The head of a platform team showed me a detection tool his group had built, and it was good work. Point it at the delivery pipeline and it would surface every aging item, every stage where the numbers had drifted, every queue growing faster than it drained. It ran across forty teams at once. It cost almost nothing to add the forty-first. He asked me what I thought the tool was worth, and I told him the honest answer, which is that the tool was worth a great deal and had also just relocated his hardest problem rather than solving it. He had automated the noticing. He had not automated the knowing what it means.

A director of operations at a mid-sized insurer walked me through her AI roadmap last spring. Fourteen initiatives, each with a sponsor, each with a slide. She was proud of it, and she had reason to be. The work was real. Somebody had thought hard about every line. Halfway down the list I stopped her and asked which of the fourteen she had decided not to do. She looked at the slide, then at me, and said nobody had asked her that. The roadmap had a column for what they were building. It had no column for what they had chosen to leave alone.

The question came from the CFO, near the end of a review that had gone well. If this system can read our whole delivery pipeline and find where the time goes, she said, why am I paying three analysts to do it by hand? It was a fair question, asked without malice. She had watched the tool surface in an afternoon what her team took weeks to assemble. She was doing the arithmetic any executive would do. Faster and cheaper, so why the people.