A few years back, I led an ERP overhaul at a mid-sized automotive components plant. The vendor sold us on AI-driven demand forecasting that would cut inventory costs by 20 percent. I was skeptical but optimistic. The demo looked great, and the board wanted quick wins. So we went live with the AI module in the first phase, right alongside the core financials.
For the first month, everything seemed fine. The system was churning through purchase orders and suggesting reorder points. Then one Tuesday I got a call from the warehouse manager. He said we had nine months' worth of a particular alloy steel sitting in the yard, and more trucks were arriving. I thought he was exaggerating. He wasn't.
We traced it back to the AI. The forecasting model had been trained on three years of our purchase history, but nobody had cleaned the outliers. Two summers prior, we'd taken on a one-off contract for a heavy-machinery client that required a huge volume of that same alloy steel. The AI interpreted that spike as a seasonal pattern and ramped up orders the following year. It didn't know the contract was a fluke. The vendor's data scientist later called it data leakage—the model learned patterns that weren't actually repeatable. I should have caught that during data preparation, but I was too focused on integration timelines and not enough on what we were feeding the beast.
That mistake taught me that AI in ERP isn't a set-it-and-forget-it tool. You need a human in the loop, especially for high-value procurement decisions. We ended up adding approval thresholds: any purchase recommendation above a certain dollar amount had to be reviewed by a buyer. That simple guardrail saved us from a million-dollar inventory bloat. The AI still helps with routine, low-risk items, but the big decisions stay with people who understand the business context.
Now I see AI modules in nearly every ERP pitch I get from vendors. Predictive maintenance, quality anomaly detection, dynamic scheduling. The technology is genuinely useful, but the hype often outruns the data readiness of most factory environments. I've walked into plants where the bill of materials hasn't been updated in a decade, and the vendor is promising real-time AI optimization. It doesn't work that way. The model is only as good as the historical data you feed it, and most manufacturers have messy, inconsistent records scattered across legacy systems.
So if you're considering an AI-powered ERP, don't start with the algorithm. Start with a data audit. Figure out what's missing, what's wrong, and what's a one-time anomaly. Then decide if you still want the AI. In our case, after cleaning the data, the forecasting improved dramatically. But I'll never forget watching those trucks roll in, each one carrying steel we didn't need, and thinking: the AI didn't fail us. We failed the AI by not teaching it the difference between a pattern and a fluke.