Accountability is the missing piece of every AI rollout

Everyone rolling out AI asks the same first question: which tool? Almost nobody asks the question that actually decides whether the rollout survives: who is accountable when the machine is wrong?I run AI systems inside my restaurant. Every one of them is treated less like software and more like a new team member: real responsibilities, real supervision, and real consequences when it gets something wrong. That framing has saved me more than once.

AI fails like an unmanaged employee

Software fails loudly. It crashes, throws an error, stops. AI fails politely. It produces something confident, plausible, and wrong, and then keeps going. Nobody notices a polite failure until it has repeated itself for a month.

Early on, one of my systems drafted a reply that referenced the wrong person. Small mistake, easy to catch, except the whole point of automation is that you stop catching things. That incident bought two permanent rules in my operation: the system verifies the name before drafting, and no two systems edit the same thing at the same time. The mistake got a name, the name got a rule, and the rule got written down. That is accountability. Not a slogan. A changelog.

The rules my ten systems live under

Every system has one human owner. Not a committee. When the catering confirmation system misses a delivery detail, there is exactly one person who finds out first, and it is not the customer.

Drafts, not sends. Almost nothing goes out without a human approving it. The few exceptions earned that trust with months of clean drafts, and even those run behind a veto window: the message sits staged, a human gets notified, and silence for a set time is the only yes.

Kill switches, written down. Any system that touches a customer can be stopped by one action my managers already know. If you cannot turn it off in ten seconds, you do not own it. It owns you.

Thresholds decide upgrades, not enthusiasm. One of my systems earned the right to send without me because we measured the human process it replaced and the human process was missing too often. The data made the promotion decision. When a system stops earning its numbers, it gets demoted back to drafts.

Retros on a calendar. A few of my automations file a short self-review on schedule. Two of my earliest ones flunked enough consecutive reviews that I shut them down. Killing an automation that is not working is not failure. It is management.

The deeper accountability

There is a second layer under all the operational stuff, and it matters more.

My systems exist to buy back hours. The accountability question for the hours is mine, not the machine’s. If AI hands me back ten hours a week and I spend them on more busywork, the whole architecture is just a faster hamster wheel.

So the last rule is a question I have to answer, not a dashboard: who got the hours? Some weeks the honest answer is a spreadsheet did. The good weeks, it is a team member who needed to be seen, a note written in someone’s first language, a dinner where I was actually present. AI should automate the ordinary so humans can amplify the extraordinary. The machine is accountable for the ordinary. You are accountable for the extraordinary.

Start with one system, one owner, one kill switch. And if you want the questions that keep the hours pointed at people, they are numbered and free at jonharmeling.com/prompts.