Roughly 73 percent of warehouse operators report difficulty filling frontline positions. Annual turnover in distribution centers exceeds 40 percent. Those two numbers get all the attention, and they deserve some of it.
But the number that quietly compounds is this one: a new hire needs two to five days of training before reaching minimum productivity. Combine 40 percent turnover with a five day ramp and a warehouse of 200 workers is running a permanent onboarding operation. Eighty new hires a year, each spending their first week below the line, each guided by tools that assume they already know the job.
That is the ramp problem. The labor market decides how many people you can hire. The ramp decides how much of their time you actually get.
The tools are the bottleneck. RF scanners and voice systems are reliable, but they were designed for experienced workers. They give one instruction at a time and the same instruction to everyone. A first day hire and a ten year veteran see identical prompts. The veteran ignores half of them. The new hire needs twice as many.
There is independent evidence that guidance built for the worker changes the numbers. A 2024 peer reviewed study (Epe et al., Logistics 2024) measured AR guided picking against RF scanning at a European logistics provider. In a controlled test, picks per hour rose 23.6 percent. In live daily operations at the same site, the gain was 10 percent. The study also compiled error averages across published research: 0.08 percent for vision guided picking against 0.39 percent for scan based picking. Those are the study's numbers, measured with generic pick by vision, not with APXR, and the live floor number is the one we model on.
The same research is blunt about why first generation AR picking stalled: ergonomics was the top adoption barrier, and overall worker acceptance landed at 77.6 percent. Heavy tethered battery packs, one sided weight, eye strain, and static workflows that treated every worker the same. The hardware and the software both ignored the person wearing them.
That reads like a failure. We read it as a to do list. Lightweight monocular devices purpose built for full shift picking. Guidance that adapts to skill level and language instead of repeating itself. Verification that blocks a wrong pick before it happens instead of flagging it after. And worker acceptance measured in every deployment, because a tool the floor rejects delivers nothing.
The ramp problem is solvable. Not by hiring faster, and not by training harder, but by putting the guidance inside the work itself. We will publish our own measured baseline versus APXR results when the pilot completes, whatever they show. Until then, the peer reviewed numbers above are the honest floor of what this category can do, and the ramp is where we aim.