BLUF — The thing blocking your operation is not whether the technology works. It is that nobody has decided who owns a decision the operation no longer makes by hand. Aviation solved this problem sixty years ago, and not by building a better autopilot. It built named roles, a standing duty to monitor, and a written threshold where a human takes the controls.
The demo was never the hard part
Notice what did not go wrong in that room. The agent did not hallucinate, miss a delivery or invent a supplier. The objection had nothing to do with capability, and another month of tuning would not have answered it. Yet tuning is usually what happens next, because accuracy is measurable and improving it feels like progress, while the actual blocker sits in an org chart, unowned.
Your operation does not have a technology gap. It has a signature gap: a decision the operation is now capable of making, that nobody is yet willing to put their name against.
Autopilot flew hands-off in 1914. an airliner was certified to land with it in 1964.
The Sperry Corporation built the first aircraft autopilot in 1912. In 1914 Lawrence Sperry, then twenty-one, demonstrated it by flying a biplane down the Seine with both hands in the air and his mechanic standing on the wing. The technology worked, publicly and unmistakably, in front of witnesses.
A commercial airliner was not certified to land using it until 1964. The Sud-Aviation Caravelle made the first automatic landing in September 1962 and was certified for low-visibility minima two years later. Half a century separated the demonstration from the permission.
Plenty of engineering happened in that window. But aviation never lacked urgency about fog, and the thing that finally unlocked it was not a better gyroscope.
What aviation built instead was an accountability model
Three things had to exist before an airline would let the machine fly the aircraft, and none of them are technical.
Named roles. Crew duties were formally divided into pilot flying and pilot monitoring. Not a shared responsibility that both would attend to, but two named jobs with different obligations, assigned before the aircraft left the gate.
A standing duty to monitor. Regulators require that at least one pilot monitor the autopilot’s behavior at all times, and expect the crew to notice when it is not performing safely and to take timely corrective action. The obligation does not pause because the automation is doing well.
A written handover point. The Caravelle’s early certification allowed automatic control down to fifty feet above the threshold, at which point the human landed the airplane. The boundary between machine authority and human authority was not a matter of judgment in the moment. It was a number, agreed in advance.
Aviation did not make the autopilot more trustworthy in the abstract. It made the accountability legible, so a specific person could answer for a specific decision at a specific moment.
Your operation needs the same three things, and probably has none of them
The translation is close to literal, which is why this parallel is worth more than an analogy.
A name against each decision, not a team and not a function. A standing duty to supervise, meaning somebody watches what the agents do while they do it rather than reviewing a sample next quarter. And a threshold written down, so the conditions under which a human takes over are decided while everyone is calm.
What changes between industries is only the unit of the threshold. In procurement it is usually money or schedule risk: this substitution goes through, above this value it comes back to the category manager. In prior authorization it is clinical. A request that matches the payer policy cleanly can be submitted automatically; one where the medical necessity is arguable routes to a clinical reviewer, and the line between those two is a documented rule rather than a feeling. Aviation used fifty feet because altitude was the variable that mattered. Yours will be a dollar amount, a confidence score, or a diagnosis category.
None of that requires a platform decision. All of it requires a decision that only you can make, which is probably why it stays open.
Nobody will run what nobody will sign.
What this looked like in practice
In a governed procurement operation we run, this was the work that actually took the time. Building the agents was the shorter half. The longer half was going workflow by workflow and settling three questions for each: who owns this class of decision, what threshold brings it back to a person, and where the record lives. Those conversations were slower than the engineering and less fun. Once they were settled the operation absorbed three times the volume against a flat headcount, and the reason it was permitted to is that every automated decision had a name behind it and a line it would not cross without asking.
I do not know how much of this structure a smaller operation genuinely needs. A four-person team can hold it informally for a long while, and I have seen that work. What I have not seen work is discovering the question late, in the meeting where you hoped to get approval.
Find one unsigned decision this week
This is an afternoon of work, not a program.
• Pick one decision your agents already make, or would make. Something real and specific: this invoice gets held, this substitution gets approved, this request gets escalated.
• Ask who signs it. Not who built it or who monitors the dashboard. If it went wrong and someone senior asked why, whose name is the answer? If you get a function rather than a person, you have found your signature gap.
• Write the threshold as a number. Above what value, what risk score, what degree of clinical ambiguity does this come back to a person? It has to be specific enough to argue about.
• Check you could reconstruct it. Take one decision made last week and try to show what was considered, what was decided and why. If that takes more than a few minutes, the accountability is not real yet, whatever the policy document says.
The permission is the project
Half a century passed between an autopilot that worked and an airline permitted to land with it, and nobody remembers that gap as a technology story, because it was not one. It was the time it took to settle who was responsible for what, clearly enough that a regulator, an insurer and a captain could all live with the answer. Enterprises are in the same gap now. It does not have to take fifty years; it mostly takes having the uncomfortable conversation before the demo rather than after it.
So the question for this week: pick the decision your operation is most ready to automate, and ask whose name goes against it. If the room goes quiet, you have found the real work. Tell me in the comments where your operation goes quiet.
Recent blogs
Secure your agents
We’d love to chat with you about how your team can secure and govern Ai agents everywhere
