AI Governance for Aerospace Manufacturing Agents

Why Aerospace AI Agents Require Runtime-Specific Governance

Failure Modes Specific to Aerospace Manufacturing Agents

• Unauthorized access to ITAR-controlled data: Agent queries against PLM or engineering systems surface controlled technical data without export authorization checks.
• Unsupervised production parameter changes: An agent modifies machining, tolerance, or process parameters without triggering AS9100 change control.
• Permission sprawl across interconnected systems: Broad, platform-wide credentials are granted to agents operating across MES, PLM, QMS, and supplier portals.
• Insufficient audit trails for regulatory review: Agent actions are logged inconsistently, falling short of FAA Part 21 and Part 43 recordkeeping expectations.
• Missing human review on safety-critical actions: Agents execute changes to configuration-controlled items without the review step AS9100 expects for equivalent human changes.

Technical Controls: Agent Identity, Permission Scoping, and Tool-Call Governance

In the absence of aerospace-specific technical standards, NIST SP 800-207's Zero Trust Architecture provides the closest documented model for scoping agent access. Its core principles (per-request authentication, least-privilege access, and continuous verification independent of network location) apply directly to agents operating across interconnected MES, PLM, QMS, and supply chain systems.

Practically, this means each AI agent should carry its own identity, distinct from the service accounts or human credentials it acts on behalf of, with permissions scoped to the specific system and action it needs rather than platform-wide access. An agent responsible for quality inspection data entry should not carry the same permissions as one supporting supply chain coordination, even if both operate within the same manufacturing environment.

Tool-call actions, particularly those affecting production parameters, engineering records, or export-controlled data, should route through a policy enforcement point capable of evaluating each call against defined rules before execution. This enforcement point can log every tool call for audit purposes and, where the action affects safety-critical or configuration-controlled parameters, require human approval before the call is allowed to proceed. NIST AI 600-1 identifies expanded system access as a specific risk category for generative AI systems, reinforcing the need to constrain what an agent can reach rather than relying on downstream review alone.

CISA and the UK NCSC's joint guidelines for secure AI system development recommend that access control, logging, and monitoring be built into the AI system lifecycle rather than layered on after deployment. For aerospace manufacturers, governance controls need to be part of the initial agent deployment architecture, not a compliance step added once agents are already operating in production.

Evaluation Criteria for Aerospace AI Governance Leaders

• Does the platform scope tool-call permissions per connected system (MES, PLM, QMS) rather than granting broad access?
• Can audit logs support AS9100 traceability and FAA Part 21/43 recordkeeping review?
• Is ITAR-controlled technical data identified and segregated from agent processing and output?
• Does the platform require human approval before agents modify safety-critical or configuration-controlled parameters?
• Can agent identities and permissions be revoked or updated across interconnected supply chain and production environments?
• Were agent permission boundaries tested in non-production environments before granting live system access?

Runtime Governance Requirements at a Glance

Agent Identity & Permissions

Scoped, per-system credentials distinct from human or service accounts.

Tool-Call Governance

Policy enforcement before agent actions reach MES, PLM, or QMS systems.

Audit Logging

Traceable records aligned with AS9100 and FAA recordkeeping expectations.

Regulatory Alignment

Mapping agent behavior to AS9100, ITAR, and FAA obligations.