The order that isn't a description
A purchase order is not a sentence about the world. It is a sentence that changes the world. When a category manager commits to eight thousand units of a toy line for week 51, the words on the system screen do not report a fact that was already true. Before the commit, no obligation existed. After it, a supplier is owed a call-off, a warehouse slot is reserved, a shelf plan is fixed. The click is a speech act in J. L. Austin's sense: not true or false, but happy or unhappy. It succeeds only if certain conditions hold at the moment it is issued, and it can misfire even when every word in it is correct.
Retail operations run on acts of exactly this kind: a purchase order, a price change pushed to point-of-sale terminals, a stock transfer between depots, a markdown instruction, a supplier call-off cancellation. Each is performative. Each depends on a state of the world being current at the instant of utterance, not merely plausible at the moment it was planned. The characteristic failure of the job is not that the category manager describes the demand curve wrongly. It is that the assortment is committed against a demand curve that has already moved, and the commitment is happy by form and unhappy by fact.
Austin's felicity, applied to a stockroom
Austin set out the theory in the 1955 William James Lectures at Harvard, published in 1962 as How to Do Things with Words. His target was the assumption that meaningful sentences are the kind you can check against the world for truth. Marriage vows, verdicts, bets and bequests fail that test and are still perfectly meaningful; they do something rather than say something. He replaced the truth test with felicity conditions: the speaker must hold the office, the form must be the conventionally accepted one, and the circumstances and consequent conduct must fit. Get any of these wrong and the act misfires — the world is unchanged, and the speaker is exposed.
A category manager's authority to commit stock is granted by the retailer's governance — a budget line, a delegated authority matrix, a supplier contract. That is the office. The purchase order form, the markdown code, the transfer instruction are the accepted conventional forms. What remains is the circumstantial condition Austin cared about least because it interested him least: is the state of affairs the act presupposes still true right now? A markdown instruction presupposes current stock cover, a live selling rate, and no competing action already taken by a rival or by the supply chain itself. When the presupposition has quietly expired, the instruction still executes — POS terminals apply it, the ledger updates — but it does something other than what was intended. That is a misfire wearing the clothes of a success.
What each generation of model would actually check
The lineage of models that might sit inside this loop differs in exactly what they can check before an act is issued.
| Generation | What it holds | What it misses in retail terms |
|---|---|---|
| Large Language Model | A corpus fixed at a training cutoff | No visibility into today's POS stream, today's inventory position, this morning's supplier notice; it can draft a flawless markdown memo with no idea whether the stock still exists to discount |
| Large World Model | A sensed scene — a store, a shift, a promotion window | Registers that the shelf is empty or the queue is long right now, but the warrant expires with the scene; it has no memory of what was already committed elsewhere in the network |
| Large Universe Model | Every stream still running — POS, inventory telemetry, supplier notices, demand signals — held as revisable beliefs with provenance and decay | Nothing structurally, on this axis; the open question becomes how many sources, how fresh, how well corroborated |
A Large Language Model can write the markdown instruction in the retailer's house style, cite the right SKUs, and use the correct approval language, while holding no fact about whether the demand curve behind it is still the one it was trained to expect. A Large World Model, watching a single store's footage or a single terminal's feed, can tell you the shelf is full and the aisle is quiet this afternoon — useful, immediate, and gone the moment attention moves elsewhere, with no record of the clearance already started two regions over. Neither is a smaller version of the same thing done badly. Each is missing a different felicity condition.
Two ways to read the category manager's job
Here the two defensible positions diverge, and the disagreement is not cosmetic.
The first position says: assortment and pricing decisions are performative acts, and the job's characteristic failure is a felicity failure, not a forecasting failure. The demand curve the plan was built against is a snapshot; the demand curve the market is running is continuous. POS streams, inventory telemetry, supplier despatch notices and competitor price signals are all evidence about whether the conditions presupposed by an already-drafted commitment still hold. On this reading, the fix is structural: intake has to be continuous, sourced, and revisable, because the thing being checked — current cover, current rate of sale, current competitor action — is itself continuous and revisable. A weekly planning cycle checked against a corpus, however well-curated, checks felicity once and calls it settled.
The second position says the opposite is closer to the truth: the category manager's authority and the supplier's obligations are fixed by contract and by the retailer's governance structure, not by how much telemetry anyone is watching. A call-off is binding because the framework agreement says so, not because a stream confirmed the shelf was empty. Enlarging the sensor surface does not create standing; it creates noise, false urgency, and a system that reacts to every wobble in a POS feed as though it were a commitment-grade fact.
The category manager's authority to commit stock was granted in a supplier contract negotiated eighteen months ago. No amount of live telemetry confers that authority. The limit on the act is legal and commercial, and no dashboard changes it.
Both positions are right about something the other position is right to insist on.
The objections that hold ground
The institutional objection is correct, and the record should not blur it. Standing is granted, never inferred from observation. A category manager can commit spend because the delegated authority matrix says so, at a threshold fixed by finance, within a contract signed by procurement. No stream, however comprehensive, makes someone a category manager. But granting the objection does not retire the intake problem — it relocates it. Authority answers who may commit. Felicity answers whether the commitment, made now, still matches a world that has not stood still since the authority was granted. The retailer's governance fixes the first question once. The demand curve reopens the second question every hour the store is trading.
The bullwhip objection also holds ground. Retail supply chains are notorious for amplifying noise: a small uptick at the till becomes an inflated reorder two tiers up the chain, because each layer reacts to a signal it cannot fully verify. Widening intake to every live stream — POS, inventory, supplier notices, scraped competitor pricing — multiplies the chances that a spoofed or distorted signal triggers a real commitment. A promotional spike caused by a data feed error, read as genuine demand, can generate a purchase order that is technically felicitous and substantively wrong. Freezing intake to a weekly, audited cycle is not laziness; it is a control against exactly this failure mode, and retailers who have been burned by phantom-demand reorders know its value.
The counter to both is that stale intake produces the classic Austinian misfire on schedule, not by accident. A markdown pushed against last week's sell-through when a supplier recall notice arrived Tuesday is not a subtle error; it is a commitment made on a condition already false. The remedy for the spoofing risk is not less intake but better-attested intake — provenance on each stream, corroboration across sources, and a decay function so that an unconfirmed spike loses weight before it can trigger a commitment. That is a property of the design, not an argument for freezing the picture.
Where the position narrows
The resolution is not that continuous intake wins outright. Governance still fixes who may act, and no volume of streaming telemetry substitutes for a signed contract or a delegated authority limit — the first objection stands undiminished. Nor does continuous intake eliminate misfire; it changes its type, from acting on stale conditions to acting on unverified ones, and that second failure is real and has cost retailers dearly in overreaction to noisy signals.
What narrows is the claim about where continuous intake earns its cost. For decisions where felicity conditions are few, stable and enumerable — a price floor, a supplier payment term, a return-to-vendor threshold — a rules engine fed by an audited, less frequent refresh is the right architecture, and building continuous multi-stream intake around it is waste dressed as rigour. The case for the third position sharpens precisely where conditions are numerous, fast-moving, and interact across a network no single rule table can hold: a national markdown cascading across a thousand stores while a competitor's clearance, a supplier recall and a weather-driven demand shift are all in motion at once. There, the category manager's real job is not forecasting a curve but continuously re-checking whether the commitment already made still matches the world it presupposes. That is a narrower claim than the one that opened the page, and it survives the objections that widened it.