Every operations manager I meet who runs an annual physical inventory hates it. Two days of shutdown, a rented crew of temps who have never seen your product, a variance number that shows up in December and surprises everyone, and a promise that next year you will start cycle counting.
Then next year arrives and the physical happens again.
The reason is almost never effort. Plenty of warehouses do count throughout the year. They just run a program that produces counts instead of accuracy, and after twelve months of that, nobody trusts it enough to cancel the shutdown.
The difference between counting and a counting program
Counting is when someone walks out with a tablet, counts a shelf, and types in a number. If it does not match, the system gets corrected and everyone moves on.
A program does three additional things. It decides in advance what gets counted and how often. It sets a threshold above which a variance triggers a look at the cause, not just an adjustment. And it measures itself with a number you could show an auditor.
Skip any one of those and you have counting.
Frequency: stratify, but not the way the textbook says
The standard advice is ABC by dollar value. A items get counted monthly, B quarterly, C annually. It is a fine starting point and it is not quite right, because dollar value is not the only thing that makes a part inaccurate.
I stratify on two axes. Value is one. Transaction velocity is the other. A $4 fastener that gets picked forty times a week touches more hands than a $900 assembly that moves twice a year, and hands are where accuracy goes to die.
A part that is high on either axis earns frequent counting. Something like:
- High value or high velocity: every 4 to 6 weeks
- Moderate on both: quarterly
- Low value and slow moving: once or twice a year
- Anything with an open variance history: back to the top of the list until it behaves
The other rule I hold to: every active location gets touched at least once a year, even the dead ones. Slow movers are where phantom inventory sits quietly for a decade.
Work out the math before you commit. Count the SKU-locations in each bucket, multiply by frequency, divide by working days. If it comes out to more counts per day than one person can do in two hours, you have designed a program that will be abandoned by March. Lengthen the intervals until the daily number is boring.
Count by location, not by item
If you tell someone to count item 44821, they will find the first place it lives, count it, and report a number. Item 44821 also sits in an overflow rack two aisles over and in a returns cage nobody thinks about.
Count location by location. Everything in bin C-14-3, whatever it turns out to be. That catches the two failures item counting never will: stock in a location the system does not know about, and stock the system says is there that is not.
It also makes the count blind by default. The counter is reporting what is in front of them rather than confirming a number the screen already showed them, and blind counts find things confirmation counts do not.
The variance threshold is the whole program
Here is where most programs quietly become useless. A count comes back off by six units, someone adjusts the system to match, and the loop closes. The number is now correct and the reason it was wrong is still running.
Set a threshold. It can be units, dollars, or a percentage, and it should be low enough that a few items a week cross it. Anything above the line does not get adjusted on the spot. It gets a recount by a different person first, because a meaningful share of variances are just miscounts, and then it gets ten minutes of someone asking why.
The causes repeat. Receiving put away a case quantity as an each. Two similar part numbers got swapped at the pick face. A kit consumed components that never got backflushed. Someone pulled a unit for a sample and told nobody. A return came back and got put on the shelf without being received.
Each of those has a fix that prevents the next fifty variances. The adjustment alone prevents nothing.
Measuring it in a way that survives an audit
The number most warehouses quote is net dollar variance, and it is the least useful one available. Being 40 units over on one part and 40 under on another nets to zero and tells you nothing except that you have two problems.
Count location accuracy instead. A location is accurate if the item, the quantity, and the unit of measure all match. Partial credit does not exist. Divide accurate locations by locations counted.
That number will be ugly the first month. Warehouses that believe they are at 98 percent routinely open at 82 when they start scoring this way, and that gap is the actual reason the annual physical still exists.
Track it weekly, by zone, and by counter. Zone-level detail tells you where your process breaks. Counter-level detail tells you who needs twenty minutes of training, which is a kinder finding than it sounds.
When you get to cancel the shutdown
Auditors do not accept cycle counting because you say you do it. They accept it when you can hand over a documented program (scope, frequency, thresholds, who counts, how variances get resolved), a full year of results showing every location was counted on schedule, and an accuracy trend that is both high and stable.
Most audit firms want to see sustained accuracy in the high nineties before they will sign off on dropping the wall-to-wall. Some will run their own sample against your numbers first. That is a good sign, not a bad one, because if your program works their sample will confirm it.
Realistically you are twelve to eighteen months out from that conversation on the day you start. That is not a reason to wait. It is a reason to start with the measurement in place, because a year of counting you cannot score is a year you will have to run again.
What this does not require
You do not need a WMS to do any of this. I have seen it run well on a barcode scanner and a spreadsheet, and I have seen it run badly inside a six figure system.
What software buys you is the scheduling and the audit trail, which matter a lot once you are counting hundreds of locations a month. What it does not buy you is the variance investigation, and that part is the program.