Share with your CTO
EasyJet’s CTO is making a pointed claim: the airline expects agentic AI to handle more than 60% of all online transactions within four years, and the companies that will execute on that shift are the ones that built composable, API-first infrastructure before the agentic wave arrived. Writing in InformationWeek, the executive argues that monitoring and interpretability for AI agents deserve the same rigor as financial controls, and that unmonitored agents compound performance problems silently until the damage is structural rather than correctable.
What this means for your business
The 60% figure is almost certainly directional rather than precise, but the direction is what matters. Enterprises that are still running monolithic booking or operations stacks face a compounding disadvantage: every quarter spent planning a future modular architecture is a quarter during which companies that already have one are training agents, capturing feedback, and widening the gap. If your current IT estate requires a system rewrite to plug in a new AI agent, you are not behind on AI deployment specifically, you are behind on the architectural prerequisite for it.
The piece’s strongest claim is also its most underappreciated one: unmonitored agents drift. This is not a compliance talking point, it is a performance argument. An agent optimizing flight load factors or passenger rebooking during disruption is making hundreds of micro-decisions per hour, and without logged reasoning and feedback loops, you cannot tell when its outputs have quietly diverged from business intent. The recurring failure mode looks like this: a model performs well in testing, gets deployed, and then degrades incrementally in production because no one built the observability layer to catch the drift before it compounds into a costly pattern. Financial controls caught this dynamic decades ago with variance reporting; agentic AI needs the equivalent infrastructure now.
The author writes from inside an airline deploying agents at scale, which makes the operational specificity credible, though it also means the framing tilts toward organizations with the runway to invest in platform engineering as a dedicated discipline. Smaller or less capitalized enterprises may find the “composable architecture as prerequisite” argument more aspirational than actionable in the near term. The genuine falsification condition here is narrow: if agent performance proves auditable and correctable within legacy monolithic stacks, the architecture argument weakens considerably. Nothing in the evidence so far suggests that is happening.
Concept deep-dive: Composable architecture
Composable architecture means building enterprise systems out of interchangeable, independently deployable modules connected through APIs, rather than one large integrated codebase. Think of it as the difference between a modular synthesizer where you swap individual components and a fixed keyboard where everything is soldered together. For agentic AI, it matters because agents need to query, write, and coordinate across reservations, customer service, and operations simultaneously, and only modular systems let you connect or upgrade those integrations without rebuilding the whole stack.
Based on reporting from CTO: Enterprise AI agents require monitoring and oversight, originally published 2026-07-31 12:22:00.
