An anon company burned $500m on Claude

There is a story doing the rounds about a company that spent half a billion dollars on an enterprise AI rollout and could not point to a single line of P&L movement twelve months later. The number may have grown in the retelling. The pattern has not.

Here is how it happens. Procurement negotiates a fleet of licences. IT integrates. Legal writes the policy. A launch email goes out with a training video attached. Then everyone waits for the productivity.

What nobody funded is the part where work actually changes. Which tasks move to the model. Which tasks deliberately do not. What a role looks like afterwards, and who decides. That is not an IT line item. It is an operating model decision, and in most rollouts it is nobody's job.

So usage settles at the level of autocomplete. The licence utilisation dashboard looks respectable, the value line stays flat, and eighteen months later the spend is quietly reclassified as a learning experience.

The fix is not more tools and it is not more training. It is redesigning the work, one task at a time, and being able to say out loud which decisions the humans keep.