Latest / The Buyer's Side with Fexingo: Procurement, Sourcing, and B2B Purchasing Decisions / How a Buyer Used Demand Sensing to Cut Inventory 28 Percent
Transcript
- Lucas: A mid-sized electronics manufacturer I've been following — let's call them CircuitCore, about $800 million in annual revenue — cut their inventory by 28 percent last year. That's roughly $4.7 million in freed-up working capital. And they did it without hurting service levels. In fact, their fill rate actually improved. Luna: Okay, 28 percent is a big number. How? Most companies I talk to say inventory cuts come with a trade-off — you stock out more often. Lucas: That's exactly the conventional wisdom. And CircuitCore was living it. They had about 45 days of inventory on hand, but stockouts on high-margin components were still running around 8 percent. So they were carrying too much of the wrong stuff and not enough of the right stuff. Classic forecast-driven problem. Luna: So they switched to a different model? Lucas: Yes — demand sensing. Instead of relying on a monthly forecast generated by the sales team and then padded by the planners, they built a system that ingests real-time data streams: point of sale data from their top 20 distributors, weather data, even social media sentiment around product launches. Then a machine learning model generates a demand signal every single day, not once a month. Luna: Every day. That's a shift. What kind of data sources are we talking about? POS is standard enough, but weather? Lucas: Weather matters more than you'd think for their product line — they make components for outdoor telecom equipment. Cold snaps drive replacement orders. Heat waves delay installations. The model picked up a 12 percent demand lift when temperatures dropped below freezing in a given region. That was invisible in the old monthly forecast. Luna: Interesting. So the machine learning model essentially re-forecasts every night, and the buyers get a new recommended inventory target the next morning? Lucas: Exactly. And here's the key — those recommendations are based on a probabilistic range, not a single number. The model outputs a 50th, 80th, and 95th percentile demand estimate. The buyer chooses the service level they want to target. For critical components, they set the 95th percentile. For low-risk items, they set the 50th. That alone cut safety stock by about 35 percent. Luna: Because they stopped using a one-size-fits-all buffer. Makes sense. But rolling this out across the whole supply chain — that had to ruffle some feathers. Lucas: It did. The biggest pushback came from the procurement team itself. Some buyers had been managing the same categories for a decade. They trusted their gut. And the model told them they were ordering three weeks too early on certain capacitor lines. One buyer fought it for six months, until the model predicted a shortage that the buyer's intuition missed. They ended up air-freighting $80,000 worth of parts. After that, the resistance softened. Luna: Ouch. So implementation took time, but the results speak. What was the actual timeline from pilot to full rollout? Lucas: About 14 months. They started with one product line — their highest-volume connector family — and proved the concept. Took six months to clean the data, train the model, and validate it against historical orders. Then they expanded to all 12 product lines over the next eight months. By month 14, every category was running on the new system. Luna: And the inventory reduction — was that a one-time step-change, or do they keep improving? Lucas: The 28 percent came in the first year after full rollout. But they're still seeing month-over-month improvement. The model learns. As it gets more data, its demand predictions tighten. They expect another 5 to 7 percent reduction this year without any service level degradation. Luna: So the real lesson isn't just 'use AI'. It's about changing the planning cadence from monthly to daily, and matching safety stock to actual demand variability. Lucas: Exactly. And it's worth noting — the technology isn't exotic. They used an off-the-shelf machine learning platform, not a custom build. The hard part was the data integration and the change management. Luna: If these conversations have moved your work forward in some small way, you might consider buying us a coffee. Lucas: It's at buy me a coffee dot com slash fexingo. No pressure at all. Just a way to keep the show ad-free. Luna: Right. And now, back to demand sensing — one thing I'm curious about. How did they handle the supplier side? Did they share the demand signals with suppliers? Lucas: They did, selectively. For their top 10 suppliers, they provided a rolling 13-week demand view updated weekly. Those suppliers could then adjust their own production schedules. One supplier — a contract manufacturer in Mexico — cut their lead time from 12 weeks to 8 as a result. That created a virtuous cycle: shorter lead times meant CircuitCore needed even less safety stock. Luna: So demand sensing becomes a collaborative tool, not just an internal optimization. Lucas: Exactly. The buyer who owned that relationship said it transformed the conversation. Instead of haggling over price every quarter, they were talking about data quality and forecast accuracy. That's a much healthier dynamic. Luna: What about the cost of the system itself? Did the savings justify the investment? Lucas: Absolutely. The software subscription was about $120,000 a year. The data integration work cost another $200,000 upfront. Total investment under $500,000. Against $4.7 million in inventory reduction, plus the airfreight savings and fewer stockouts, the payback was under six months. Their CFO told me it was the best ROI project they'd run in five years. Luna: That's a strong endorsement. Any regrets? Things they'd do differently? Lucas: Two things. First, they wish they'd involved the suppliers earlier. Some of their best data — like distributor sell-through — came after they asked, not before. Second, they underestimated how much data cleansing was needed. POS data from different distributors came in different formats. It took three months just to normalize it. Next time, they'd budget more time for that phase. Luna: Good lessons. So for a procurement team listening today — what's the first step? Lucas: Start with one product line, one data source. Pick your highest-volume, most volatile category. Get daily POS data from your top distributor. Build a simple model in Excel or a cloud platform. Compare its weekly demand predictions to your existing forecast. The delta will tell you whether demand sensing is worth the investment. Nine times out of ten, it is. Luna: And if that delta is small? Lucas: Then you might not need demand sensing — your forecasting is already good. But for most companies, the forecast error is 30 to 50 percent for volatile items. That's a lot of wasted inventory. Luna: Worth checking. Lucas, thanks — this was a good one. Lucas: Thanks, Luna. Next episode we'll look at how a buyer used dynamic discounting to improve supplier payment terms without straining cash flow.