White paper
Estiba Limited · estiba.io
The audit count is not going away. What it cannot produce is the accuracy an operation runs on: where inventory error concentrates in multi-site retail, how much counting is enough, and what it costs.
Summary
Most multi-site retailers verify their stock once or twice a year, by counting everything. That count exists for audit and reporting, and it should stay. What it produces is a single number, months late, with no record of what was wrong or why, and no business can run its shelves on that.
Inventory error is not spread evenly across a catalogue. It concentrates in the lines that are handled most, which are the same lines a business can least afford to be missing. That makes uniform counting the wrong shape for the problem.
This paper sets out a different method: a selected set of lines counted every day, in twenty minutes before the doors open, by the staff already on shift, with each discrepancy explained and signed for before the session closes.
Twenty lines a day, six days a week, is more than six thousand line checks a year in a single shop. A full count, once or twice a year, gives each line one or two. That is the whole of the arithmetic, and the inputs are above it.
It includes the measured trade-off between how sensitive such a method should be and how often it will raise a false alarm, which is the question anyone adopting one has to answer and which is rarely published.
It requires no new hire, no integration project and no new hardware. Whether it displaces a stocktaking invoice you already pay depends on why that count is run, which is a question for you rather than a claim for us.
01 · The instrument
The annual or twice-yearly stocktake is usually run by a third party, overnight or on a Sunday, across the whole catalogue. It produces a variance figure and a corrected opening position.
Four things it cannot give you, by design.
The result is a business that knows its stock was wrong and cannot say what went wrong, when, or whether it has stopped.
The number is accurate on the night and decays from the following morning.
So “it was within tolerance” and “we have no idea” describe the same state of knowledge at the year end. The reassurance is produced by the instrument rather than found by it.
This is not a failure of execution. A full count is an expensive instrument used rarely, and rarity is what makes it blind to causes. Nothing about counting harder, or more carefully, or with a better contractor, changes any of the four.
None of which is an argument against the stocktake. It is an audit instrument and it answers an audit question: what the position was, on a date, to a standard somebody else sets. The error is asking it for operational accuracy as well, because accuracy on the shelf is a property of how a business runs rather than of how it reports, and a measurement taken twice a year cannot hold it.
And there is a question a reader can put to their own business in a minute. What share of your fastest-moving lines is not where the record says it is? Most businesses cannot answer it at all. That is not a gap in anybody's arithmetic. It is what an instrument used once a year leaves behind, and it is the reason for the method set out below.
02 · Concentration
A full count treats every line as equally likely to be wrong. It is not.
And this is established rather than asserted. DeHoratius and Raman examined nearly 370,000 inventory records across 37 stores of a single retailer and found 65% of them inaccurate, with the variation within a store explained by SKU-level factors including item cost, selling quantity and method of distribution. Research published by ECR across seven of Europe's largest retailers found about 60% of the SKUs analysed affected by record inaccuracy. The second finding matters more here than the first: the variables that predict inaccuracy overlap heavily with the variables this method selects on.
Much inventory error is produced by movement, and by more than one kind of it. A unit becomes mis-stated when somebody moves it, carries it, puts it back in the wrong place, takes it to a fitting room, transfers it between shops, books it in against the wrong code, or damages a label. A line that sits on a shelf for a season is handled once. A line that sells well is handled continuously.
So the probability that a line is where the record says it is falls as the line moves. The rate at which something sells is a proxy for how often it has been touched, and how often it has been touched is the mechanism by which the record and the shelf come apart.
Not every mechanism is reached by that proxy. Receiving errors, damaged labels and wrong master data do not scale with how fast a line sells, and the limits at the end of this paper say as much. The mechanisms that do scale with movement are the ones a weighted selection can reach, and they are enough for the method to work.
The conventional reason to prioritise a line is its value: being wrong about an expensive item costs more than being wrong about a cheap one.
The second reason is this one, and it is independent of the first: the lines that move fastest accumulate the most opportunities to go wrong.
They also happen to be the lines whose absence is most expensive, because an item that was going to sell this week is the one a customer is standing in front of you asking for.
Value and handling often point at the same shelf, though not always: a high-value line that barely moves is precisely where the two diverge, and the weighting exists so that it and a cheap line that moves constantly can both earn attention. That is the argument for a weighted selection, and it is the reason a random sample and a full count are both the wrong shape: one ignores the concentration, the other pays to cover ground where nothing is happening.
03 · The method
If error concentrates, then the counting should too.
What follows is a cycle count, and that is the honest place to start. Counting a subset of lines on a rolling basis rather than the whole catalogue at once is long established and well described. The claim here is not the subset. It is four things about it: which lines are selected and on what basis, that it runs every working day rather than periodically, that a difference cannot be closed without a stated cause, and that what results is an auditable record rather than an adjustment.
Each morning the system selects a set of lines and presents them as that day's work. The selection is weighted by what a line is worth, how fast it has been selling, and what its own history of discrepancies looks like. Lines that have been wrong before come round sooner. Lines that have been counted recently and were right come round later.
The session is designed to fit the twenty minutes a shop already has before it opens, and to be run on a phone by whoever is on shift. No specialist, no supervisor, no visit.
Twenty minutes is the target for a clean session. A session with differences takes longer, because explaining a difference is part of the work rather than an exception to it.
Twenty lines a day, six days a week, is more than six thousand line checks a year in one shop. A stocktake run once a year gives each line a single observation; run twice, two.
And those six thousand are not spread evenly. They are concentrated, by the weighting above, on the lines that are worth most, move fastest and have been wrong before. The lines a business can least afford to be wrong about are looked at repeatedly rather than annually, and the lines that sit still are looked at rarely, which is the correct allocation of a fixed twenty minutes.
The interval also decides how long a difference waits. If differences arise evenly through the year, the expected time between one occurring and being seen is half the gap between looks. Across 616 lines at twenty a day, six days a week, that gap averages thirty-six days, which puts the expected wait at about eighteen. A single annual count puts it at about 182.5. The weighting shortens it further on the lines carrying the most value and movement, and lengthens it on the lines that sit still, so these are the flat-allocation figures and the conservative ones for the lines a retailer cares about.
A session cannot be closed while a discrepancy is unexplained.
When the count and the record disagree, the person standing in front of the shelf is asked what they found, and the session stays open until each difference carries a reason and the name of the person who gave it. The reason is chosen by the operator, never suggested by the system, because a reason that was offered and accepted is not evidence of anything.
This is the whole difference between a system that can count and a system that counts.
Most inventory platforms have the capacity to record a stock adjustment, and most have a field for a reason. A field is optional. A gate is not, and the gate is what removes the need for anyone to be conscientious at seven in the morning.
Over a quarter this produces something a stocktake never produces: a dated series. The same line, counted repeatedly, with what was expected, what was found, where it was found, who looked, and why they thought it differed.
04 · Measurement
Any method that flags a line as behaving unusually has two dials. How far from normal a line must run before it is flagged, and how many consecutive periods it must do so. Set them loosely and the manager is handed a list nobody reads. Set them tightly and the thing you built the system for passes unnoticed.
Nobody can choose those values sensibly without knowing what each setting costs. That number is rarely published, so we measured it.
We generated books of synthetic lines in which the answer was known by construction: some built to be decaying, the rest built to be healthy, with realistic period-to-period variation. Because the truth was planted, every flag can be scored as correct or not. We then ran the rule across a range of settings.
This is a measurement of the detector, not a description of any retailer. No figure below says anything about how much stock goes missing anywhere.
| Outcome | Share |
|---|---|
| Planted decay caught | 83% |
| Healthy lines flagged | 3.7% |
The burden of that 3.7% depends on how many healthy lines there are. In a book where most lines are behaving normally, even a small false-positive rate produces more flags than findings. No precision figure appears here for that reason: precision depends on how much decay a real book actually contains, and that is exactly what we have not measured. It is also the clearest argument against shipping either setting as a default.
Requiring three consecutive periods at 1.35 times loses more detection than requiring two consecutive periods at 1.60 times. Waiting for more confirmation is the more expensive way to buy confidence, which is the opposite of what most people assume.
And the variation in the lines themselves dominates both dials. How lumpy ordinary trading is for a given business affects the outcome more than either setting. Which means the right values are not universal and should not be shipped as a default, they belong to the retailer, who is the only one who knows how lumpy their own weeks are.
It does not show that any particular retailer will catch 83% of anything. It shows what the rule does when the answer is known, which is the necessary first step and not the last one. The figures that matter in a shop can only come from a shop, and the method for producing them is the subject of the pilot described at the end.
05 · The record
The output of a stocktake is a corrected position. The output of this method is a corrected position and a record of how it got that way.
Each entry carries the line, the location it was expected in and the location it was found in, the quantity expected and the quantity counted, the reason the operator gave, who gave it, and when. Nothing is stored that the person at the shelf did not confirm.
Questions that a full count cannot answer, and that this one can:
The last of those is the one that changes a business.
A cause that surfaces can be dealt with; a correction that repeats silently is paid for every time it happens, and nobody ever learns the price.
And one distinction that is worth more than it sounds: the record separates found, in the wrong place from never found. Those are different problems with different remedies, and in a single annual variance figure they are indistinguishable. A business that can tell them apart aims its existing effort better.
Time to conclusion is recorded at session and shop level, because that is a property of the method and tells you whether it is working.
Nothing is recorded that ranks one member of staff against another. The signature on a count exists to say who looked, so that a question can be asked later. It exists to identify a witness, never a culprit, and the whole method depends on the people doing the counting having no reason to make the record anything other than true.
06 · Cost
Estiba is priced per location, per month, in bands by the number of lines in the stock file. There is no implementation fee, no professional-services line and no minimum term beyond a month.
Four things it does not add, each of which you can verify before speaking to us.
You already know what you pay a stocktaking firm each year, and how many locations it covers. We publish a price per location per month. The subtraction is one line and we are not going to do it for you, because a figure you calculate is worth more than a figure we assert.
What we will not do is put a number on what inventory error is costing you.
We would have to invent the inputs, and a paper that invents its inputs should not be trusted about anything else in it.
If you want that figure, the structure is simple and the inputs are yours: the proportion of your fastest-moving lines that are not where the record says, the margin you make on them, and how often a customer asks for one. And as above, most businesses cannot fill in the first of those at all.
07 · Limits
A paper that only lists capabilities is a brochure. These are the limits, stated so that nobody discovers them in month two.
It holds no view about what will sell, and it will not tell you what to buy. Everything it reports is about what was observed.
It records what is present and what the person who looked believed had happened. It is not a security system and it does not replace one. What it can do is separate differences that were found in the wrong place from differences that were never found, which makes existing effort better aimed.
Nothing in the record ranks one person against another, and nothing is measured per person per line. This is a design decision rather than an omission, because a system that grades the people entering the data teaches them to manage the data.
though it will show you where it is wrong.
Estiba is built, tested against a catalogue of 3,339 lines, and ready to be run. The first pilots are in preparation, and the figures in this paper are from measurement of the method rather than from a client's shop. We would rather say that than let a reader assume otherwise.
08 · Trying it
We bring the demonstration; you bring nothing. No stock file, no system access, no preparation.
If it is not for you, say so and we will leave it there.
References
Nicole DeHoratius and Ananth Raman, Inventory Record Inaccuracy: An Empirical Analysis, Management Science 54(4), April 2008.
Yacine Rekik, Aris A. Syntetos and Christoph H. Glock, Measuring the Sales Impact of Improving Inventory Records, ECR, 2019.
Read next
Interactive walkthrough · 90 secondsOne line’s life
What happens to a single line between one stocktake and the next.
Essay · four partsWhat retail knows about its own stock
The case built from peer-reviewed studies rather than from our own figures.
Short answers · thirteen questionsQuestions we are asked
Including the ones that are hardest to answer.
The siteestiba.io
What the method is, who it is for, and how it runs in a shop.