When the lawyer in the Avianca case handed the judge a brief stuffed with fake case law invented by ChatGPT, his defence came down to one simple idea: the tool got it wrong. The judge would have none of it, and he was right. You can delegate a task to a machine. You never delegate responsibility for what you sign.
This distinction is the most important of all, and it is also the easiest to forget in the comfort of daily use, when the tool helps a thousand times in a row without a hitch.
Two things we confuse too quickly
To delegate a task is to tell the tool: do this work. Draft this rough copy, sort this data, propose these options, summarise this document. Fine, AI excels at that, and you should use it widely for what it does well.
To delegate responsibility would be to say: decide in my place, and answer for it if it is wrong. And that, no machine can carry. It has nothing to lose, nothing to answer for, nothing that commits it. Responsibility does not transfer to what has no skin in the game. You can hand over the act, never the duty of having done it well. That duty stays with the person who signs, who sends, who decides.
The Avianca case became famous precisely because it settled this confusion in public, and harshly. But it plays out on a small scale, with no courtroom, every time a professional takes an AI answer as their own without checking it, and finds out too late that they alone had to answer for it.
The more a decision commits you, the cleaner the human part must be
The practical consequence is clear and comes down to one rule. The more a decision commits you, the more conscious the human part must be. Not necessarily slower, but clear.
On reworking an internal email, the stake is nil, let it go. On a contract clause, an amount, a fact you put to a client, a figure you present to an investor, checking is not optional, it is the counterpart of the signature. Taking something as your own means becoming responsible for it. So you check what must be checked, you go back to the source before you quote, and you keep your hand at the moment that commits you.
This rule has a virtuous side effect. It forces you to rank your uses by level of stake, and so to know, each time, what you are doing. It is a discipline of decision, not distrust of the tool.
What is left, precisely, to a leader
I repeat it to my students and to the leaders I work with, and it is almost reassuring. AI is a formidable operator and a poor person in charge. Handing it the work, yes, and more and more. Handing it the fact of being right, never.
Far from a constraint, this distinction marks out what still makes, and lastingly, the value of a person in an organisation. As the machine absorbs execution, what gains value is precisely what it cannot carry: owning a choice, answering for it to others, holding firm when things go wrong. Responsibility is not a burden the AI frees us from. It is the core of what is left to us, and that is where a leader should invest, not step back. A machine with nothing to lose will never really decide for someone who has everything to lose.