Who absorbs the change
Physical AI has been installed on the American selling floor at every scale, from a single trial store to a chain-wide rollout. Read as a record rather than as a set of opinions, it sorts itself, and it sorts on one question: how much of the change does the customer have to absorb?
When is it cheaper to send a robot than to power a shelf?
Two ways to know what is on a shelf. A fixed sensor draws continuously, so its cost per observation grows with how long you wait between looks. A mobile platform pays a fixed energy price per trip no matter how rarely it goes. Move the interval and the answer changes sides. Both inputs are ranges in the review, so what you get is a crossover band rather than a number.
The same record, sorted by who has to absorb the change
Eighteen graded deployment records, grouped by how much the customer's own path through the store has to change. Pick a tier to see what reached scale and what was withdrawn. Note that the ordering is not by how advanced the technology is.
The workforce half: flat in aggregate, changing shape
The employment series does not show the collapse the format debate implies, and it does not show stability either. The total holds while the composition moves, which is a different problem and has a different answer.
| Series | Reading | Grade |
|---|---|---|
| US retail employment, June 2026 | 15,459,700, near its February 2020 level | verified |
| Retail's share of national employment | Fell from 10.19% to 9.72%: the aggregate held while the rest of the economy grew | verified |
| Cashier jobs projected to 2034 | 313,600 fewer | verified |
| Cashier openings a year | 542,600, alongside that decline | verified |
| The entry-level role | Changing shape rather than disappearing: about half a million people a year enter a role whose content is rewritten faster than a vendor-specific curriculum can be written, accredited and taught | author's judgement |
What would settle it
The review names what it could not locate. Each of these is a measurement someone already holds, and publishing any one of them would resolve a hypothesis this page can only show as a range:
| Measurement | Which question it decides |
|---|---|
| Orders per site per week for centralised robotic fulfilment | Operators say order density rather than machine capability separates a $2.585bn charge from a further build-out. This review did not locate a published density. |
| A published accreditation cycle time | The claim that curriculum is outrun by the systems it teaches rests on comparing two durations, and only one of them is published. |
| A published service life for the systems in the ledger | The same comparison needs the other duration. A curriculum built on competencies that run across vendors does not depend on it; one built on a named system does. |
Every figure on this page is read from The future of commerce (TR-2026-31) and nothing is added to it. The energy figures are computed in the review from published cell, panel and datasheet powers rather than measured on a floor, and the page says so wherever it shows one. Constraint vocabulary: the eight.