• A policy is not a control

    Where a client's data goes during an AI engagement, and how you would prove it. Intake, processing, and the part nobody writes about: what happens when the engagement ends.

  • Control limits, not thresholds

    Data observability alerts on thresholds, and thresholds either scream or say nothing. A pipeline is a process. Processes have control limits. Manufacturing settled this in the 1920s and wrote it down.

  • Everyone publishes the target

    Every analytics business case ends with a number. Almost nobody publishes what fraction of it landed. Two of mine, and the four things that decide the difference.