You plug in the AI tool. Intake forms get summarised automatically. Tickets route themselves. First drafts appear before anyone asks for one. For a fortnight it feels like magic — hours reappear on the calendar, admin quietly vanishes.
Then the first exception lands. A summary reads the situation wrong. A ticket lands in the wrong queue. A drafted reply goes out in a tone nobody would have chosen. Someone has to catch it, fix it, and work out what should have happened instead.
Nobody wrote that person’s name down anywhere.
The decision didn’t disappear
Here’s what’s actually happening. The tool didn’t remove the decision that used to sit with a person. It moved that decision further from view, and multiplied how often it shows up. When a person handled the work manually, they made a judgement call on every borderline case, in real time, and adjusted quietly as they went. When AI handles the work, the borderline cases still exist. They just surface as exceptions, dozens a week, with no one named as accountable for catching them.
Automation is very good at doing the defined part of a job faster. It isn’t good at deciding what to do when the job was never fully defined in the first place. Feed it an undocumented process, and it doesn’t tidy the process up. It runs the mess at scale, with more confidence than the mess deserves.
The exception wearing an automation
Picture a mid-sized distributor that automated its purchase-approval workflow, routing anything under a set value straight through and flagging the rest for a person to review. Sensible, on paper. Except nobody had ever agreed, in writing, what counted as a flag-worthy exception versus routine spend — so the tool inherited one person’s private sense of “normal,” running at a scale that person had never operated at before. Within a month, spend was moving faster, and nobody could say with confidence who was actually approving what.
Or picture a professional services firm that plugged an AI assistant into client intake to save its coordinators from repetitive admin. The tool drafted client-facing emails, competently, in a voice nobody had approved and with commitments nobody had signed off on. That wasn’t the tool’s fault. Nobody had ever written down what “ready to send” meant. A person doing that job had caught the gap by instinct. The tool had no instinct to catch it with.
Neither founder did anything reckless. Both moved early on something worth automating. But a tool can only run the operating model you hand it. If the decision rights and the definitions were never made explicit, automation doesn’t surface that gap gently. It runs straight through it, repeatedly, before anyone notices.
What actually closes the gap
Fixing this isn’t a technology problem. It’s an ownership problem wearing a technology costume. Three things close it.
First, map the decision before you automate the task. Take whatever process you’re about to hand to AI and write down every point where a person currently makes a judgement call, not just where they do the mechanical work. Decide, on paper, which of those judgement calls the tool can make on its own, which ones it should flag, and who the flag goes to. If you can’t answer that last part by name, the tool isn’t ready to switch on yet.
Second, treat the exception queue as the real job, not the overflow. Whoever reviews what the tool flags is doing the highest-judgement version of that role, not the leftover version. Staff it and evaluate it accordingly, rather than handing it quietly to whoever’s desk is closest and calling the rest “automated.”
Third, put a thirty-day audit on the calendar before launch, not after something goes wrong. Pull a sample of what the tool actually did, decided, or sent, and check it against what a person would have chosen. Adjust the rules, not just the settings. Most AI implementations that stall out weren’t badly built. They were switched on once and never checked again.
None of this slows the automation down. It’s what makes it hold up once it’s running real volume, not just the pilot. If you’re about to bring AI into how your operations actually run, or you already have and the exceptions are piling up faster than anyone expected, let’s talk it through: https://calendar.app.google/nJyTsHMsJJ5TnCzBA