Every airline has an AI strategy. Fewer have a clear answer for who runs the model once it’s live, and the shift from models that advise to agents that act makes that question urgent. This piece is about that gap.

Why the strategy isn’t the problem

Walk into most airline AI reviews today, and the deck looks familiar: a strategy document, a steering committee, a named AI lead with a dotted line to the CIO. And still, model after model gets stuck between the pilot that worked in a demo and a production system nobody quite trusts yet.

That is not a strategy problem. Strategy tells an airline what to do with AI: reduce delays, resolve invoice discrepancies, or speed up refunds. It does not say who owns AI, who is accountable when something breaks, or how the work continues once the pilot team moves on. That is a harder question, less about willingness than the difficulty of follow-through.

And it is getting harder rather than easier. A model that recommends produces a suggestion someone can ignore. An agent that acts —releasing a stand, rebooking a passenger, adjusting a crew pairing — produces a consequence. The moment AI moves from advising to acting, the question of who owns it stops being a governance formality and becomes an operational control.

BCG’s guidance, drawn from more than a thousand AI implementation programs, puts the ratio at roughly 70/20/10, weighted heavily toward people and processes over the algorithms themselves. That is the shape of the problem: not what to do, but who does it, and how.

Why aviation specifically stalls here

Aviation makes this harder than most industries. Ownership is scattered across the airline, the airports it flies into, MRO shops, ground handling partners, and OEMs. No single team owns AI end to end, because no single team owns the operation end to end either. An agent that touches a turnaround touches five organizations’ systems and at least three organizations’ liability.

Underneath that sits a layer of legacy systems predating most executives now trying to modernize around them, and a regulatory picture still taking shape, from EASA’s path for approving machine learning in safety-related systems to the EU AI Act’s obligations for aviation. Neither will be settled before airlines need to ship, so compliance has to get built into delivery work, not bolted on afterward.

From strategy document to operating model

The useful reframe is simple, even if building it isn’t. A strategy tells you whether to invest in AI. An operating model tells you how the work actually gets run once you have.

The pattern shows up well beyond aviation. McKinsey found that nearly two-thirds of enterprises have experimented with AI agents, but fewer than ten percent have scaled them to real value and identified data limitations at the top of the reasons why. Close behind sits the operating model: changing how the organization is actually run. In most airlines, that’s the item with no owner.

BCG’s recommendation lands in a similar place: a central function that orchestrates delivery across silos and supplies change-management capability most airlines lack. Call it an AI Center of Excellence if that’s the language you use. The label matters less than what it owns:

  • Intake and prioritization, so someone can say yes, and no, to a use case against defined criteria
  • Guardrails, enforced inside the delivery pipeline rather than checked after the fact
  • Data stewardship, with a named owner for what a model or an agent is allowed to touch
  • Compliance checkpoints, tied to regulatory triggers rather than a calendar date

That structure only works with real teeth: it holds the intake queue, the release gate, and a budget, not just a seat at the table. And it only works if it stays thin, one intake path, one set of guardrails, with the work staying close to the team that understands it.

What this looks like in practice

A fuel invoice reconciliation model might rank above a nice-to-have chatbot for gate agents, not because it’s flashier, but because it ties directly to discrepancies, finance is already paying people to chase down manually.

Or take predictive maintenance. A model that flags a likely engine fault carries risk either way: ground the aircraft too often and turnaround times collapse, miss a real signal, and the exposure is worse. A human already signs the release to service, a regulatory requirement, not a design choice. What the operating model decides is everything around that signature: the confidence threshold that triggers a review, and what happens the day the model’s behavior drifts from approved data.

That’s the difference between a governance framework on a slide and one that’s actually running: a named owner for the data, a checkpoint instead of an annual audit, and a system that catches drift before a headline does.

Continuing the conversation at World Aviation Festival

This is exactly the ground DataArt’s Dmytro Baikov will cover on stage at the World Aviation Festival in Lisbon, on October 14 at 13:20, on the panel Creating Safe & Effective Foundations for AI Transformation within the Airline. The session covers what leaders building safe, reliable AI frameworks should prioritize, how to balance innovation with risk, how airlines can measure AI maturity, and how to build platforms that won’t buckle under tomorrow’s technology.

If your airline has a strategy but no clear owner for what happens after a pilot works, that’s worth talking through in person. Come find Dmytro at the session, or stop by DataArt’s booth, 1-168B.

Article by DataArt