AI Governance for Airlines and Aviation Operations

Where Airline AI Agents Touch Operational Systems

Airlines are placing agents in functions that were historically managed through structured, human-reviewed workflows. These agents now read from and write to systems that carry safety and regulatory weight.

Operational Use Cases and the Governance Gap

Airlines are deploying AI agents into crew scheduling, predictive maintenance, flight planning, and ground operations. These agents interact with maintenance logs and scheduling platforms that matter for safety and compliance. The governance question is not whether an airline has a written AI policy, but whether each agent’s access is identifiable, scoped, and enforceable at the moment it acts. A policy document does not stop an agent from calling an API it should not call. Runtime controls do.

Technical Controls Required for Agent Governance

Effective governance rests on five technical controls applied consistently across operational systems.

1. 1

   Distinct Agent Identity

   Each AI agent interacting with scheduling, maintenance, or flight-planning systems should carry a unique, verifiable identity rather than a shared service credential.

2. 2

   Scoped, Least-Privilege Access

   Permissions should be limited to the specific data fields or functions an agent needs, such as read-only access to maintenance logs versus write access to scheduling records.

3. 3

   Runtime Policy Enforcement

   Access decisions should be enforced at the point of tool call or API invocation, not solely through pre-deployment policy documents.

4. 4

   Action-Level Audit Logging

   Logs should capture what an agent accessed, changed, or recommended, at a level of detail sufficient to support safety investigations or regulatory review.

5. 5

   Segmentation by Risk Proximity

   Agents operating in administrative systems should be separated from any agents with proximity to safety-critical operational technology.

Safety Management Systems Were Not Built for Software Agents

FAA regulation 14 CFR Part 5 requires certificated air carriers to maintain a Safety Management System covering hazard identification, risk assessment, and safety assurance. ICAO Annex 19 establishes equivalent State Safety Programme and SMS obligations internationally. Neither framework contains provisions specific to AI agent identity, permissions, or tool-call governance.

These SMS structures are process- and documentation-oriented, built around human-driven hazard reporting cycles rather than real-time technical enforcement of software behavior. EASA’s AI Roadmap 2.0 and its machine learning concept papers introduce trustworthiness and assurance concepts, but they are oriented toward certified avionics and air traffic management systems, not enterprise agents operating in scheduling or maintenance back-office environments. Airlines should not assume this guidance extends to their operational agents without explicit confirmation.

Risks of Excessive or Ungoverned Agent Permissions

When AI agents are granted broad, system-level access instead of task-scoped permissions, the resulting exposure is not theoretical. An agent with unrestricted access to maintenance records can alter or misinterpret data that feeds dispatch decisions. An agent with unmonitored write access to scheduling systems can introduce crew assignment errors without a clear record of how the change occurred.

Because current SMS obligations may be interpreted by regulators or auditors to extend to AI-driven operational decision support, even without AI-specific rules, an airline’s inability to reconstruct an agent’s actions during a safety investigation is itself a governance failure, independent of whether the underlying AI output was correct.

Implementation Path for Airlines

Given the absence of a regulatory mandate, airlines should start by building an internal inventory of deployed AI agents, their permissions, and the operational systems each can reach. This inventory work should be coordinated between IT security teams and safety and compliance teams, since agent governance spans cybersecurity access control and aviation safety risk management.

Existing SMS hazard-identification processes can be extended to capture agent-related risks rather than requiring a separate reporting structure. Agent behavior should be validated in non-production environments before any access is granted to live maintenance or scheduling systems, and human-review checkpoints should be defined for agent actions that affect dispatch, maintenance scheduling, or flight-planning outputs.

• Inventory deployed agents, identities, permissions, and reachable systems
• Coordinate ownership between IT security and safety or compliance teams
• Extend existing SMS hazard identification to agent-related risks
• Validate agent behavior in non-production before live system access
• Define human-review checkpoints for dispatch, maintenance, and flight-planning actions

None of this constitutes regulatory compliance on its own, but it establishes a documented, defensible governance baseline while aviation-specific AI rules remain unsettled.