Turn the store into a fulfillment cell.

The incumbent format is not dead. CVS ended 2025 with 8,979 locations, filled 1.8 billion prescriptions, and produced $21.5 billion in front-store revenue.¹ It is used here as the reference format rather than the first customer: the drugstore and convenience footprint is the shape of store this system is aimed at.

Those two halves of the business matter separately. The front store — packaged essentials, no pharmacist involved — is the part a machine can plausibly serve. The prescription counter behind it is a regulated business carrying obligations a vending mechanism does not discharge, and treating them as one problem is the quickest way to build something that cannot be permitted. The next section draws that line before anything else.

The opportunity is narrower still: locations and hours where the cost of a staffed browsing store is larger than the service customers actually need.

Two errors surfaced in simulation that would otherwise have surfaced in hardware. A 0.32 m bottle gripped at its centre can never clear a 0.30 m tote rim, so tall items have to be gripped near the top and the arm plane raised for them. More seriously, the geometry as first drawn required 2.65 m of horizontal reach, while the largest collaborative arm on the market reaches about 1,750 mm.¹⁷ The design as drawn could not have been bought.

The layout is being reworked so the carousel presents totes closer to the arm and required reach falls under roughly 1.2 m — moving the specification out of a class that does not exist and into one that is widely stocked. That is the argument for simulating first, stated plainly: the cheapest place to find a fatal dimension is before a purchase order.

The model includes hardware, site work, financing, power, insurance, payment processing, packaging, remote monitoring, field maintenance, restocking, spoilage, depreciation, downtime, and inventory working capital. VenHub, the closest public benchmark, markets a $250,000–$400,000 system occupying 160–600 square feet with a claimed 1.5-minute pickup time.³

One finding shapes the whole model: throughput is unlikely to be the binding constraint. A cell fast enough to serve a site will spend most of the day idle, so cost per pick is set by how many orders the location actually generates, not by how quickly the arm moves. Below a certain daily volume a staffed counter is simply cheaper. That is why the beachhead is defined by hours that must be covered rather than by order count, and why utilisation — not cycle time — is the number the pilot has to answer.

The cost gate is expressed as installed cost deliberately. A 5 kg-class arm with the reach this layout needs is inexpensive and widely stocked, but controllers, end effectors, integration, commissioning and training routinely add 30–60% on top of the arm price.¹⁶ A quote that covers the arm alone understates the project by roughly a third, and that error is the most common way automation payback models flatter themselves. Leasing is modelled alongside outright purchase — comparable arms lease in the region of $800–$1,500 a month over three to five years, which changes the pre-revenue cash profile more than the sticker price does.²⁰

For scale: a human picker is generally benchmarked near 150 items per hour against roughly 600 for an automated cell.¹² Ocado's production robotic pick quotes a 630 units/hour theoretical maximum and still runs a 24/7 remote pilot to resolve exceptions.¹³ Any credible plan here budgets for human intervention rather than assuming it away.