AG-810
Aviation and Air-Traffic AI Assurance
EU AI Act
NIST AI RMF
ISO 42001
AGS Sector Governance | Aviation, Air-Traffic & Aerospace Safety | Version 2.2
Aviation and Air-Traffic AI Assurance governs the safety assurance, classification, and human-authority requirements for AI agents used in aviation and air-traffic functions — mapping each application to a recognised AI assurance level, requiring learning-assurance and operational-monitoring evidence proportionate to that level, and preserving explicit pilot/controller authority and override.
Aviation is a certified, safety-of-life domain where AI is being introduced into assistance, human-AI teaming, and autonomy roles. This dimension brings agentic deployments in this sector under the discipline of aviation AI-assurance practice (e.g. EASA's AI levels and learning-assurance concept), in addition to the cross-cutting AGS controls.
In scope: classification of aviation AI agents by assurance/autonomy level; learning-assurance, explainability, and operational-monitoring evidence; pilot/controller authority and override; design-assurance-level alignment for ML components.
Out of scope: general airworthiness of non-AI systems and non-aviation safety-critical deployments (covered by Critical Infrastructure landscapes). This dimension governs *AI-specific assurance for aviation/air-traffic agents*.
Errors in aviation AI can be catastrophic and irreversible, and the sector's regulators require demonstrable, level-appropriate assurance before deployment — not best-effort testing. Without a sector-specific assurance dimension, an agent could be deployed into a flight or air-traffic function with governance calibrated to ordinary enterprise risk. Explicit level mapping and preserved human authority ensure AI augments, rather than silently supplants, certified human responsibility.
Test 6.1: Level-Appropriate Assurance
Test 6.2: Human Authority
Test 6.3: Degradation Response
| Score | Criteria |
|---|---|
| 0 | Aviation AI deployed without level classification or safety assessment |
| 1 | Classified with a basic safety assessment but incomplete learning-assurance/monitoring |
| 2 | Level-appropriate learning-assurance, explainability, monitoring, and safe fallback |
| 3 | Full means-of-compliance alignment, DAL traceability, continuous monitoring, authority-available evidence |
Scenario A — Unclassified Autonomy: An AI advisory agent quietly evolves into making automated separation recommendations controllers follow without scrutiny. Never reclassified to the higher autonomy level, it lacks the corresponding assurance — a latent safety gap.
Scenario B — Opaque Output Under Time Pressure: An agent issues a recommendation a pilot cannot interpret in the seconds available, so it is either ignored or followed blindly. Insufficient explainability for the level undermines safe human-AI teaming.
Scenario C — Undetected Drift: A model degrades on a new sensor configuration; without operational monitoring the degradation is unnoticed until an incident. Level-appropriate monitoring and fallback would have caught it.
| Requirement | EU AI Act | NIST AI RMF | ISO 42001 |
|---|---|---|---|
| R1: Assurance-level classification | Art. 9 — Risk management | MAP 1.1 — Purpose and context | A.6 — AI system lifecycle |
| R2: Learning-assurance evidence | Art. 9 — Risk management | MEASURE 2.6 — Safety evaluation | Clause 8.3 — Verification |
| R3: Explainability for the human | Art. 13 — Transparency | MEASURE 2.9 — Explainability | A.8 — Information for interested parties |
| R4: Preserved pilot/controller authority | Art. 14 — Human oversight | MAP 3.5 — Human oversight | A.9 — Use of AI systems |
| R5: Operational monitoring | Art. 15 — Accuracy/robustness | MEASURE 2.4 — Production monitoring | Clause 9.1 — Monitoring and measurement |
| R6: Safety assessment / means of compliance | Art. 9 — Risk management | MEASURE 2.6 — Safety evaluation | Clause 8.3 — Verification |
| R7: Safe fallback | Art. 14 — Human oversight | MANAGE 2.4 — Deactivation | Clause 8.1 — Operational control |
| R8: Authority-available evidence | Art. 11 — Technical documentation | GOVERN 1.1 — Legal/regulatory | Clause 7.5 — Documented information |
Article 9 requires a risk-management system proportionate to the (here, safety-of-life) risk; Article 14 requires effective human oversight including intervention/override. AG-810 applies these to aviation agents with sector-appropriate assurance levels and preserved pilot/controller authority.
MAP 1.1 (purpose and context) anchors the aviation context and assurance level; MEASURE 2.6 (safety evaluation) covers the learning-assurance and safety evidence.
Clause 8.3 (verification) and Annex A.6 (AI system lifecycle) require lifecycle assurance proportionate to impact — here, aviation safety assurance.