FeaturedLogisticsNews

Your inventory reads 95%. Your records, one by one, don’t.

By 15 de October de 2026No Comments

Your inventory reads 95%. Your records, one by one, don’t.

📅 October 2026
⏱️ 7 min read
✍️ Lorena Horowicz

Warehouse aisle with the figure: 65% of inventory records did not match the physical count

Ask for your operation’s inventory accuracy figure and you’ll probably get a reassuring number. The median across companies reporting to APQC’s benchmark sits at 95%. That number is an average, and your replenishment agent doesn’t decide on averages.

Two ways to measure the same thing, two different answers

Aggregate accuracy compares the total the system reports against the total the count produced. It’s a subtraction between two large numbers, and in that subtraction the deviations cancel: the item that read high covers for the item that read low. The result can look excellent even if no individual record is right.

The other way is to open the records one by one and ask, for each, whether the system’s number matches what was on the shelf. Almost nobody publishes that. The reference work is still Nicole DeHoratius and Ananth Raman’s, published in Management Science in 2008: they audited nearly 370,000 inventory records across 37 stores of a single retailer.

65% of those records did not match the physical count.

The two figures don’t contradict each other. They measure different things, and that’s worth stating before someone uses one to dismiss the other. What matters is which of the two describes the conditions an agent actually works in.

What changes when an agent decides

While a person makes the call, the error gets absorbed without anyone recording it. The buyer knows that item always reads high because of returns, that the count in that location hasn’t reconciled since the last refit, that the supplier ships in cases and the system books units. None of that is written down. It lives in three people’s heads.

An agent has none of that knowledge and no way to ask for it. It reads the record, takes it as true, and suggests. If the record was wrong, the suggestion is wrong.

And here is the real problem, which is about trust rather than precision. When a suggestion goes wrong because the source record was wrong, the operator doesn’t conclude that the record is wrong. They conclude the agent is useless. From then on every suggestion gets checked by hand. The agent keeps running, the licence keeps being paid, and a person keeps making the decision.

On paper the project is live. In practice it automated nothing.

Three operations, the same brake

A consumer goods distributor running four distribution centres launched a replenishment agent. Three weeks in, the purchasing team stopped looking at it: it over-ordered in two warehouses and under-ordered in the other two. The ERP counted cases and the WMS counted loose units, and nobody had decided which one wins.

A logistics operator moving refrigerated freight tried to automate truck assignment. Its TMS held eleven different spellings of the same customer, inherited from three migrations. The routing engine treated them as eleven destinations.

An auto parts manufacturer tested demand forecasting on 18 months of history. The history was complete, but they had changed their item coding without keeping the mapping. The model read a 100% collapse in 400 SKUs and the spontaneous appearance of 400 others.

In all three the investment was already made and the project stalled. In none of them was the model at fault.

How old the number is, and why it still holds

The study is from 2008 and it measured retail, not industrial logistics. RFID, cycle counting and modern WMS have spread since then, and it’s reasonable to assume the number would be better today. It’s equally reasonable to assume it wouldn’t be 100%.

What hasn’t changed is that almost nobody measures accuracy this way, and that the indicator companies do report doesn’t answer the question automation depends on. So the practical conclusion isn’t to trust a study from almost twenty years ago. It’s to measure your own.

Measure your automation ceiling this week

You can do this without buying anything and without pulling in IT.

Take your 20 fastest-moving items. Pull available stock for each one from the WMS and from the ERP, on the same day, at the same hour. Write both values into a sheet and count how many match exactly.

If 18 of 20 match, you have a base for automating replenishment decisions. If 11 match, any agent you install will be wrong in almost half the cases and the team will stop using it.

Run the same check on locations and on units of measure. Three spreadsheets, half a day of work, and you have a more useful diagnosis than any demo.

How Kovix handles it

In Kovix SCP, integration with WMS, TMS and ERP happens before the planning engine. It isn’t a later tuning phase: it’s the condition for the engine to start at all.

That means one definition of item, location and available stock, with conversion rules written down and versioned. When the agent suggests a purchase, the suggestion carries the detail of which records produced it and which system each one came from. If something looks off, you can see where.

The operational difference is simple: the team argues about the decision, not about the number.

Next steps

Run the 20-SKU check before you evaluate any AI tool. If the result is good, you’ll be able to ask far more of your vendor. If it’s bad, you just saved yourself a project that was going to stall.

Sources

DeHoratius, N. and Raman, A. (2008). Inventory Record Inaccuracy: An Empirical Analysis. Management Science 54(4), 627-641.
APQC, Open Standards Benchmarking: Inventory accuracy, median across 8,660 companies.

Let’s run the 20-SKU check together

30 minutes on your own data. You leave with the number and with what to fix first.

Book the 30 minutes

Leave a Reply