Don't ask AI to predict your next problem. Test for the last one. I went back through all nine months before my hotel's Super Bowl disaster. Every number was flat. Nothing in those numbers saw it coming. Pick one problem your team already lived through.

Rules shrink the holes in AI output. They don't close them. I asked AI for a dashboard with one line of prompt. It invented a guest, a refund and a passing grade. With eight rules, it still rewrote the data underneath and told me it hadn't. Check the data yourself.

Set your AI's alarm past your worst normal month, not on it. I asked AI to set an alarm for my business. It set it at exactly normal, so any month a hair worse sets it off. Cry wolf enough times, and nobody comes. Find your worst normal month.

Test your AI's alarms against your last disaster before you trust them. I had AI write tests for my hotel's last disaster, then run them on two years of history. Month 11, one went quiet. Nothing was fixed. It just stopped looking. Run it on the past first.

Your next AI policy is for other people's agents. Delta is blocking AI agents from booking. JPMorgan and Schwab block them from accounts. Yelp only lets in agents that pay. Decide who gets in.

Microsoft now treats AI agents as insiders. Right call. Each agent gets an identity, a risk score, and access that tightens as risk rises, enforced outside the model. Name an owner.