Aviation's lesson for businesses running agents is not about technology. It is that authority follows current, tested competence on a specific type of machine, checked again and again, with the junior person trained to challenge. Copy that structure, without the cost of simulators, and agent oversight stops depending on luck and goodwill.
Why look to aviation at all?
Because aviation solved a problem businesses now have. A highly automated system does most of the routine work, and a small number of trained people supervise it, intervene at defined points and take over when it stops. The automation is certified. So are the people. Neither is trusted on the strength of the other.
Businesses adopting agents often get half of this. They certify the agent, or at least test it, and assume people will cope. Aviation's answer is that a certified aircraft flown by an uncertified crew is not a safe flight. A business running a certified agent with untrained reviewers is in the same position. The structural reasons this needs a platform layer at all are set out in why businesses need an agentic operating system; this piece is about the people half.
Lesson one: qualify people per type, not in general
Pilots hold ratings for specific aircraft types. Experience on one does not authorise them to fly another, because the systems, limits and failure modes differ.
Agents differ the same way. A reviewer who handles a payments agent's held releases at depth 4 is doing different work from one approving a correspondence agent's drafts at depth 2. The bounds differ, the red lines differ, and so do the ways each agent fails. "AI training" in general is the equivalent of a licence with no type on it. Train and assess people for the agent and depth they will oversee, and let a business adopt the practice of tying approval rights to a current certificate for that depth.
Lesson two: competence expires
Aviation does not assume that a pilot who passed a check a year ago can still handle an engine failure. Proficiency is tested on a recurring cycle, and lapsed currency grounds the pilot until it is restored.
Governed agents already work like this. An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. People who approve its actions deserve the same discipline: a certificate with an expiry date, recertification on a cycle, and earlier refreshers when the agent gains a capability or the reviewer returns from a long absence. A reviewer certified on the agent as it stood a year ago is, in aviation terms, current on a type that no longer exists.
Lesson three: know what mode the machine is in
A familiar hazard in automated cockpits is mode confusion: the crew believes the automation is doing one thing when it is doing another. Training puts heavy weight on always knowing the current mode.
The agent equivalent is depth and certificate state. Is this agent executing routine work today, or only drafting because its certificate lapsed yesterday? Was this action held because of the customer's own limit, or because a source was external? The verdict names which bound decided, and reviewers must read it every time. A reviewer who assumes yesterday's mode will approve as routine something the system deliberately stopped.
Lesson four: rehearse the abnormal before it happens
Crews practise failures they may never meet in service, so that the response is drilled rather than improvised. For abnormal situations they reach for a short, tested reference rather than memory.
Businesses can copy both habits cheaply. Write a quick reference guide for the abnormal states of each agent: certificate lapsed, source unavailable, refusal disputed by a customer, suspected fraud in a held payment. Then rehearse them. A walkthrough of each scenario with the reviewers and team lead, using real past cases, costs an hour and shows exactly where people hesitate.
Lesson five: train the junior to say no
Crew resource management grew from the recognition that accidents often follow a junior crew member who saw the problem and did not speak up. Modern training makes challenge a duty, with agreed words for it.
Agent review needs the same culture. The agent's draft looks confident; the senior colleague approved a similar one yesterday; the queue is long. A reviewer who has not been trained and permitted to refuse will approve. Make refusal with a written reason a normal, expected outcome, review the refusals in team sessions, and treat a reviewer who never refuses as a training need, not a star.
Lesson six: debrief, and feed it back
Aviation's safety record rests partly on reporting and learning from near misses. Every event feeds the next round of training.
The agent's record is a ready-made source. Every overturned approval, every refusal and every edit is filed. Use them. Corrections that keep recurring belong with whoever owns the agent's brief, and the people best placed to write the brief are usually those who wrote the process, which is why the people who wrote the process should teach the agent and the reviewers. Control functions can do the same: teaching internal auditors to bank corrections that improve an AI agent turns audit findings into better agents instead of filed reports.
What is the trade-off?
Aviation training is expensive, slow and heavily regulated, and businesses cannot copy its cost. Full-motion simulators, fixed training hours and examiner networks are out of reach for a finance team of twelve.
The structure is cheap; the expense was never the point. Per-agent qualification, expiring certificates, drilled responses to the abnormal, a culture of challenge and a feedback loop from the record can all run on a team's own cases with a few hours a month. The risk is the opposite one: borrowing the vocabulary of aviation without the discipline, and calling a slide deck a type rating.
What should you do differently?
Pick one agent and apply the six lessons to the people around it. Name the reviewers and certify them for that agent and depth. Put an expiry date on their certificates. Write the quick reference for its abnormal states and walk through each one. Make refusal normal. Review the record monthly and send recurring corrections to the agent's owner. The guide to training teams for AI agents covers what each group must learn and the order to train them.
Questions people ask
Is aviation too different from office work to be a useful model?
The stakes differ, but the structure of the problem is the same: automation doing routine work, people supervising at defined points. The habits transfer even though the hazards do not. Borrow the habits, not the drama.
What is the cheapest aviation habit to adopt first?
The walkthrough of abnormal states. It needs a room, an hour and a handful of past cases, and it shows quickly whether reviewers know what to do when an agent's certificate lapses or a payment is held for an unusual reason.
Should reviewers have a checklist for every held action?
For high-consequence actions, yes, kept short. A checklist for a held payment at depth 4 should fit on one screen. For low-consequence drafts, judgement and good reasons matter more than a list.