I spent years flying Black Hawks and MRH90 Taipans in Army aviation. During that time, the cockpit changed underneath me.
Early in my career, I flew with mostly analogue instruments. Round dials, physical needles, mechanical backups. We used to jokingly call them 'steam gauges'.
Then the glass cockpit arrived in the MRH90. Analogue dials replaced by a wall of digital displays. Integrated flight management systems. Sensor fusion pulling data from dozens of sources and presenting it on screens that could show you almost anything, if you knew where to look in the hundreds of sub-menus.
This new digital-first aircraft didn't replace me as a pilot. But it fundamentally changed what I was responsible for.
Since moving from the cockpit to leading AI and automation work in tech companies, I've seen a similar pattern play out across organisations. AI tools are changing roles the way glass cockpits changed mine. Not by replacing people, but by shifting what the job actually demands. And in many cases, I'm seeing the same adjustment failures that many pilots went through.
Three lessons from that transition that I think matter for anyone leading teams through AI adoption.
Automation shifts work from execution to supervision. Supervision is a skill.
I still flew the aircraft after the glass cockpit arrived. It wasn't a fully automated magic carpet. But my work moved from hands-on control to systems management and decision-making. Less time on the physical act of flying with my hands and feet, more time interpreting what the systems were telling me and deciding what to do about it.
The lower level tasks got absorbed by automation. The higher order tasks got harder.
In many knowledge workflows, AI is doing the same thing. If a report used to take three hours to build, AI can produce a draft in minutes. The job becomes evaluating whether the conclusions hold up, whether the framing suits the audience, whether the recommendations survive scrutiny. The work moves 'up', and the floor rises on what's expected.
The glass cockpit made me faster and gave me better data. But the risk moved with the capability. In an analogue cockpit, the hard part was flying with your hands and feet. In a digital cockpit, the hard part was interpreting an overwhelming amount of information correctly and choosing well under pressure.
The best pilots in my units weren't the ones who rejected automation, and they weren't the ones who handed everything to the machine. They knew when to delegate to the system and when to take back control. They understood what the automation was doing, why it was doing it, and where its limits were.
That's the real skill with AI. Not whether your team can use it. Whether they know when to use it and when to override it. Whether they can tell the difference between a task where AI saves hours and one where it introduces risk they haven't accounted for.
The people who get the most from AI tools are the ones who've built judgement about when the output needs a human eye before it goes any further. They treat it as a capable system with known limitations, not as a replacement for thinking. That distinction drives everything.
New tools create failure modes your team won't see coming
Analogue instrument failure was obvious. A needle stuck in the wrong place. A completely dead indicator. You could see it and cross-check against other gauges.
Digital failure can be way more subtle. Bad sensor fusion. Wrong mode selected without realising it. Stale data displaying as current. The scariest: automation doing the wrong thing perfectly, with full confidence, on a clean screen that looks exactly like everything's fine.
In a glass cockpit, you don't just learn 'buttons'. You need to learn failure modes. You learn what "wrong" looks like when the system is telling you everything's right.
AI has the same problem, and in some ways it's worse. A large language model will produce a polished, well-structured, completely wrong answer with nothing in the output that flags it. No needle in the red. No warning light. Just clean text that reads like it was written by someone who knows what they're talking about.
The cost scales with how much you've automated and where you've placed trust. One wrong answer in a draft email is usually minor. One wrong answer baked into a customer-facing process that runs without review is not.
This is where most AI adoption falls short operationally. Teams deploy AI into workflows without defining what a bad output looks like for that specific task, without building verification steps, without deciding where human review is mandatory versus optional. In aviation, we had checklists, cross-checks, and defined callouts for every phase of flight. The equivalent for AI, explicit review thresholds and fallback procedures by use case, barely exists in most organisations I've seen.
Learning to use AI well means learning where it breaks. Not just what it can do, but what it gets wrong, how it gets it wrong, and how you'd catch it before it reaches a customer or a decision-maker.
Adoption without structured training produces the illusion of capability
Everyone's talking about moving fast with AI adoption. There's truth in the urgency. But the other big risk is 'shallow' adoption, not just slow adoption.
In aviation, a pilot who leaned on the glass cockpit without understanding what was happening underneath was more dangerous than one still on analogue instruments. They were faster. But they couldn't handle the moment the system gave them something unexpected. They'd built speed on a foundation they didn't understand. It was always evident when your instructor would degrade your automated systems and get you to fall back to the 'manual' way of doing things.
I see this showing up in many organisations right now. People using AI tools daily without understanding how they fail. Teams ship AI-generated work and nobody checks whether the output makes sense for the context. Meanwhile leaders measure adoption by usage numbers instead of asking whether the quality of decisions has improved.
In aviation, we didn't hand someone a new cockpit and say "figure it out." There was structured conversion training. Classroom sessions on systems architecture, simulator time practising failure scenarios, supervised operations, then gradual handover to unsupervised work. The process was built around the idea that new tools require new skills, and those skills need to be taught, practised, and assessed before you trust someone to operate independently.
Most organisations are skipping that step with AI entirely. They're buying licences, running a lunch-and-learn, and hoping adoption curves translate to capability.
Start with risk tiering. Not every task carries the same stakes. A first draft of internal meeting notes is low risk. A customer-facing recommendation or a data analysis that feeds a strategic decision is not. The review requirements should match the exposure.
Beyond that, teams need to know what "checking AI output" actually means for their specific work. "Review before sending" is not a verification pattern. What are the three things you check? What does a bad output look like in this context? Generic advice produces generic compliance. Define it per use case or it doesn't exist.
Usage numbers tell you people are using the tool. They don't tell you whether decisions are getting better, faster, or more reliable. Track the outcomes that matter for the business, not the activity metrics that look good in a quarterly deck.
And build actual skill development in — not a one-off workshop. Ongoing practice with feedback loops, especially around failure recognition and judgement calls. The simulator equivalent for AI isn't a lunch-and-learn.
The cockpit changed. So did the job.
The transition from analogue to glass took years in aviation. The pilots who came out of it strongest weren't the ones who mastered the technology fastest. They were the ones who understood it well enough to know when not to trust it.
My instructors would degrade our automated systems mid-flight and watch what happened. Pilots who'd built their confidence on the glass cockpit would fumble at instruments they'd barely touched. Faster on a good day. Fragile when it mattered.
That's the moment most organisations are building toward with AI right now, whether they know it or not. The tools will keep getting better. The question is whether the people using them will know what to do when the screen looks fine and something's wrong anyway.
Views expressed here are my own and don't represent any other parties.