Your inventory accuracy is not one number

Someone asked you what your inventory accuracy is. You said 95%, or 80%, or "pretty good," and everyone nodded.

The number was meaningless — not because you were wrong, but because "inventory accuracy" is not one measurement. It is at least four, they disagree violently, and almost nobody says which one they mean.

The same store, four legitimate answers

Auburn University's RFID Lab worked through this in a whitepaper worth reading if you own this problem. They took one real store, on one day, and measured it four accepted ways:

• Total Magnitude: 73%
• Exact Match: 19%
• Sum of Absolute Variance: 65%
• SMAPE: 76%

Same shelves. Same day. Nineteen percent or seventy-six percent, depending entirely on the formula.

The reason is that the formulas ask different questions. Exact Match asks "is the count exactly right for this SKU?" — so it treats a SKU that is one unit off the same as one that is completely gone. Total Magnitude asks about total units across the building, which lets an overage in one SKU quietly cancel a shortage in another. Neither is wrong. They measure different things.

Auburn's own note on Exact Match is the useful one: it "assigns the same value to a SKU that is nearly out-of-stock and a SKU that is only missing one unit."

A benchmark you cannot reproduce is a number someone sold you.

The practical consequence is that teams set an accuracy target, hit it, and see nothing improve on the floor — because they optimised a formula rather than the health of the inventory.

What the research actually establishes

Set the benchmark question aside. Three findings are solid, and they matter more than any target.

Records drift, and by more than people assume. DeHoratius and Raman examined nearly 370,000 inventory records across 37 stores of one retailer and found 65% of them inaccurate (Management Science, 2008). That is peer-reviewed, with no vendor selling anything on the back of it. It is also one retailer, with fieldwork from the early 2000s — so treat it as evidence that drift is large and normal, not as your number.

Bad records roughly double your stockout rate. This is the causal finding, and it is the one to remember. In the FMCG out-of-stock study by Gruen and Corsten, items whose recorded on-hands matched reality had an out-of-stock rate of 4.1%. Items where the record was wrong ran at 8.9%. The industry average out-of-stock rate in that work was 8.3% — or, in their plainer phrasing, "for every 13 items one wants to buy, one will be out of stock." (Published 2007 by GMA/FMI/NACDS; the research was funded by a grant from Procter & Gamble, which you should know before quoting it.)

Trading-partner data is worse than anyone admits. Auburn's RFID Lab and GS1 US ran over a million items through five retailers and eight brand owners. Without RFID, 69% of orders shipped between brands and their retail partners contained data errors. The line in that study that should bother you more than the percentage: partners "commonly accept these process errors and attempt workarounds, which often create additional errors and costs."

Phantom inventory is the mechanism

Here is how a good business ends up with empty shelves and a system that says everything is fine.

The recorded on-hand quantity is higher than what is physically there. Because the system thinks stock exists, the reorder point never trips. Nothing gets ordered. The item runs out. Sales for that item fall — which the forecast then reads as reduced demand, so it orders even less next cycle.

Gruen and Corsten name this directly: "a self-fulfilling prophecy in demand forecasting." Their work attributes 47% of out-of-stock events to store forecasting — the single largest root cause, ahead of anything happening in the warehouse.

Two more of their findings are worth sitting with, because both contradict the instinctive fix:

• Store staff who manually override computer-assisted ordering tend to make it worse — "store personnel underperform even imperfect CAO demand forecasts."
• Improvements evaporate when attention moves on: "When the measurement stops, people go back to their old way of operating, and the OOS return to their previous levels."

Neither of those is solved by buying software. They are solved by making the correct action the easy action, which is a design problem.

The spreadsheet question, answered honestly

Every distributor gets told to get off spreadsheets. Usually by someone selling the alternative.

The defensible version of the argument comes from Ray Panko's academic work on spreadsheet error rates. In field audits of real organisational spreadsheets from 1997 onward, 91% contained errors. In controlled experiments across 998 subjects, 51% of spreadsheets contained errors — on models of only 25 to 50 cells.

The finding that should actually worry you is about confidence rather than correctness. Developers estimated a median 10% chance they had made an error. 86% had.

Panko names two failure modes that map exactly onto distribution. Hardwiring: someone types a number over a formula cell, correct for them that day, silently wrong for everyone afterward. Omission: the model never accounted for the factor at all, so no amount of checking the arithmetic will find it.

None of that means spreadsheets are unprofessional. It means they fail quietly, and quiet failure is expensive when the output is a purchase order.

If you distribute food, there is a deadline

The FDA's Food Traceability Rule under FSMA Section 204 requires Key Data Elements captured at Critical Tracking Events, traceability lot codes, and the ability to produce records to the FDA within 24 hours for foods on the Food Traceability List.

Enforcement will not begin before 20 July 2028 — Congress directed the FDA not to enforce before then, after the agency proposed extending the original January 2026 date.

Twenty-eight months is a comfortable runway and a bad reason to wait. A 24-hour lot-level traceability demand is not satisfiable from a spreadsheet at any error rate, and the businesses that treat 2028 as a build deadline rather than a compliance date get a working system out of it either way.

What to actually do

Pick one formula and write it down. Publish it internally. Any vendor benchmark quoted at you without a formula attached should be ignored.

Measure the drift, not just the total. Track the direction of your errors. Systematic overstatement is a phantom-inventory problem and will show up as stockouts. Symmetric noise is a counting-discipline problem. They have different fixes.

Count by velocity, not by calendar. ABC classification exists because counting your fastest movers weekly and your slow movers annually catches more error per hour spent than counting everything twice a year.

Fix the capture, not the count. Every recurring discrepancy traces to a moment where something moved and nothing recorded it. Receiving without a scan. A transfer done by memory. A credit never entered. The count is a symptom.

Then, and only then, consider software. Because a system fed by the same broken capture points will produce the same wrong numbers, faster and more expensively.

The honest note

Most of the statistics circulating in this corner of the internet do not survive being looked up. "The average retailer has 63% inventory accuracy" traces to nothing — it appears to be an inverted, rounded echo of the DeHoratius and Raman finding, passed between blogs until it acquired the authority of repetition. The widely-cited "NRF says shrink is 1.68%" refers to a survey the National Retail Federation stopped publishing after 2023. The "70% of ERP implementations fail" figure attributed to Panorama Consulting does not appear in Panorama's report; what the 2026 edition actually says is that more than a quarter of organisations went over budget, most often because of "the unexpected need for additional technology" — which Panorama attributes to poor system selection rather than bad luck.

We mention this because you are about to be sold something with numbers attached, and the fastest way to judge a vendor is to ask where a figure came from and watch what happens next.

ops.example.com/warehouselive

Picking

In transit

Delivered