Unexpected item in the bagging area.
If you've used a self-checkout, you've heard it. Usually right after you did exactly what the screen told you to do.
That's when the attendant walks over. She taps her badge on the screen, clears the error, keys in the code for the loose avocados you couldn't find, approves your coupon with the ripped barcode, and is back at her podium before the next machine starts beeping.
Self-checkout was supposed to replace the cashier. Instead it showed us how much the cashier had been handling all along. The produce codes they knew by heart. The coupon that wouldn't scan. The customer who set a purse down on the scale. None of that was in the job description.
Humans compensate for bad systems. AI tends to expose them.
I thought about that attendant this week while reading Fortune's coverage of its AIQ Summit, where chief people officers from HPE, Palo Alto Networks, and Lennar talked about what AI is doing to the way work gets organized. The article cited ServiceNow research showing that only 9 percent of organizations have made meaningful progress building complex autonomous workflows. That number didn't surprise me as much as I expected.
A lot of companies are turning AI on and hoping for the best, as if the results will just appear. But plenty of workflows only work because people have been making them work through brute force. They know which spreadsheet is current. They know the procedure hasn't been updated since the system changed. They know Mary in Accounting has the answer when something doesn't match, even though nobody can quite explain how Mary knows.
AI isn't that forgiving. Hand it inconsistent data, unclear standards, or undocumented exceptions and it won't work around them the way people do. It amplifies them. A person looks at a strange result and thinks, "That can't be right." An automated workflow moves it to the next step.
AI doesn't just automate workflows. It audits them.
HPE's Stacy Dillow made a similar point at the summit, urging leaders to rethink the work itself instead of retrofitting AI into the structures they already have. Palo Alto Networks' Danielle Gonzalez said they're hiring for curiosity and for people willing to unlearn what they thought was true. I contend our processes need the same unlearning.
AI implementation isn't the same thing as AI transformation. Implementation is turning the tool on. Transformation takes intentional workflow design, clear data and process standards, governance, change management, and a way to measure whether any of it is working. Then it takes follow-through.
The good news is that the problems AI exposes were already there. The people doing the work were just so good at handling them that nobody noticed.
So before handing a process to AI, go find the people who keep walking over to fix it. Ask what they check that isn't in the procedure, and which steps they'd skip if anyone let them. Sometimes the answer will be automation. Sometimes AI takes one part and a person keeps the rest. And sometimes a few steps should disappear before anything gets automated at all.
Otherwise all we've built is a faster way to hear "unexpected item in the bagging area."