A wall-to-wall count buys you one accurate day a year. A rolling count buys you an accurate week, every week, without closing the dock.
Key takeaways
You stop counting the building and start counting locations. Cycle counting is described in Microsoft’s warehouse management documentation as a warehouse process that you can use to audit on-hand inventory items: work is generated for one location at a time, a counter goes to it, and the result is reconciled. Because only a location is in play, the rest of the building carries on shipping.
The second half of the answer is a configuration detail that decides whether the theory survives contact with the dock, and it is worth quoting exactly. In Microsoft’s warehouse management documentation, when counting work is created, the on-hand inventory that is associated with this location isn’t blocked, and is therefore available for reservation and outbound processing, even though open counting work exists. That is the mechanism. If your system blocks the location, you do not have cycle counting — you have a series of small shutdowns.
The vendor documentation describes the process as three steps: create cycle counting work, process the cycle count, and resolve differences in the counted value.
Triggering is where the design decisions live. A threshold indicates the quantity or percentage limit of inventory items, and the system automatically creates cycle counting work when the threshold limit is reached. A plan creates cycle counting work either immediately or periodically through a batch job. And a worker can simply count what is in front of them: spot cycle counting — the worker counts items in a warehouse location at any time, even if no open cycle counting work exists for that location.
The threshold trigger is the underrated one for a small building. Counting a pick face when it drops below a set quantity means you are counting when the location is nearly empty, which takes seconds and produces the cleanest possible answer.
Count when the location is small
Microsoft’s documented example runs like this: there are 60 items in a location that has a cycle counting threshold of 40; during a sales order transaction, a worker picks 25 items from the location and puts them in a staging location, and because the new item count, 35, is less than the threshold quantity, the system automatically creates cycle counting work for the location. Counting a nearly-empty face takes a fraction of the time that counting a full one would, and the arithmetic is harder to get wrong.
Cadence should follow risk, not the alphabet. High-velocity pick faces, high-value items, anything lot- or serial-controlled, and any location that has produced a variance in the last quarter should come round more often than bulk reserve that has not been touched since spring. A plan can be set to control the maximum number of cycle counts that are created for items in different locations, and the number of days before a warehouse location is counted again, which is the lever for exactly this.
Two practical rules make the cadence stick. Count into the shift, not after it — a count at the end of a day competes with despatch and always loses. And count locations, not items: an item scattered across four locations produces four honest counts and one confusing total.
The count is not an internal hygiene exercise; several Canadian regimes assume you can produce a defensible inventory position on request.
For food, the Safe Food for Canadian Regulations require documents that identify the food by common name, lot code or other unique identifier, and the name and principal place of business of the person by whom or for whom the food was manufactured, prepared, produced, stored, packaged or labelled, kept for two years. And the response window is short: on request, traceability documents must be provided to the Canadian Food Inspection Agency within 24 hours of the CFIA making the request, in English or French, and if provided electronically, in a single file, in plain text, capable of being imported into and manipulated by standard commercial software.
For imported goods held under customs control the clock is longer. A bonded warehouse licensee must keep records for the period of six years after their removal, including records concerning inventory and the transactions that occur while the commercial goods are in the bonded warehouse. And the CBSA expects the physical building to support that record: a licensee must ensure goods are identified in such a manner so as to enable an officer to locate the goods and check them against the appropriate documentation.
And underneath everything, the tax record: books and records must be kept until the expiration of six years from the end of the last taxation year to which the records and books of account relate.
The single most common failure in cycle counting is treating a variance as an instruction to change the book. It is not; it is evidence that a transaction was wrong, and the transaction is what needs finding. That is why counted differences land in a review state rather than posting straight through — after a cycle count, any items that have differences in the counted value have a work status of Pending review.
Adjusting quietly and repeatedly is also a commercial risk beyond the warehouse. Buyers of businesses ask precisely this question — Treadstone’s sister law firm addresses how you know inventory on the books is actually still there and, less comfortably, what happens when an inventory count turns out to have been deliberately inflated.
A Saskatchewan 3PL runs a single shift, 4,000 pick locations and a food customer. It sets a plan so that every A-velocity pick face is counted at least every 30 days and every lot-controlled location at least every 14, and sets thresholds so that any pick face dropping below a quarter of its stocking limit generates a count immediately.
At 10:20 a counter is directed to a lot-controlled face. Picking against that location continues, because the on-hand isn’t blocked, and is therefore available for reservation and outbound processing, even though open counting work exists. The count comes back three cases light. Rather than posting an adjustment, the supervisor pulls the last four transactions against that location, finds a pick confirmed to the wrong lot the previous afternoon, corrects that transaction, and the location reconciles. The lot record stays intact — which matters, because the customer’s recall obligation runs on the lot code or other unique identifier.
The judgement in cycle counting is thin; the pattern recognition is not. Ranking locations for counting by their variance history, velocity and value rather than by a fixed rotation. Reading a counter’s recorded result against the last few transactions in that location and proposing which transaction is the likely culprit, so the supervisor investigates instead of searching. Drafting the variance narrative for the customer. Assembling a traceability extract in a single plain-text file when a customer or the regulator asks, since that is the form the rules ask for.
What it does not do is decide. Whether a variance is an error, a loss or a theft, whether to write inventory off, and whether a lot should be held are decisions a person makes and signs.
That is a question for your accountant and, if you are audited, your auditor — a rolling programme can support the year-end position, but the obligation to keep records adequate to determine your tax position is unchanged, and books must be kept until the expiration of six years from the end of the last taxation year to which the records and books of account relate.
Somebody other than the person who last picked that location, wherever headcount allows. The system supports either: work can be user directed, where the worker specifies a cycle counting work ID, or system directed, where the system assigns a cycle counting work ID to the worker.
It should be countable, not ignored. Counting workflows provide an Add LP or item button, which lets workers count on-hand inventory that the system doesn’t expect, though with a few limitations.
A 30-minute call is enough to tell you whether AI pays for itself in your back office.
