Why Your Stock Figure Is Wrong, and What Actually Fixes It

The system says 400. The storeman climbs the rack, counts it twice because the first count felt wrong, and comes back down with 260. Somebody says the system is broken. Somebody else says the storeman cannot count. By that afternoon a purchase order has gone out for 140 pieces, and three weeks later those 140 arrive and stand next to the 260 that were always there and the 90 that turn up in October behind a pallet of something else.

Nobody in that story is lying and nobody is lazy. The system is a ledger. It reports the transactions it was given, in the order it was given them, and it holds no opinion at all about the ones it never received. The 400 is not a statement about the rack. It is a statement about the paperwork. The gap between those two things is not mysterious, and it has causes with names.

The number is load-bearing, which is the whole problem

An inventory figure is not a report. It is an input to four other decisions, and each of them inherits its error. Purchasing buys against it, so an overstated figure becomes a stockout and an understated one becomes cash on a rack. Planning nets against it, so every material requirement calculation carries the same distortion, multiplied by the bill of material. Sales promises against it, which is how a customer is told Thursday for something the store cannot find on Wednesday. And finance values it, so the closing stock number, the cost of sales beneath it and the gross margin above it all rest on the same arithmetic — which is why the annual write-off lands in one month for errors accumulated across twelve.

The causes, and how each one announces itself

Inventory error is not one phenomenon with one fix. It is six or seven distinct failures producing a single symptom.

Issues to production that were never booked

The most common cause in any plant, by a distance. Material leaves the store because a machine is waiting, the supervisor takes it, and the issue document is raised later, or in a batch at the end of the shift, or not at all. The physical stock falls. The system stock does not.

The signature is one-directional: the system is always higher than the rack, and the gap is worst on the components used every day by everybody. It is not a costing problem, though it looks like one at month end. It is a store-control problem, and no amount of counting corrects it because the counting does not touch the cause.

Returns that came back without a document

Material goes to the floor, the job is cancelled or the quantity was over-issued, and the balance walks back to the store in someone's hands. There is no return document because nobody thinks of a return as a transaction — it is just putting something back where it came from.

The signature is the inverse of the first: the rack is higher than the system, on items issued in bulk and part-used, often in a location the item is not supposed to live in.

Units of measure that do not agree with each other

You buy adhesive in drums and consume it in litres. You buy cable in kilograms and sell it in metres. You buy fasteners in boxes of 500 and issue them by the piece.

Every one of those conversions is a place for a factor to be wrong, and when it is wrong it is wrong permanently and invisibly. A drum booked as 200 litres when it holds 205 produces no variance anybody notices, only a drift that is absorbed at the annual count and starts again the next day. Cable factories meet the sharp version, because the conversion between kilograms and metres depends on cross-section, cores and stranding and usually lives in a spreadsheet somebody built years ago. The tell is an item that never counts correctly, with an error roughly proportional to throughput.

Scrap that never became a transaction

Material that entered production and was rejected is consumed material. If it goes in the skip without a scrap transaction, the system believes it is either still in the store or still in work in progress, depending on where the process stopped. Startup scrap, purge, offcut and trim are the same problem in a different costume, and in process plants they are large — a moulder purging a machine between a black run and a white one is consuming kilograms of resin against no order at all.

Stock that is not yours, and stock of yours that is not here

Material sitting at a subcontractor — at the galvaniser, the plater, the coater, the CNC shop — is still your stock. It is not in your building, it has not been consumed, and unless the system has a location that represents that subcontractor, it is either double-counted or missing entirely. Three tons of steel at the galvaniser with no record of which job they belong to is a stock error and a costing error at the same time.

The reverse: consignment stock from a supplier, and customer-owned material held for a job, are physically in your racks and are not your inventory. If they are booked in as ordinary stock, your balance sheet carries somebody else's goods. If they are not booked at all, nobody knows how much is there and the customer's material gets used on another customer's order.

Receipts against the wrong item, and stock allowed to go negative

Two look-alike codes and a receipt entered at speed leaves one item 200 over and its neighbour 200 under, netting to zero — which is why it survives every review that looks at total value. And a system that permits an issue of stock it does not hold will accept a physically impossible transaction and keep the arithmetic tidy while the record comes untethered from the world.

Cause	Direction of the error	Where it concentrates	What it is actually a symptom of
Unbooked issues to production	System higher than physical	Fast-moving components, consumables	Store control, not counting
Unbooked returns from the floor	Physical higher than system	Items issued in bulk, part-used	No document for putting something back
Unit-of-measure mismatch	Either, proportional to volume	Bulk material, cut goods, packs	An item master decision nobody revisits
Untransacted scrap and purge	System higher than physical	Process plants, changeover-heavy lines	Scrap treated as a fact of life, not an event
Material at subcontractors	System higher, or double-counted	Anything sent out for a process	A missing location in the model
Consignment and customer-owned	Either, and unpredictable	Distribution, moulding, fabrication	Ownership never modelled separately from custody
Wrong-item receipts	Net zero, gross large	Similar codes, similar descriptions	Item coding and receiving discipline
Negative balances	Meaningless in both directions	Everywhere, once permitted	A configuration decision taken by default

Why the annual wall-to-wall count corrects nothing

Once a year the plant stops, everyone counts, the differences are adjusted, and the auditors are satisfied. This is a valuation exercise. It is not an accuracy exercise, and treating it as one is why the same conversation happens again next year.

Three reasons it does not work.

It corrects the number and not the cause. An adjustment posts a difference and closes it. Nothing in the process asks why the difference existed, so every mechanism that produced it is still running on the Monday after.

Accuracy decays from the moment the count ends. If the underlying processes generate error at a steady rate, the count sets the clock back to zero and the drift resumes immediately. By the middle of the year you are as wrong as you were before, and you have no way of knowing it, because the only measurement happens in December.

And it is biased by what the counters can see. Counting every line in a building over two days, under pressure to finish, produces its own errors — and when the count sheet shows the system quantity, human beings converge on it. A count that is not blind is not a count; it is a confirmation.

Two definitions of accuracy, and the one that flatters you

Net value variance compares the total value counted with the total value on the books. It is the measure finance cares about and it is close to useless operationally, because errors cancel. A store that is 200 over on one item and 200 under on its neighbour reports perfect accuracy under this measure while being wrong twice.

Location-level count accuracy asks a harder question: for each item in each location counted, did the physical quantity match the system quantity within the tolerance set for that item? One item, one location, one pass or fail, no netting. This is the number that predicts whether a picker will find what the screen promised, and it is always lower than the finance measure — often dramatically so on the first honest measurement.

Measure both. Report the second one to operations. A plant that reports 99 per cent value accuracy and 72 per cent location accuracy is describing itself correctly and most people only hear the first figure, which is the same trap as a dashboard that computes a number perfectly from a stale basis.

Tolerance should be set per item class, written down, and never adjusted after the count to make the number look better. Zero tolerance on a bin of washers counted by weight is theatre; one per cent on a serialised device is negligence.

ABC, and the second axis most people leave out

Cycle counting means counting a subset every day, forever, with the frequency driven by how much the item matters. The classification everyone knows is by annual consumption value. That alone is not enough, because value is not the only reason an error hurts: a cheap gasket that stops a line has an error cost far above its stock value, and so does anything with a long lead time, anything under batch or expiry control, and anything a customer would notice.

Run the classification on two axes. Value gives you A, B and C. Criticality — line-stopping, long lead time, regulated, serialised, high theft risk — promotes items upward regardless of value. An item that is C by value and critical by consequence is counted as though it were A. This is a five-minute decision in a spreadsheet and it is the difference between a counting programme that protects the business and one that protects the balance sheet.

The loop that actually converges

Counting on a schedule does not improve accuracy on its own. What improves accuracy is what happens after a variance is found, and it has to be a loop with a named owner.

Count blind. The count sheet or the handheld shows the item and the location and no quantity. If the counter can see the expected figure, the count is contaminated.

Recount before adjusting. Every variance outside tolerance is counted a second time by a different person. A meaningful share of first-pass variances are count errors, and adjusting on the first pass injects new error into the record.

Adjust only with a reason code. No adjustment posts without a cause from a short, fixed list: unbooked issue, unbooked return, unit-of-measure error, scrap not recorded, wrong item received, damage, ownership error, unknown. "Unknown" must exist, and its share is itself a diagnostic — a plant where 60 per cent of adjustments are unknown does not have a process, it has a habit.

Review the reason codes weekly, not the variances. This is the step that is always skipped and it is the entire point. Nobody learns anything from a list of items that were wrong. Everybody learns something from a chart showing that 40 per cent of this month's adjustments were unbooked issues from one department. That chart names a process to fix, and fixing it removes a whole category of future variance rather than one line of it.

Count the failures again, sooner. An item that failed its count goes back into the schedule at a higher frequency until it passes twice consecutively. Errors cluster; the item that was wrong last month is the best predictor of the item that will be wrong this month.

The loop is dull, and its dullness is why it works. It also costs almost nothing: one or two people counting a defined list for the first hour of each shift, which is less total effort than a wall-to-wall count and spread across the year.

What honest looks like by item class

Targets belong per class, stated as location-level accuracy within a written tolerance. The shape below is what a disciplined plant can hold; the figures depend on your tolerance policy, so set the two together rather than copying either.

Class	Typical share of lines	Typical share of value	Count frequency	Tolerance	Accuracy expectation
A by value	10 to 20 per cent	70 to 80 per cent	Monthly, sometimes weekly	Tight, often zero on discrete units	High nineties
B by value	20 to 30 per cent	15 to 20 per cent	Quarterly	Small percentage	Mid nineties
C by value	50 to 70 per cent	Under 10 per cent	Twice a year	Percentage, or by weight	Low nineties
Serialised or batch-controlled	Varies	Varies	As A, regardless of value	Zero on the lot identity	Effectively total on identity
Critical spares and long lead time	Small	Small	As A	Tight	High nineties
Bulk consumables issued unmeasured	Large	Small	Measured by consumption reasonableness rather than counted	Not applicable	Not a counting problem

The last row matters. Some items should not be under a counting programme at all, because the cost of counting them exceeds any error they can produce. Gloves, fasteners in a bin, welding consumables — the control on these is a periodic reasonableness check against production volume, not a count. Putting them into the cycle plan is how a counting programme dies of its own weight.

The system settings that make the loop possible

Counting will not converge if the transactions underneath it can still lie. Four changes, none expensive, all of them arguments with somebody.

Block negative stock, everywhere, with no exception granted to any user. This is unpopular for about three weeks and then it is invisible, because it forces the receipt to be booked before the issue, which was always the right order.

Model every place material can physically be as a location, including the subcontractor's premises, the quality hold area, the returns bench and the goods-in cage before inspection. Material that exists somewhere the system cannot name will be wrong permanently.

Model ownership separately from custody, so consignment stock and customer-owned material sit in locations that are visible operationally and excluded from valuation. In a moulding plant this extends to the customer's tooling, which is custody with an amortisation agreement attached and usually recorded nowhere.

Fix the units of measure in the item master before anything else. The purchase unit, the stock unit, the consumption unit and the sales unit, with the conversion defined once and used everywhere. This is the least glamorous work in an implementation and it is where the accuracy of every downstream figure is decided, which is why it belongs at the front of the sequence rather than in the cleanup phase.

And make the scrap transaction as easy as the skip. If recording scrap takes four screens and finding the skip takes four steps, you already know which one wins.

What convergence actually looks like

A plant that starts this in January does not have accurate stock in February. What happens is more gradual: the reason-code chart stops being dominated by one cause, then by the next, and each disappearance is a process that got fixed. The volume of adjustments falls before the accuracy figure rises.

Somewhere in that period a subtler thing happens. People stop going to the rack to check before believing the screen. That behaviour — the physical walk that everybody in a low-accuracy warehouse does automatically and nobody records as a cost — is the real expense of a wrong stock figure, and it is the first thing to disappear when the number becomes trustworthy. Count the people doing it, multiply by the minutes, and you have the running cost of not fixing this.

The question of whether any of this needs a warehouse management system is separate, and the answer is more often no than vendors admit — that argument is set out in the honest test for whether you need a WMS. Buying one before the counting loop exists installs faster movement over the same wrong numbers.

If the counting has been tried before and drifted away twice, the missing piece is usually the weekly review of reason codes and an owner with authority over the departments causing the variances. That is a governance design, not a software one, and it is the first thing we look at on an ERP implementation where inventory is the reason the project exists.