I spent the better part of last year rolling an AI demand forecasting module into a mid-sized automotive parts manufacturer's ERP. Going in, I thought the hard part would be the algorithms. It wasn't. The hard part was convincing the production manager that the system wasn't trying to replace his thirty years of gut instinct.
We started with a pilot on one product line: brake calipers. The AI would look at historical orders, supplier lead times, machine availability, and even weather data to suggest weekly production volumes. The first month, its suggestions were off by about 18 percent. Not terrible, but the production manager, Dave, just pointed at a rainy week in April and said, 'Your fancy system didn't know that our biggest customer shuts down for plant maintenance every spring.' He was right. The AI had learned a pattern, but it was the wrong pattern.
That's when I realized we'd skipped the most important step: teaching the system what actually matters to the people on the floor. We added a feedback loop where Dave could flag bad forecasts and explain why. We fed in maintenance calendars, customer shutdown schedules, even local events that affected worker absenteeism. Within three months, the forecast error dropped to under 6 percent. But more importantly, Dave started using the system voluntarily. That's the real milestone no one puts in a requirements doc.
The same thing happened with quality control. Our ERP had a new AI module that could flag potential defects by analyzing sensor data from the CNC machines. It sent alerts about spindle vibration and tool wear. The first week, it flagged forty-seven 'anomalies.' Most were false alarms. The quality engineers started ignoring the alerts, which is the worst possible outcome. So we tuned the thresholds, learned which sensors mattered, and slowly rebuilt their trust. By the end, the system caught a bad batch of raw material before it hit the line, saving us maybe $60,000 in scrap.
What I've learned is that AI in manufacturing ERP isn't a magic box. It's a tool that needs context. It needs the people who've been running these machines to tell it when it's being stupid. And it needs a boring, unglamorous amount of data cleanup. I once found three different naming conventions for the same raw material across plants. The AI couldn't reconcile them, and neither could I, until we spent two weeks standardizing the master data.
If you're thinking about adding AI to your ERP, start small. Pick one process, one line, one problem. Get the data right. Bring the operators in early. The technology is the easy part; the human part is where it succeeds or fails.