Latest / The Buyer's Side with Fexingo: Procurement, Sourcing, and B2B Purchasing Decisions / How One Buyer Used AI Demand Forecasting to Cut Safety Stock 40 Percent
Transcript
- Lucas: So here's a number I keep coming back to: about forty percent of inventory in a typical manufacturing company is safety stock. That's buffer against uncertainty. And most of it is set using spreadsheets with a formula from the nineteen sixties. Luna: The fixed safety stock formula. I've seen it—base it on a multiple of standard deviation of demand, and then just leave it there forever. Lucas: Exactly. One buyer at a mid-sized electronics contract manufacturer decided that wasn't good enough anymore. They were carrying about twenty million dollars in safety stock across five hundred SKUs. That's capital sitting on shelves earning nothing. Luna: What did they do instead? Lucas: They built a machine learning demand forecast. Not a massive data science project—they had a team of two data engineers and one procurement analyst. They used a gradient-boosted tree model, trained on three years of point of sale data from their top customers. Luna: Gradient boosting—that's the same family of models that powers a lot of credit scoring and ad targeting. Why that over a neural network? Lucas: Interpretability and data efficiency. With only three years of history, a deep net would have overfit. Gradient boosting gave them feature importance scores—so they could show suppliers which variables actually drove demand. Things like weather, promotion calendars, even local sports schedules. Luna: Wait—sports schedules? Lucas: Yeah. One of their big retail customers saw a fifteen percent demand spike on game days for certain electronic accessories. Fans buying portable chargers and earbuds before heading to the stadium. The old forecast never captured that. Luna: So the model picked up on that pattern. Then what? They didn't just flip a switch, I'm guessing. Lucas: No, they ran a six-month parallel test. The model generated a recommended safety stock level each week, but they didn't act on it—they just compared what the model would have done against what actually happened. They tracked stockouts, excess, and write-offs. Luna: And the results? Lucas: The model would have reduced safety stock by forty percent on average—from twenty million down to twelve million—while actually improving service levels by about two percentage points. Fewer stockouts because the model anticipated demand shifts faster. Luna: So eight million dollars in freed working capital, plus better customer service. That's a pretty compelling case. Lucas: It is. But the implementation itself took another six months. They had to integrate the forecast output into their ERP system, and they had to retrain their buyers to trust the model instead of their gut. Luna: That trust piece is huge. How did they get buy-in? Lucas: They started with the lowest-risk SKUs—commodity components with stable demand. Once buyers saw the model predicting accurately week after week, they expanded to higher-variance items. And they built a simple dashboard that showed the model's confidence interval. If the interval was wide, the buyer could override. Luna: So human judgment still had a seat at the table. That's smart. Lucas: Exactly. The model wasn't a black box. They also had to renegotiate with suppliers. Some of them had been used to the old, higher safety stock levels—they were essentially holding extra inventory on consignment. When the buyer reduced buffer, some suppliers pushed back, worried about rush orders. Luna: How did that play out? Lucas: The procurement team shared the forecast data with key suppliers. They gave them read-only access to the model's output so suppliers could plan their own production. That actually strengthened the relationships—suppliers saw fewer emergency orders and could optimize their own lines. Luna: So transparency helped. Let me ask—what about the downside? Did they ever have a bad forecast? Lucas: Yes. During a chip shortage, the model failed because the supply constraint became the dominant factor, not demand. It predicted demand would drop because prices went up, but actually customers kept ordering because they couldn't get chips anywhere else. The model didn't account for that panic buying dynamic. Luna: Interesting. So the model trained on historical demand, but not on supply shocks. That's a classic blind spot. Lucas: Right. They had to add a manual override flag for supply-constrained periods. And they started feeding in external data—like lead times from suppliers and even news sentiment about key components. The next iteration of the model included that. Luna: So the system improved over time. What's the net result today, a year in? Lucas: They hit the forty percent reduction target across about seventy percent of their SKUs. The remaining thirty percent still have higher buffers because of volatility or low data volume. But overall, they freed up about eight million in working capital, and their inventory turns went from four to six. Luna: That's a real, measurable impact. And they did it without a huge tech stack. Lucas: Exactly. The total cost of the project was about three hundred thousand dollars—mostly data engineering time. The ROI was massive. And it's a model that procurement teams at smaller companies can replicate. You don't need a data science PhD. Luna: I want to ask about the human side. Did any buyers lose their jobs because the model made decisions? Lucas: No. In fact, their roles shifted. Instead of spending hours updating spreadsheets, buyers focused on supplier relationships, strategic sourcing, and the exception cases the model flagged. They became more valuable. Luna: That's the ideal outcome. And you know, it's stories like this that make me think about what we're doing here on the show. We spend a lot of time digging into these real cases—and it's only possible because listeners support us directly. Lucas: Yeah, that's a good point. We don't run ads, and we don't have a big corporate backer. It's just a handful of listeners who chip in through buy me a coffee dot com slash fexingo. That's literally what funds us making this many episodes. Luna: And we're grateful for it. If these conversations have moved your work forward in some small way, that's the whole point. Lucas: Alright. Back to the electronics buyer. One more thing I want to highlight—they didn't try to automate the entire replenishment process. They still let buyers place the orders. The model recommended a quantity, but the buyer had to hit 'submit.' That preserved accountability. Luna: Interesting. So they kept the human in the loop deliberately. Lucas: Yes. And it worked. No single catastrophic failure because someone trusted the model blindly. The question I'm left with is: how many other procurement teams are still using that nineteen-sixties formula, when a gradient-boosted tree could be running in a few weeks? Luna: Probably a lot. But hopefully this story nudges a few of them. Lucas: That's the idea. Thanks for listening.