Point of viewFuture of airlines

Governing AI in a safety-critical business

Airlines can't let an opaque model decide a maintenance finding, a crew assignment or a passenger's refund. The answer is not to avoid AI but to give every agent an autonomy level, an owner and an audit trail.

5 min read By · Point of view

Key takeaways

  • Airline AI spans very different risk levels, from competitor-fare monitoring to airworthiness evidence.
  • A five-level autonomy scale — observe, suggest, act with approval, act within limits, autonomous — makes those differences explicit.
  • Safety-critical work stays at observe or suggest; a licensed person signs every decision.
  • Every agent needs an inventory entry, evaluations, a named owner, a kill switch and a log of what it did.

AI in an airline is not one risk. A model that summarises competitor fares and a model that suggests a maintenance finding should not be governed the same way. Treating them alike either blocks the low-risk work or under-controls the high-risk work.

Make autonomy explicit

We assign every agent one of five autonomy levels. At level 0 it observes and reports. At level 1 it suggests, and a person does the work. At level 2 it prepares the full action and a person approves every case. At level 3 it acts on its own inside policy limits and escalates everything else. Level 4 is fully autonomous with sample review — a level we rarely recommend for airline work.

Exhibit 1

Where airline agents sit on the autonomy scale

Examples from our airline agent designs

LevelExample agentWho decides
0 · ObservePredictive maintenance watcherMaintenance control
1 · SuggestRecovery options plannerOCC duty manager
2 · Act with approvalRefund & compensation drafterRefunds team lead
3 · Act within limitsFare rule extractorReviewer for exceptions

Note: Illustrative designs; autonomy is agreed per agent with operations, safety and finance.

For executives

What this means for your bank

  1. Keep an inventory of every model and agent, with its autonomy level and owner.
  2. Keep safety-critical agents at observe or suggest; require licensed sign-off.
  3. Evaluate agents before and after go-live, and review overrides.
  4. Give operations a kill switch that halts all agents.
Put it to work

How DaasLabs can help

See how autonomy levels, approvals and the kill switch work in practice.

See governance & controls

Set up AI and agent governance with our Data Governance, Privacy & Regulatory Data service.

Learn more

Baseline your AI maturity across 12 capability layers.

Take the maturity assessment
Stay informed

Get new airline insights in your inbox

New perspectives on AI, data and transformation in airline — a few times a month. Browse all insights.