Michael Israel, who runs technology for the Kraft Group, has one rule for AI across the company's stadium, paper mills, packaging plants and real estate: each project has to name the behavior it will change. His examples are a camera that checks whether factory workers are wearing their safety gear, and a system that tells stadium staff a bathroom line is getting long during a concert.
An alert is only as good as the person who answers it
Cameras that spot missing safety gear are not new, and neither is software that counts people in a line. What the rule recognizes is that spotting is now cheap and responding is the scarce part. Every alert costs somebody a few seconds and a walk across the floor, and one person has only so many of those in a shift.
Picture a warehouse that points cameras at its loading dock to catch people without a high-visibility vest. On a busy day the system might flag 150 cases. A shift manager who also has trucks to schedule can walk to perhaps a dozen. By the second week the phone buzzes in the manager's pocket and nobody looks at it. Set the same cameras to send three alerts a day, each naming a place and a time, and to put the rest in a weekly count for the team meeting. Now the manager trusts the alert and acts on it, and that is the behavior change.
Hospitals show how to fix it
The hospitals did not solve alarm fatigue with better sensors. They solved it by deciding who owns the settings. Many cut the noise by monitoring only the patients who needed it, setting limits for each patient, and checking the settings at the start of every shift. Boston Medical Center set its cardiac monitors to sound at crisis level instead of warning level, and its nurses reacted faster.
That is the lesson for any manager buying AI that watches something. The tuning is the work, and it never finishes. Someone has to own the threshold, look at the false alerts every week, and be allowed to turn alerts off.
The person who answers has a new job
Israel's team has started meeting with human resources so that job descriptions match the new tools. The stadium employee who receives the bathroom-line alert now has a duty the old job never had: get there within minutes. If that person is still rated only on the old duties, the new one is invisible. If they are rated on every alert, including the ones no person could have reached, they will learn to resent the system the way nurses learned to resent the beeping.
So the change reaches pay, training and how performance is judged, well before it reaches a dashboard. It fits what we argued earlier, that AI pays off when companies plan what staff do next.
Before any AI that watches something goes live, write three lines: who gets the alert, what they do in the next ten minutes, and how many alerts a day they can handle. If one of those lines is blank, the tool will produce a dashboard and nothing else.