The idea was great. A manual process was eating hours every week, so someone set up an automation. Maybe a workflow tool, maybe AI, maybe both. Form comes in, data gets moved, emails get sent, records get created. No more copying and pasting.
Then the problems started. Wrong emails going to the wrong people. Records created with missing information. Duplicates. A customer getting a reminder for something they’d already paid. And the errors are happening faster than anyone can catch them, because now nobody’s looking.
What the owner thinks
“The automation is broken.” Or, more broadly, “AI isn’t ready for a business like ours.”
Sometimes a tool is the problem. But in most cases I’d bet the automation did exactly what it was told. It just got told to copy a process that wasn’t ready to be copied.
What I’d investigate
I’d go back to how the process worked before it was automated, and ask the people who used to do it by hand:
- Was it done the same way every time? Or did each person do it a little differently?
- Were the inputs consistent? Did the information arrive complete and in the same format, or did someone clean it up as they went?
- What were the exceptions? What did people quietly handle that nobody wrote down? “Oh, if it’s that kind of customer, I’d skip the reminder.”
- Who owned it? If something went wrong, who noticed and fixed it?
Usually, the answers reveal a lot of invisible human work. People were catching errors, filling gaps and making judgment calls without anyone realizing it. The automation removed those people and kept the mess.
What’s underneath it
Automation copied an unstable process. When a person does a messy process, they absorb the mess. When software does it, the mess goes straight through, at speed, every time.
Automation makes a process faster. It doesn’t make it better.
What I’d change
- Pause or narrow it. Turn off the parts causing the most damage. Keep the parts that are working.
- Stabilize the manual version. Decide the one way it’s done. Make the inputs consistent, often by fixing the form or the place the information starts.
- Write down the exceptions. Every “if it’s this kind, do that” becomes a rule or gets routed to a person.
- Automate the boring, stable part. The steps that are the same every time. Leave judgment to people.
- Keep a human check. Especially at first. Someone reviews what the automation did, on a schedule, so errors get caught in days, not months.
A small example of what I mean: an automated reminder goes out to every customer with an open invoice. By hand, the office would have skipped the customer on a payment plan and the one who already called to say the check was in the mail. The automation doesn’t know either of those things. Nobody ever wrote them down, so the software never had a chance.
What I’d leave alone
Automations that are running stable steps well. Not every automation is a problem. If something has worked quietly for months, leave it alone.
The tradeoff
Going back to manual, even partly and even temporarily, feels like moving backward. Somebody spent time setting this up, and the business was excited about it. But a few weeks of stabilizing usually beats months of cleaning up errors that the automation keeps creating.
How I’d decide what to automate next time
A few questions before automating anything:
- Is the process done the same way every time?
- Are the inputs complete and consistent?
- Are the exceptions known and few?
- Is there someone who will notice when it goes wrong?
If the answer to any of those is no, fix that first. Then automate.
This is especially true with AI. AI is very good at doing a clear task quickly. It’s not good at guessing what your business meant when nobody decided.
So, one exercise: pick one automation in your business and ask the person who used to do it manually what they used to catch. Their answer will tell you whether it’s safe.
If you’re trying to decide what to automate and what to fix first, bring it to a Systems Fit Call.
