When I ran the cafe we printed two dockets for every order. One went to the kitchen. The other sat on a spike by the till and got binned at close, unread, every single day. It had been printing for years because the woman who set up the system wanted a paper backup for a till that crashed back in 2016. That till got replaced twice. The docket kept printing.
Most AI shops would look at that and speed it up. Print it faster, then email the owner a daily count of dockets nobody reads. Now you've got a useless step running at machine speed, and everyone claps because the dashboard is green.
Thats the whole problem with how this gets sold. Automate first, at speed, because speed demos well. All you've done is take a broken process and run it broken faster.
The order that actually works is slower at the front and it goes like this. Get inside the business first. Stand on the floor and watch the real thing happen, not the version in the manual. Then question every step until someone can tell you who wanted it and why. Not "we've always done it," an actual name and an actual reason. Delete anything that cant answer. Simplify whats left. Make it fast by hand. And only then, once a step has earned it, do you automate.
Automation comes last. It's the reward a process gets for surviving the other four.
Half the steps in most small businesses are that binned docket. Somebody set them up for a reason that stopped being true years ago, and they just kept running because nobody had the time to stop and ask. Point AI at that pile without looking and you don't fix anything. You just make the mess permanent and give it a login.
Here's what the good research actually says. A 2025 study by Erik Brynjolfsson and his colleagues, published in the Quarterly Journal of Economics, put a generative AI assistant in front of about 5,000 customer support agents. Productivity went up around 14% on average, and the least experienced workers jumped by roughly 34%. The tool made good people better at the part of the work that was already worth doing. Nobody got sacked. That's augmentation, and it only works when the underlying work is sound. Bolt the same tool onto a broken workflow and all it does is help you produce rubbish quicker.
So we never lead with the robot. First few days on a job I'm just watching, same as watching a new hand on the coffee machine to see where they waste movement. Which forms get filled in twice. There's always a report that gets built every Monday and opened by no one. And an approval step that exists because one customer complained in 2019 and the founder never took it back out. You'd be amazed how much just falls away once you ask out loud who it's for.
By the time we automate anything, half of what the owner thought needed automating is already gone. Deleted. The stuff that's left is lean enough that a machine can carry it without carrying the rot.
That docket, for the record. We didn't automate it. We turned the second copy off. Saved a roll of paper a week and one small daily annoyance, and it cost nothing but somebody finally asking why.
Get inside first. The automation can wait, and it'll be worth more when it comes.