This weekend I hunted a tough spot. The climb went straight up over brush and standing water, and I was strapped to the tree the whole way.
That strap does almost nothing for you on the bottom step. Slip a foot off the ground and you step back down. Twenty feet up, it's the only thing that matters. A trauma center in West Michigan studied 193 hunters hurt in treestand falls and found injuries got worse once the fall passed twenty feet. Fewer than 3% of those hunters were wearing a harness.
AI works the same way.
A demo is the bottom step. It runs on test data, a few people see it, and when it slips nobody gets hurt. Production is the stand. Now it's answering real customers, texting real phone numbers, and touching real records. A fall from there costs money and trust, and sometimes it costs a lawyer... That's the fall from the stand.
Engineers learn this early in their careers, or at least we hope we do. Working in production means handling everything carefully and keeping backups and a rollback plan, because something will break. AI doesn't change that rule. It raises the height.
The Wisconsin DNR puts the hunting version plainly: stay connected "from the time you leave the ground to the time you get back down." The strap matters most at the top, and that's exactly why you clip in at the bottom. Nobody clips in mid-fall.
Here's what tied off looks like before an AI system goes live.
1. It can only reach what its job needs.
Its own login, with access limited to the records and tools for that one job, and nothing else. In IBM's 2025 breach study, 97% of the breached organizations that had an AI security incident said they lacked proper access controls for AI.
2. The rules live in the system, not in the prompt.
Telling an AI to ask for consent, honor an opt-out, or stay away from certain data is a request. It will follow it most of the time. A rule built into the system around it, one the AI can't get past, is a strap. Consent before it calls or texts, an opt-out that always works, and carrier registration for business texting all belong there.
3. Every change can be undone.
Backups of what it touches. A way to roll back a bad change, or switch it off, in minutes. Tested before you need it, not the day you do.
4. It leaves a record.
What it did and why, so when a customer, an auditor, or your own team asks, you can answer.
None of this shows in a demo, and this is the real work of running any automation in production. Deciding the tie-offs before the climb, while it's still cheap to do it. It's a leadership job, because the person who builds the demo is rarely the person who gets the call when it falls.
Before your next AI system goes live, ask one question: if this falls from the top, what catches it?
If you're about to put something in the stand, I'm glad to look at the tie-offs with you.
What's the highest thing your AI is touching today?