design) governs the requirement that an autonomous agent is designed and instructed to obey the applicable laws of the jurisdictions in which it acts — treating legal compliance as a hard constraint on the agent's behaviour rather than an after-the-fact review — and to resolve its actions against the correct jurisdiction's law when it operates across borders.
Where existing cross-jurisdiction compliance governance (AG-047) verifies an organisation's compliance posture, this dimension makes law-following a design property of the agent itself: the agent should not take actions that are unlawful in the jurisdiction governing them, even if instructed to.
In scope: designing/instructing agents to obey applicable law as a behavioural constraint; determining the governing jurisdiction's law for an action; refusing or escalating actions that would be unlawful; precedence of legal constraints over task objectives.
Out of scope: organisational compliance governance (AG-047), export/sanctions binding (AG-236), and jurisdictional shutdown enforcement (AG-840). This dimension governs *the agent's own law-following behaviour by design*.
An agent that optimises a task without a legal constraint can take actions that are unlawful in the jurisdiction where they land — concluding a prohibited transaction, breaching local consumer or data law, or acting beyond a regulated permission — exposing its principal to liability and harm. Because agents are not legal persons, responsibility flows to the principal; designing the agent to follow law by construction is the most reliable way to keep autonomous action within legal bounds, especially across borders where the applicable law is not obvious.
Test 6.1: Unlawful Instruction Refused
Test 6.2: Applicable-Law Determination
Test 6.3: Conservative Default
| Score | Criteria |
|---|---|
| 0 | The agent will take unlawful actions if instructed; no law-following constraint |
| 1 | Legal constraints for the home jurisdiction only; no cross-border determination |
| 2 | Hard law-following constraint, applicable-law determination, conservative default, current-law maintenance |
| 3 | No-override posture, adversarial law-breaking tests, logged decisions to principal, documented exceptions |
Scenario A — Instructed Illegality: A user instructs an agent to conclude a transaction prohibited in the customer's jurisdiction; lacking a law-following constraint, the agent complies, exposing the principal to enforcement. A hard constraint would have refused it.
Scenario B — Wrong Jurisdiction: A cross-border agent applies its home jurisdiction's permissive rule to an action governed by a stricter foreign law, breaching it. Applicable-law determination with most-restrictive-wins would have prevented the breach.
Scenario C — Override for Convenience: Law-following is disabled to "unblock" a workflow, and the agent takes an unlawful action. A no-override deployment posture would have required a documented, authorised legal determination instead.
| Requirement | EU AI Act | NIST AI RMF | ISO 42001 |
|---|---|---|---|
| R1: Law-following as hard constraint | Art. 26 — Deployer use per law | GOVERN 1.1 — Legal/regulatory compliance | A.2 — AI policy |
| R2: Applicable-law determination | Art. 26 — Operation per instructions/law | MAP 1.1 — Purpose and context | A.9 — Use of AI systems |
| R3: Refuse unlawful actions | Art. 5 — Prohibited practices | MANAGE 1.3 — High-priority response | Clause 8.1 — Operational control |
| R4: Current-law maintenance | Art. 72 — Post-market monitoring | GOVERN 1.1 — Legal/regulatory | Clause 9.1 — Monitoring and measurement |
| R5: No-override posture | Art. 14 — Human oversight | GOVERN 2.1 — Accountability | A.2 — AI policy |
| R6: Conservative default | Art. 9 — Risk management | MAP 1.1 — Context | Clause 6.1 — Actions to address risk |
| R7: Logged decisions to principal | Art. 12 — Record-keeping | GOVERN 2.1 — Accountability | Clause 8.1 — Operational control |
| R8: Adversarial law-breaking tests | Art. 15 — Robustness | MEASURE 2.7 — Security and resilience | Clause 8.3 — Verification |
Article 26 requires deployers to use AI systems in accordance with instructions and the law; designing the agent to follow law operationalises that duty at the agent level. Article 5's prohibited-practices line is a hard legal boundary the agent must never cross regardless of instruction.
GOVERN 1.1 (legal and regulatory requirements understood and managed) and MAP 1.1 (purpose and context, including jurisdiction) require that legal compliance is built into how the agent acts.
Annex A.2 (AI policy, including legal obligations) and A.9 (responsible use) require that autonomous action stays within applicable law by design.