Every store runs on an inventory number, and almost every store knows the number is wrong. The system says there are eighteen units on the shelf; the shelf has eleven. The report says a SKU is in stock; the customer is staring at a gap. The count was accurate the day someone took it, and it has been drifting ever since: eroded by theft, miscounts, unlogged damage, deliveries that never got reconciled, and the simple fact that nobody can recount a thousand SKUs every night.
This is the quiet tax on retail operations. Reorders get made off a number nobody trusts. Out-of-stocks happen on items the system swears are stocked. And the single largest source of shrink (product that left without selling) hides inside the drift, indistinguishable from an honest miscount. An inventory you cannot trust is not just an accuracy problem. It is a loss-prevention blind spot.
The problem is not counting. It is staying current.
Stores already count. They cycle-count, they do annual stocktakes, they spot-check the high-value aisles. The problem has never been the count itself; it is that a count is a snapshot, and a store is a movie. The moment you finish counting, the number begins to decay, and by the time the next count comes around the drift has compounded into a figure that bears only a rough resemblance to reality.
What a store actually needs is not a better snapshot. It needs the movie: a count that updates continuously as product moves, so the number on the screen matches the product on the shelf at any given moment, not just on count day.
Watching the shelf, not guessing at it
The Inventory agent treats the shelf as something to observe directly rather than infer. Using the cameras already in the store, it watches stock levels on the shelf and the depletion of each facing over the course of the day. A cooler door that opens and a six-pack that leaves is a movement the agent sees. A facing that empties faster than usual is a velocity it tracks. A gap that opens on the shelf is a stockout it flags, the moment it happens, not when a customer points at it.
This is perpetual inventory that is actually perpetual, because it is grounded in what the cameras observe rather than in a count that is already going stale. The number on the screen tracks the shelf because it is reading the shelf.
The join that catches loss
Observing the shelf is half of it. The other half (and the part that makes the inventory trustworthy rather than merely current) is the join. The Inventory agent crosses what the cameras saw move against what the POS recorded as sold. When those two agree, the movement was a sale. When they disagree, the gap is loss.
This is the heart of why a single-stream inventory tool can never be trusted. A tool that only watches the shelf sees product leave and assumes it sold. A tool that only reads the POS sees what sold and assumes the rest is still there. Neither can tell a sale from a theft, so both quietly fold shrink into “movement.” Crossing the streams is what separates the two: four packs left the cooler, two rang at the register, and the gap is no longer a mystery to be discovered at stocktake. It is a flag, today, that feeds the Revenue Integrity Score.
From a trustworthy count to an automatic reorder
An inventory number is only worth as much as the decision it drives, and the decision that matters most is the reorder. When the count is current and trusted, reordering stops being guesswork. The agent knows real on-shelf levels, it knows true sell-through (sales, not movement inflated by shrink), and it knows the velocity of each SKU. From those, it drafts a reorder for every item running short.
The operator approves rather than authors. Instead of walking the aisles with a clipboard and a hunch, they review a proposed order built from what actually sold and what is actually on the shelf, adjust what they want, and send it. The work shifts from construction to confirmation, the recurring theme of an operating system. The agent also reconciles incoming deliveries against what was ordered and what arrived, so the drift does not start over the moment a truck pulls up.
Planograms, freshness, and the long tail
Because the agent is already watching every facing, it catches the slow problems too. A planogram that has drifted (product creeping into the wrong slot, a promotional end-cap that emptied and never got reset) surfaces as a deviation from the intended layout. Slow-moving SKUs that tie up shelf space without earning it become visible. The long tail of inventory problems that no team has time to police by hand gets policed automatically, as a byproduct of the same observation that powers the count.
One agent, one shared loop
None of this lives in a silo. The Inventory agent reads the same closed loop as the rest of Argus, which is why the Customer agent can answer “do you have this in stock” with a live, accurate yes or no. It is reading the same shelf data. It is why the forecasting layer can project demand off true sell-through instead of inflated movement. A trustworthy inventory number is not just useful to the people ordering stock; it makes every other agent in the store smarter.
That is the difference between an inventory app and an inventory agent on an operating system. One gives you a number to distrust. The other gives you a number you can act on, and then acts on it for you. Argus is in private beta with gas station, convenience, and grocery operators. If your reorders are running off a count nobody believes, talk to us.