Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How a Warehouse Cut Picking Errors by 90 Percent
Transcript
- Lucas: If you order groceries online, there's roughly a one-in-two-hundred chance the warehouse picks the wrong item. That's the industry average. And for a big player like Ocado, handling hundreds of thousands of orders a week, that adds up to a lot of angry customers and a lot of wasted food. Luna: One in two hundred doesn't sound terrible until you do the math. For a 50-item order, that's a 22 percent chance of at least one mistake. Lucas: Exactly. And Ocado's board decided that was unacceptable. So in 2018, when they opened their newest automated fulfillment center outside London, they also launched a project to cut picking errors by 90 percent. Today we'll look at how they did it—specifically, the combination of slotting algorithms and pick-by-vision headsets that got them from one error in two hundred to one in two thousand. Luna: Let's start with the slotting part. What's dynamic slotting and how does it differ from what most warehouses do? Lucas: Most warehouses still slot by category—all the pasta together, all the soup together. The problem is that similar-looking products sit next to each other, which drives mis-picks. You reach for the tomato basil and grab the tomato oregano instead. Dynamic slotting breaks that rule. It uses an algorithm that places items based on how often they're ordered together, not their category. Luna: So the algorithm basically says, 'If you always buy milk and cereal together, put them close to each other in the pick path.' Lucas: Right. And it also considers error history. Items that get confused frequently get separated by at least 50 feet. The algorithm reassigns slots every night based on the day's order data. So the next morning, the layout has shifted slightly to minimize both travel time and the chance of grabbing the wrong thing. Luna: That's a lot of data. How long did it take them to tune that algorithm? Lucas: About six months of iteration. They started with a simple rule set: separate look-alikes, group co-ordered items, put heavy items low on the shelves. But the early version actually increased errors because the algorithm moved too many skus too often, confusing the pickers. They had to add a stability constraint: no more than 15 percent of slots change per night. Luna: So the pickers still had some consistency day to day. Smart. Lucas: By month three, dynamic slotting alone had cut errors by about 40 percent. But the big leap came when they added the hardware. Pick-by-vision headsets. Pickers wear a lightweight headset with a tiny screen that shows the exact location and quantity of the item to pick. It overlays a green box around the correct bin and confirms the pick with a gesture. Luna: Wait—so it's not using barcode scanning at all? Lucas: It does both. The headset scans a barcode on the shelf to confirm location, then the picker's gaze is tracked to confirm they looked at the correct bin. The system only logs the pick if the eye-tracking and barcode match. That double confirmation is what drove the error rate down. Luna: I imagine the early adoption was rough. Anyone who's worn a headset for eight hours knows it gets hot. Lucas: Yeah, comfort was a huge issue. The first headsets weighed about 400 grams and workers complained of neck strain. Ocado worked with the vendor to cut the weight to 280 grams. They also added an anti-glare coating because the warehouse lighting caused reflections that confused the eye tracker. It was a classic ops problem: the technology worked in the lab but failed in the real environment. Luna: So what was the actual error reduction? From the baseline of one in two hundred—where did they land? Lucas: After six months of full deployment, the combined system—dynamic slotting plus pick-by-vision—brought errors down to one in two thousand items. That's a 90 percent reduction. But here's the detail that matters: the remaining errors weren't random. They were concentrated in two scenarios. First, when a picker had to pick multiple units of the same item. The system logged one pick and they assumed it was done. And second, when the picker worked more than five hours straight without a break. Luna: That's fascinating. So the algorithm didn't just reduce errors—it revealed the root causes of the ones that persisted. Lucas: Exactly. They used that data to add a mandatory micro-break every 90 minutes—just three minutes to reset attention. And they changed the confirmation step for multi-unit picks so the headset requires a separate scan for each unit. That pushed the error rate even lower, to about one in 2,500 within another quarter. Luna: Let's talk about the cost. Those headsets aren't cheap. Did the savings justify it? Lucas: Absolutely. The warehouse was processing about 50,000 orders a week. At the old error rate, that meant 250 mis-picks per week. Each mis-pick cost roughly $12 in refunds, replacements, and customer service labor. That's $156,000 a year in direct costs, not counting lost repeat business. The headset system cost about $800 per unit, and they needed 200 units for the warehouse. So $160,000 upfront. It paid back in just over a year. Luna: And the slotting algorithm was mostly software, so marginal cost near zero. Lucas: Right. The bigger win was customer retention. Ocado estimated that a customer who received a wrong item was 30 percent less likely to order again within a month. By cutting errors, they kept more customers. That's the kind of ops metric that doesn't show up on a P&L immediately but compounds hugely over time. Luna: This is exactly why I love operations stories. It's not just tech—it's the interplay of human behavior, system design, and feedback loops. Lucas: And speaking of feedback loops, we run this show without any ads. That's a deliberate choice—we want these conversations to be useful, not interrupted. If you've found value in episodes like this one, and you want to support us keeping it that way, you can do so at buy me a coffee dot com slash fexingo. No pressure, just an option for those who appreciate the ad-free approach. Luna: Yeah, it's a small thing, but it lets us focus on the details we just talked about, like the $156,000 number and the 90-minute break insight. Lucas: Exactly. So back to Ocado: the real takeaway isn't the headset. It's the methodology. They treated error reduction as a system problem, not a training problem. They measured what was actually happening, introduced two complementary interventions, and then used the remaining errors to diagnose further improvements. Luna: Which is essentially the scientific method applied to warehouse operations. Lucas: Right. And they're now rolling this out to their other fulfillment centers. The second warehouse took only four months to hit the same error rate because the algorithm and the headset configuration were already tuned. That's the power of transferable process knowledge. Luna: So for someone listening who runs a warehouse or a distribution center, what's the one thing they should try on Monday? Lucas: I'd say start with an error audit. Don't just count errors—categorize them. Wrong item, wrong quantity, wrong location. If you find that 80 percent of errors are wrong-item picks, then look at what's next to what. You might find that moving a few high-conflict items 50 feet apart cuts errors by a third without any new technology. That's cheap, fast, and it builds the case for bigger investments down the road. Luna: And if you do decide to go high-tech, make sure the humans are part of the design process. Ocado's comfort improvements with the headset were driven by worker feedback. Lucas: That's the operational lesson that ties it all together. Technology amplifies good processes, but it doesn't replace understanding what's actually happening on the floor. The next time you place a grocery order and everything shows up correct, there's a good chance an algorithm and a headset played a role. But also a team of people who helped design a system that works for them. Luna: Always good to remember that behind every piece of optimization, there's someone who has to make it work in practice. Lucas: Alright, that's it for this episode. We'll be back next week with another operations deep dive.