AI capabilities aren't overestimated; the way firms expect to convert them is. The parameter arms race benefits superstars, while profit remains in ordinary processes.

My second AI wave. The first one froze. The killer had two hands. The first was the rule bottleneck. The second was worse: the answers never added up to a world. Each rule worked only in its own context, and even a narrow domain needed a practically endless supply of questions. That's the real difference this time. The machine learns patterns from everything we've written. It arrives with a world, imperfect but present, instead of demanding one rule at a time.

The third systemic risk of the agent era, and the one almost nobody is pricing in: coupling. When competing firms deploy agents trained on similar data, with similar objectives, the market doesn't diversify. It synchronises.

The question your enterprise was built to answer is dead. The winning strategy is no longer "build and scale". It's "sense and respond".

This is a book about LLMs, AI agents, and AI integration into a firm, not about what happens inside the transformer. It is for the people accountable for a firm's integrity: the ones who'll sign the AI budget, own the operational risk, and answer for the consequences long after the vendor's demo team has moved on.