Three years ago, every Fortune 1000 stood up an AI Center of Excellence. Most are now quietly being restructured or wound down. The pattern of what's replacing them tells a clearer story about how AI gets done at scale than any analyst report — and it has more to do with platform engineering than with centers.

Two years of sales copilot rollouts have generated enough data to look past the vendor case studies. The lift is real. It is also more modest, more uneven, and more dependent on workflow design than the original pitch suggested. The question now is not whether to deploy — it's how to capture the lift that actually exists.

A new category of AI agent operates inside the browser, on behalf of a user, against the apps that already exist. It doesn't require API access, vendor cooperation, or a custom integration. It is the most boring-looking and most disruptive shift in enterprise software in years.

Most AI systems handle the easy 80% of cases well and the hard 20% badly. The decision that separates great deployments from frustrating ones is not how to make AI handle more — it's how to design the handoff to a human at the right moment, with the right context, before trust breaks.

The promise of AI was supposed to be that the models would compensate for messy data. The opposite has turned out to be true. Bad data fails louder, more visibly, and more publicly with AI than with anything that came before it — and most organizations still have not absorbed the implication.

AI literacy is now on every CHRO's roadmap. Most programs being rolled out are a mix of awareness videos, vendor demos, and prompt cheat sheets — and three months later, behavior hasn't changed. A small number of programs are producing measurable shifts in how people work. The difference is structural.