What AI Still Can't Do in a Crisis
There is a great deal AI does well in normal operations. Spotting anomalies, summarising noise, drafting the first version of almost anything. A crisis is the moment that competence runs out, and it runs out quickly.
It has never seen this before
Models learn from what has already happened. A crisis is the part that hasn't. A shock nobody modelled, two systems failing at once, a supplier collapsing for a reason that was never in the data.
Pushed outside its training, a model does not stop. It produces something fluent and wrong, which is more dangerous than producing nothing. In a room where people are already short of certainty, a confident wrong answer travels fast.
AI is also a thing that can fail
It is worth saying plainly. AI now sits inside your operational estate, and it can be attacked like any other part of it. Poisoned training data, permissions granted to an agent that nobody reviewed, a model quietly changing a number that a report depends on.
That makes it a resilience question, not only a technology one. If an automated process sits inside a critical service, it belongs on the dependency map with everything else.
Machines can find the signal. Someone still has to own the decision, the explanation, and the consequences. In a crisis, that accountability does not transfer.
Discuss this scenario with Mitali Thapar
Confidential executive discussions for C-suite leadership and audit committees.