The dangerous AI project is the one that works
The dangerous AI project is the one that works.
Failures are cheap. They end. The working one stays. The demo is real, the accuracy is good, the team is proud, and nothing in the business changes.
Picture a distributor that builds a model that classifies inbound supplier emails with 94% accuracy. It runs every day. The two people who used to sort those emails still sort them, because the model does not file anything and nobody redesigned the step after it. The project is a success on every slide and a cost on every P&L.
Technical success and economic relevance are separate tests, and the second one has to be written down before the first one starts.
The question is not "does it work." It is: when this works, what stops happening? Which hours disappear, which decision gets made faster, which error rate drops, and who is going to stop doing the old thing on the day the new thing ships.
If the answer is "we'll see how people use it," you are funding a science project with a production budget.
Kill the working one. It is harder, because it works. That is also why it will still be running, and still costing money, three years from now.