Picking accuracy has a denominator problem before it has a process problem
Two warehouses can both report ninety nine percent accuracy while one of them is four times worse than the other. The difference is the unit being counted, and until that unit is fixed, every improvement project is measuring noise.

The denominator decides what the number means
Picking accuracy is reported three ways in practice: by order, by line and by unit. The three are not interchangeable. A single wrong unit in a twelve line order fails the order measure completely and barely moves the unit measure, so the same event produces a one percent and a zero point one percent depending on the denominator.
Order level accuracy is the measure a merchant feels, because a failed order is a failed delivery and a support ticket. Line level accuracy is the measure a warehouse can manage, because a line is a task and a task has an owner. Unit level accuracy is the measure that matters for high value goods, where one missing piece is a real loss.
The correct practice is to publish one primary measure and two supporting measures with the conversion stated. A warehouse reporting ninety nine point five percent by order while quietly omitting that the average order has three lines is reporting something close to ninety eight point five percent by line, and the gap between those two numbers is where the refunds live.
Whatever the denominator, it has to be fixed before any improvement is claimed. Changing the measure mid year produces an improvement that is arithmetic rather than operational, and it is the most common way a picking project reports success while the complaint rate does not move.
Single order picking and batch picking fail differently
Single order picking walks the same aisle once per order and hands the goods to packing with no sorting step. Batch picking walks the aisle once for many orders and sorts at the end. Batch picking cuts travel dramatically and introduces the one failure mode that single picking cannot have: a unit placed in the wrong order's tote.
That failure mode is invisible at the pick face because the pick itself was correct. The unit was taken from the right location in the right quantity and scanned against the right task. It became an error ten metres later, when it went into the wrong compartment, and no scan at the pick face can catch it.
Which is why batch picking needs a scan at the put to order step rather than only at the pick step. The second scan is the only checkpoint that sees the sort. Operations that adopt batch picking to save travel and keep the single scan configuration are buying travel savings with accuracy.
The batch size itself is a variable worth managing. Larger batches amortise travel further and increase the sort error rate, because the tote grid gets denser and the compartment labels get smaller relative to the number of open orders. There is an optimum and it is usually lower than the maximum the system allows.
Where the scan sits changes what it can catch
A scan at the pick face confirms that the right item was taken from the right location. A scan at the packing bench confirms that the right item reached the right order. Both are called picking verification and they catch entirely different errors.
Pick face scanning is cheap and early. It catches a wrong location, a wrong SKU in a location, and a quantity shortfall while the picker is still in the aisle and can correct it in seconds. It cannot catch a unit that was swapped between totes, because at the pick face that unit was correct.
Packing bench scanning is later and more complete. It sees the whole order content and can compare it against the order line by line, which is the only way to catch a swap. The cost is that the correction now requires a trip back to the aisle or a substitution decision at the bench.
The configuration that catches the most errors per unit of labour is both, with the bench scan used as verification of content and the pick scan used as verification of location. Operations that can only afford one should choose the bench, because it catches the larger share of customer visible failures.
Location labels are a readable interface
A location label is not signage. It is the interface between a human and a database, and it is read hundreds of times a day under poor light by people who are mostly looking at the shelf rather than the label. A label that requires a second look is a label that will eventually be misread.
Three properties decide whether it will be read correctly. Character size relative to the viewing distance, contrast against the shelf, and a consistent mounting height. A label at knee height on one bay and shoulder height on the next forces the picker to search, and searching is the moment attention leaves the pick.
Check digits are the cheapest reliability improvement available. A location code with a trailing check digit that the scanner validates will reject a mis-keyed code before the pick is confirmed. Without one, a single wrong character produces a confident pick from the wrong place.
Label condition is a maintenance task rather than a one time installation. A torn label, a label behind a pallet, or a label whose barcode has been scuffed by a forklift is functionally absent. The monthly walk that replaces them is cheaper than the accuracy loss they cause.
Similar SKUs need physical separation, not training
The most reliable way to reduce look alike errors is to stop putting look alikes next to each other. Two SKUs that differ only by colour, by size suffix, or by a single character in the part number will be confused at a rate that no amount of training reduces, because the confusion is perceptual rather than procedural.
Slotting rules should treat a pair of look alikes as a conflict, the same way they treat a fast mover that is too heavy for a high shelf. The resolution is either distance, a different zone, or a deliberate colour coding of the shelf label that makes the difference visible before the part number is read.
A second technique is to break the pattern in the name. If the warehouse display name carries the distinguishing attribute first rather than last, the eye reaches it before the shared prefix. This is a label change rather than a training change, and it costs nothing to deploy.
The third technique is a hard stop in the scanner. When two SKUs are flagged as a look alike pair, the scan of one requires a confirmation of the other attribute. That friction is applied only to the pairs that need it, which keeps the cost proportional to the risk.
Substitution needs an authorisation, not a judgement
A short pick is a decision point, and the decision has an owner. If the picker decides to substitute a similar item, the warehouse has made a commercial choice on the merchant's behalf, and the customer will discover it at unboxing.
The correct design is an authorisation code rather than a policy. The short pick goes to an exception queue, the merchant or its rules engine decides whether to substitute, cancel or partially ship, and the decision comes back to the warehouse as an instruction with an identifier.
Substitution rules can be pre-authorised for the cases where speed matters more than consent. A rule that permits a colour substitution on an accessory but not on a main unit is a legitimate rule, and it is legitimate precisely because it was written down before the order existed.
What cannot be delegated is the definition of equivalence. The warehouse knows the SKU, the merchant knows the product. An equivalence list maintained by the merchant is the only version of this that survives contact with customers.
Three exception types, three routes
Short picks, damaged units found at the face, and missing locations are three different exceptions that get collapsed into one queue in most warehouses. Collapsing them is why the queue becomes a backlog: they need different evidence and different owners.
A short pick needs a count confirmation and a decision on the remainder. A damaged unit needs a condition note and a write off or return to vendor. A missing location needs a cycle count request, because a location that is empty when the system says otherwise is an inventory accuracy signal rather than a picking problem.
The route matters because the feedback loops are different. Short picks improve slotting and replenishment. Damaged units improve handling and packaging. Missing locations improve cycle counting. A single exception queue delivers all three signals to one person who can act on none of them.
The queue also needs a service level of its own. An exception that sits for six hours has already missed the cutoff, so resolving it quickly is pointless. The exception has to be resolved within the window that still allows the order to ship, and that window is minutes rather than hours.
Audit sampling should follow the picker, not the calendar
A fixed ten percent audit is the standard answer and it is the wrong one. It spends the same effort on a picker with a clean six month history and on a picker in their second week, and it produces a number that cannot distinguish the two.
A better design is risk weighted sampling: new pickers, pickers returning from absence, new SKUs, new locations and any picker whose recent error rate moved get a high audit rate. Established pickers on established SKUs get a low rate that still exists, because a zero audit rate is a blind spot rather than a saving.
The audit has to be blind to be honest. If the auditor sees the expected content before opening the tote, the audit measures compliance rather than accuracy. Blind audits cost slightly more and they are the only version whose number can be trusted.
Audit results should feed three places: the picker's record, the slotting rules and the SKU master data. An error pattern concentrated in one location is a slotting problem. An error pattern concentrated in one SKU is a master data problem, usually a duplicate or a bad image.
Error attribution has four buckets
Every picking error belongs to one of four causes: master data, location, person, or equipment. The first two are systemic and the last two are individual, and the split between them is what decides whether the response is a process change or a conversation.
Master data errors show up as pickers doing exactly what the screen said and still being wrong. A duplicate SKU, a wrong image, a unit of measure that does not match the shelf, or a case pack that changed without a system update all produce this pattern.
Location errors show up as errors concentrated in a zone or an aisle. The cause is usually a label problem, a slotting conflict between look alikes, or a location that physically holds two SKUs because a putaway was never confirmed.
Person and equipment errors are the residual, and they are the ones that get the attention because they are the easiest to name. A warehouse that attributes most errors to people is usually a warehouse that has not looked at the first two buckets carefully.
The number you promise the customer
The externally promised number should be the order level accuracy, because that is the unit the customer experiences. The internally managed number should be the line level, because that is the unit with an owner. Publishing both, with the conversion stated, keeps the two from drifting apart.
The promise also needs a definition of failure. An order that arrives complete but late is a delivery failure rather than an accuracy failure, and merging the two produces a number that improves when the carrier improves and deteriorates when the catalogue grows.
Trend matters more than the level. A stable ninety nine percent is a managed process. A number that swings four points month to month is an unmanaged one, and the swing usually traces to seasonal labour rather than to any change in method.
Finally, the number should be reported with its sample size. A week with two hundred orders and one error is a ninety nine point five percent that means very little, and reporting it without the denominator invites decisions that the data cannot support.
The operational form of these checkpoints, with the exception route written beside each scan, is the order picking accuracy checkpoints with a defined exception path, and the routing is what makes the measurement actionable.
- 01Fix the accuracy denominator and publish the conversion to the other two.
- 02Add a put to order scan wherever batch picking is used.
- 03Scan at both the pick face and the packing bench, or choose the bench if only one is affordable.
- 04Validate location codes with a check digit.
- 05Slot look alike SKUs apart rather than training against them.
- 06Put the distinguishing attribute first in the warehouse display name.
- 07Route substitution through an authorisation code, not a picker judgement.
- 08Split the exception queue into short pick, damage and missing location.
- 09Weight audit sampling by picker tenure and recent error rate.
- 10Keep audits blind to the expected content.
- 11Attribute errors to master data, location, person and equipment before responding.
- 12Report the promised number with its sample size.