As AI moves from drafting emails to making or shaping real decisions, a question tends to get skipped in the excitement: when the AI gets it wrong, who is accountable? It is not a philosophical question. It is an extremely practical one, and a business’s inability to answer it cleanly is usually a sign it has handed the machine more than it should.

The machine cannot be accountable

Start with the uncomfortable basic. An AI cannot be held responsible for anything. It has no stake, no consequences, no skin in the game. When it produces something confidently wrong, you cannot discipline it, it cannot apologise meaningfully, and it will do the same thing again tomorrow. Accountability is a human property. So if AI is involved in a decision that matters, the accountability did not disappear; it stayed with a person, whether or not anyone has actually named that person. The danger is the gap between “the AI decided” and “but a human is answerable,” because in that gap, real harm goes unowned.

“The computer said so” is not an answer

Every business deploying AI into consequential decisions needs to be able to finish this sentence: when this goes wrong, the person accountable is ____. If the honest answer is “nobody, really, the AI did it,” that is not a workflow, it is a liability with the safety off. A customer harmed by a confident error does not accept “the model made a mistake,” nor should they. Someone authorised the model to make that call without adequate checking, and that someone owns the outcome. Naming them in advance is not bureaucracy. It is the thing that forces sensible limits on where the AI is allowed to act unsupervised.

Accountability shapes the design

The useful part is that asking “who owns it when it’s wrong?” up front changes how you deploy AI, for the better.

• It draws the line. Decisions where you can tolerate the occasional confident error can be more automated. Decisions where you cannot need a human firmly in the loop. The accountability question tells you which is which.
• It mandates the check. If a named person is answerable for the output, they need a real ability to review it before it goes out — which means designing the checkpoint in, not bolting it on.
• It kills the worst pattern. “Let the AI handle it end to end and hope” does not survive the question. Hope is not an accountability model.

Keep a human where it matters

None of this is anti-AI. It is the condition for using AI responsibly on anything that counts. Let it draft, suggest, summarise and accelerate freely on the low-stakes work where a mistake is cheap and catchable. But for any decision where being wrong has real consequences, keep a named human accountable, with the means and the mandate to catch the error before it lands. The point of AI is to make that person faster and better-informed, not to relieve them of the responsibility.

So before you let AI make a call, answer the question plainly: who owns it when it’s wrong? If you cannot, you have not found a place to deploy AI. You have found a place you are not ready to.