Why AI-Powered Piece Picking Should Start with Data, Not Demonstrations
“The industry has become very good at demonstrating robots. It is still surprisingly bad at deciding whether to buy them.”
Every few months another AI-powered robotic picking system makes headlines. The demonstrations are impressive: robots identify unfamiliar parcels, adapt their grasp, and improve through increasingly sophisticated vision models. The technology is advancing quickly, and rightly attracts attention.
Yet after years spent on both sides of this decision — inside a sortation equipment group, and inside the robotics ventures trying to sell into one — I believe the industry’s biggest challenge is no longer whether AI can make robots smarter.
It is whether we are making smarter investment decisions.
Too many automation projects still begin with a demonstration rather than an understanding of the operation they are meant to improve. Vendors show what their robots can do. Operators need to know whether those capabilities solve the right problem, at the right cost, under their own conditions. Those are very different questions.
The companies that gain the most from AI-powered automation will not necessarily buy the smartest robots. They will make the smartest decisions about where those robots belong.
The labour problem is real — but urgency distorts judgement
Sorting, inducting and piece picking are physically demanding jobs, built on repetitive work, night shifts and seasonal peaks. Across Europe, operators compete for an increasingly scarce labour pool while wage inflation continues without solving the recruitment problem. And when shortages become acute, urgency starts replacing analysis. A compelling demonstration suddenly becomes sufficient evidence for an investment worth several hundred thousand euros. It should not.
AI has changed what robots can do
Recent advances have genuinely expanded robotic capability. Vision systems recognise a far wider range of objects, adaptive grasp planning handles greater variation, and models improve with exposure. But this creates a quiet misconception: as robots become technically capable of handling more items, organisations assume the investment case strengthens automatically. Technical capability and business value are not the same thing, and the gap between them is where automation programmes fail.
Robot readiness is a property of your data, not your robot
Most operators have never characterised their own parcel flows in the terms an AI-powered system actually cares about. Weight bands matter, but so do depth, surface material, how an item presents at the pick face, and how the mix shifts between seasons.
During a validation programme I oversaw at Beam, BEUMER Group’s venture arm, we ran seventy-six expert interviews across forty-four companies in eighteen markets — mostly carriers, postal operators and sortation equipment makers. The biggest surprise had nothing to do with robotics. Operators were confident they understood their own flows, and the data repeatedly said otherwise: facilities running well below theoretical capacity without knowing it, and volume moving through express handling that nobody had accounted for. The data existed. It was generated, and then never interrogated.
Against that backdrop, a proper assessment tends to surprise the people who commissioned it. In one recent parcel hub assessment, more than ninety-nine per cent of items fell inside the handling envelope of commercially available systems. The capability question had been answered before the project began, and was never the interesting one.
The interesting question is rate. At that same hub, the sharpest hourly peak did not fall at Christmas. It fell in an ordinary week, roughly fifteen per cent above anything the peak season produced, because that volume arrived compressed into a narrower window. A fleet sized on peak-season daily averages would have been correctly sized on paper and undersized in practice, in February.
Simulation matters more in the AI era, not less
AI increases the number of credible solutions. As capability converges across vendors, demonstrations become steadily less useful for comparison — every system looks strong on its own curated material. Simulation becomes the only objective instrument left, letting operators test competing configurations against their own data before a hardware choice introduces commercial bias.
Used properly, it produces answers you would not have guessed. In one venture we simulated a carrier’s real volumes across four sortation sites, expecting to confirm that meaningful gains required shifting large quantities. The opposite held. The useful effects appeared at volumes far smaller than anyone in the room had assumed, which changed the entire shape of the business case. No amount of vendor discussion would have surfaced that. Only running our own numbers did.
The business case lives in the exceptions
Even where an AI-powered system handles ninety-nine per cent of parcels, the remainder often determines the economics. Oversized items, damaged packaging, unreadable labels and audit exceptions still require a person. Labour savings do not accrue proportionally to volume automated. They accrue when a shift can genuinely be staffed differently — and half an operator is not a saving. The exception path deserves as much design attention as the picking cell itself, and usually receives almost none.
Better AI still requires better decisions
Three questions are worth answering before any proof of concept begins. What does our item mix actually look like, in the terms a robot cares about? What is our true limiting hourly throughput, and in which week does it occur? And what happens to the parcels the system cannot process?
There is a fourth point, about what a good analysis is permitted to conclude. That four-site venture ended with my own recommendation not to proceed. The problem was real and the approach worked, but the enabling conditions were not yet in place. It was an uncomfortable answer and the correct one, and it saved a great deal of money that would otherwise have been spent discovering it slowly.
An evaluation that can only return yes is not an evaluation. AI will keep making robotic piece picking more capable, and the distance between what these systems can do and what any given operation should buy will widen rather than close. The operators who come out ahead will not be the ones with the smartest robots. They will be the ones who could tell the difference.
About Jesper Bang-Olsen
Jesper Bang-Olsen is Senior Director and Head of Venture Development at Beam, the venture arm of BEUMER Group, where he builds and validates new businesses at the intersection of logistics technology and automation. He is also Growth and Venture Partner at www.Logibot.eu, which develops hardware-agnostic software for evaluating, training and deploying robotic systems in parcel and warehouse environments. His background spans the postal, parcel and airport sectors, covering operations, commercial strategy and technology adoption, and he is an INSEAD alumnus. He works with European carriers and logistics operators on structuring data-led business cases for automation.
