Cycle Counting With AI: Getting Your Counts Right

Mistral released a model this week that it says outperforms frontier competitors at pinpointing objects within an image. Object localisation is an unglamorous capability with one very practical use for businesses holding stock: it is the thing counting from a photograph depends on. Which matters, because most small businesses still find out how wrong their inventory numbers were once a year, in a weekend nobody enjoys.

Why the annual count is the wrong instrument

A full count in March tells you your numbers were out by four percent somewhere across the previous twelve months. It does not tell you when, or on which items, or why, because receiving notes from July are filed, the person who moved the pallet has left, and nobody remembers the damaged carton.

What usually follows is a write-off and a resolution to be more careful. The discrepancy gets accounted for and the mechanism producing it stays exactly where it was, which is why the same variance appears next year at roughly the same size.

Cycle counting changes the timing rather than the effort. Count a slice continuously, so that every item is counted on a schedule, and a discrepancy surfaces within weeks of occurring, while the paperwork and the memories are still retrievable. The count becomes a detector rather than an annual reconciliation, which is the same distinction as finding a mechanism instead of a trigger in root cause analysis.

Count by value and movement

Counting everything equally wastes the budget on the long tail. Sort your items by annual value and by how often they move, then count the top group monthly, the middle quarterly and the remainder once a year.

For most businesses that puts roughly a fifth of the item list on a monthly cycle and keeps the daily count small enough that one person can do it before opening. A count that needs a dedicated afternoon will be skipped in a busy week, and a cycle count skipped is worse than no programme because the schedule now lies to you.

Where image counting helps and where it does not

Situation	How well counting from a photo works	What to do
Cartons on a pallet, single layer	Well, if the face is visible	Photograph, then confirm depth manually
Shelf face in a shop or store room	Well for distinct items	One photo per bay, consistent angle
Equipment or vehicles in a yard	Well, and quickly	Good use of the capability
Dense identical stacks	Poorly, and confidently	Count by weight or by container
Mixed bins and loose fasteners	Poorly	Scale-based counting instead

Occlusion is the hard limit. A photograph shows you the front of a stack, and the count depends on what sits behind it, so a model that is excellent at locating visible objects can still be badly wrong about quantity. Expect a fast estimate of the visible face rather than a number you can post to the ledger.

Dense identical stacks deserve particular caution, because the output looks the same whether the model counted carefully or approximated, which is the confidence problem that shows up wherever fluent output replaces visible effort.

Scan for identity, photograph for quantity

A barcode gives certainty about what an item is. A photograph gives a fast estimate of how many. Combining them plays to both: scan one unit to establish identity, count the face from an image, and set a variance threshold above which the location gets recounted by hand before anything is written to the system.

That threshold is the whole control. Without it, an approximate count becomes an authoritative number and your inventory record quietly gets worse while the counting gets faster. Five percent is a reasonable starting point, tightened on high-value items.

The count is the detector, not the fix

Discrepancies come from a short list of causes. Receiving errors. Damage and write-offs that never got recorded. Items picked and put back in the wrong place. Units that left without being deducted. Theft.

Counting more often fixes none of those directly. It tells you one of them is happening soon enough to work out which, and that only pays off if somebody looks at the variances. A cycle counting programme that records discrepancies into a report nobody opens has added work and changed nothing.

Give the variances an owner and a weekly ten minutes. Where the footage exists, searchable site and camera archives make the what-happened-here question answerable rather than theoretical, which is the use covered in video search for business.

What accurate counts unlock

Reorder automation, demand forecasting and anything that promises to optimise stock all assume your recorded quantities are close to true. Fed numbers that are four percent out in unknown places, they produce confident recommendations built on fiction, which is the dependency spelled out in AI inventory management for small business.

Counting is therefore the unglamorous prerequisite for the interesting projects, and it is worth doing in that order. The forecasting side is in demand forecasting with AI, and the wider question of moving goods without buying robots is in warehouse automation.

Starting next month

Pull your item list and sort by annual value. Put the top fifth on a monthly schedule and divide it across working days so the daily count takes fifteen minutes. Record every variance with the date, the item and the location. After a month, look at where the variances cluster, because they will cluster, and the cluster is your actual problem.

Add image counting after that, on one location type where the items are distinct and the face is visible, with a manual recount threshold from day one. Testing it against counts you already trust is the version of this that produces a decision rather than an impression, as in designing a proof of concept that decides.

Frequently Asked Questions

What is cycle counting and how is it different from a stocktake?

Cycle counting means counting a small part of your inventory continuously, so that every item gets counted on a schedule, instead of counting everything once a year. The practical difference is timing: a stocktake tells you in March how wrong your numbers were across the previous twelve months, while cycle counting catches a discrepancy within weeks of it happening, when the cause is still findable.

How do you decide what to count and how often?

Count by value and by movement rather than evenly. Items that are expensive or move constantly get counted monthly or more; slow, cheap stock can go a year. A common split counts the top-value fifth of your items monthly, the next layer quarterly and the long tail annually. The point of the schedule is that discrepancies on the items that matter surface fast.

Can AI count inventory from a photograph?

For some kinds of stock, increasingly well. Models have become notably better at locating and distinguishing objects within an image, which is the capability counting depends on. It works best where items are visually distinct, reasonably separated and consistently presented, such as boxes on a pallet, units on a shelf face or equipment in a yard. It works poorly on dense identical stacks, overlapping items and anything where the count depends on what is behind what you can see.

Does AI counting replace scanning or barcodes?

No, and treating it as a replacement is how these projects fail. A barcode gives you certainty about identity; a photograph gives you a fast estimate of quantity. The useful combination is scanning to confirm what an item is and image counting to speed up how many, with any variance above a threshold triggering a manual recount rather than being written straight into the system.

What causes inventory discrepancies in the first place?

Receiving errors, unrecorded damage and write-offs, items picked and returned to the wrong place, units sold without being deducted, and theft. Counting more often does not fix any of these; it tells you they are happening while the paperwork, the camera footage and people’s memories are still available. The count is the detector, and the fix is always upstream of it.