How it works · Accuracy
How accurate is camera based stocktaking?
Sentispec sites hold 99%+ stock accuracy after go live, and individual sites reach up to 99.99%. The figure is set against a controlled baseline count during deployment and signed off by your team rather than by ours, so you never have to take the number on trust.
Definition first
What the number is a percentage of.
An accuracy figure only means something once you know what it was measured against. Sentispec measures location level accuracy, and every position the run covers sits in the denominator.
- It is location level accuracy. The unit records which pallet sits in which position in the racking, and whether a position is occupied or empty. Accuracy is the share of positions where the physical state matches what the WMS holds. It is not a count of cartons, and it is not stock value.
- Empty positions count in the denominator. A system that only checks the locations your WMS says are full will not find a pallet the WMS does not know about, and it will report a flattering number. Every position covered by the run is included, occupied or not.
- The unit of measurement is a pallet in a position. Both parts have to be right. A pallet that exists but sits one bay along is a discrepancy, not a match, because a picker sent to the recorded position finds nothing.
- Accuracy is stated after go live, not at best case. Every published figure is measured on a working site after implementation, during normal operation, not from a controlled demonstration run.
What this figure is not. It is not a claim about your WMS being right. It is a measurement of the difference between your records and your racking, and at go live most of that difference has been accumulating for years. The first count at a new site typically finds more than the steady state number, which is the point of running it.
Method
How accuracy is measured, step by step.
Measurement is built into the five day deployment rather than bolted on afterwards, because a number with no baseline cannot be defended in an audit.
Calibration and label validation
Day 1 and 2 of deployment. Your label format is validated against real recordings and racking navigation is checked aisle by aisle, so the reader is proven on your stock before any figure is quoted.
Controlled baseline
A set of positions is counted both ways, by the unit and by hand, and the two results are compared line by line. Disagreements are opened as images and resolved individually.
Acceptance test sign off
Day 4 to 5. Your team, not ours, accepts the result. That sign off is the point at which the site has a documented accuracy figure with a method attached.
Continuous monitoring
After go live, every run updates the figure. Accuracy becomes a trend line on the dashboard rather than an annual event, with four weeks of hypercare while the habit forms.
The baseline is counted twice because a recount is the only test that settles a disagreement. When the unit and the hand count differ, one of them is wrong, and the recording says which. That step regularly finds errors in the manual count. The full week is set out on deployment.
Evidence behind every discrepancy line
A hand count produces a number on a sheet. If a line is challenged three weeks later, there is nothing to look at, so the only way to settle it is to count again. That is the hidden cost of manual stocktaking, and it is not in the labour line.
Because every discrepancy line opens the image that produced it, the same challenge is resolved by looking at the label. An auditor can sample the list and check the evidence themselves rather than accepting a percentage. One pharmaceutical site generates auditor approved reports automatically for exactly this reason.
Every accuracy figure Sentispec publishes carries the site, the sector and the number of pallet locations behind it. The benchmark page states the method for all ten.
Three weeks after the count, a challenged line is settled by opening the photograph of the pallet behind it. That takes seconds, not an aisle walk.
In practice
What 99%+ means on a Tuesday morning.
Percentages hide the size of the working problem. Converted into open discrepancies, the difference between 99% and 99.9% is the difference between a team task and a ten minute job.
| Accuracy | Positions wrong at any time | What that feels like |
|---|---|---|
| 95% | 1,500 | Pickers routinely find the wrong thing. Searches are normal work. |
| 99% | 300 | Exceptions are a managed queue. Short shipments are rare and explainable. |
| 99.9% | 30 | A daily list one person clears. Stock is trusted by planning. |
| 99.99% | 3 | Discrepancies are individually investigated events. |
Holding the number is a different job from reaching it, and it is the reason continuous counting beats an annual event. A driver covers 500 to 1,000 pallets in 20 minutes, driven during a break, so a site can touch every location several times a year without releasing a shift. That is what keeps a site at 99%+ all year rather than restoring accuracy once and watching it drift. See continuous cycle counting.
One three warehouse hub in Eindhoven replaced two annual wall to wall counts of 90,000 pallets, previously 3,600+ hours of manual work, with 80+ hours, and holds 99%+ accuracy all year round rather than twice a year.
The range across the ten published sites
Accuracy across ten published sites.
The lowest figure below is 99%+ and the highest is 99.99%+, measured after go live at sites between 5,000 and 85,000 pallet locations.
| Location | Pallet locations | Accuracy after go live | Sector |
|---|---|---|---|
| Mechelen, Belgium | 65,000 | 99.99%+ | Retail |
| Geel, Belgium | 7,000 | 99.99%+ | Pharma |
| Łódź, Poland | 15,000 | 99.99% | Manufacturing, retail and fashion |
| Willebroek, Belgium | 29,000 | 99.99% | Retail and fashion |
| Duisburg, Germany | 30,000 | 99.9% | Retail and fashion |
| Bremen, Germany | 35,000 | 99.5% | Retail and fashion |
| Venlo and Roermond, Netherlands | 85,000 | 99%+ | Retail and fashion |
| Horsens, Denmark | 40,000 | 99%+ | Retail |
| Kolding, Denmark | 25,000 | 99%+ | Food |
| Cuxhaven, Germany | 5,000 | 99%+ | Manufacturing |
Willebroek recorded a nine point accuracy gain against its previous baseline, which is the more useful figure if your starting point is not already close to 99%. One global 3PL in retail and fashion measured 99.97% while scanning 100,000 pallets a month.
Limits
What a camera count does not tell you.
Four things a location level camera count does not measure. If one of them is the gap you need closed, the fix is somewhere else in your process.
- It does not verify what is inside a pallet. The unit reads the label and the position. If a pallet is labelled correctly and holds the wrong goods, or holds fewer cases than the label claims, a location level count will report a match. That is a receiving and put away control, not a stocktaking one.
- It does not read a label it cannot see. A label turned inwards, shrink wrapped over, damaged or obscured by an overhanging load is reported as unread rather than guessed. Unread positions are findings, and the image is there to look at, so the driver or an operator resolves them without recounting the aisle.
- It does not count what the run did not cover. Accuracy applies to the positions driven. A run that skips an aisle, a block stacked floor area or a mezzanine has not measured them, and the report says which positions were covered.
- It does not correct your master data. If a location code does not exist in your WMS, or two locations share a code, reconciliation surfaces the conflict but the fix is in your system. Several sites treat the first count as a master data exercise as much as a stock one.
First pass read rate: typically 99.x% for valid barcodes. That is the proportion of valid, readable labels the unit resolves on the first pass down the aisle, before anyone reviews an image. A damaged or obscured barcode is the same obstacle for Sentispec as it is for a handheld scanner. Neither can read what is not there, and a site with a printing or wrapping problem will see it in both methods. The difference is that an unread position here is a finding with an image attached, so somebody can look at it rather than walk back to the aisle.
Where one racking position carries several labels, the unit reads the labels visible in the image and reports against the position, which is why the day two step of the five day deployment is LPN translation validation against your own label layout rather than a generic mapping. Sites with mixed loads settle that behaviour before the acceptance test rather than after go live.
Questions about the number
What sceptical buyers ask.
How accurate is camera based stocktaking compared with a hand scanner?
Sites reach 99%+ accuracy after go live with a camera count, against a hand scanner process that moves 50 to 100 pallets per hour and has no evidence trail behind any line. The accuracy gain comes less from the camera reading better than a person and more from counting often enough that errors are caught within days. A once a year count is accurate for about a week.
Would we take your word for the accuracy figure?
No, and you should not have to. The baseline is a controlled count run by your team during deployment, your team signs off the acceptance test, and every discrepancy line afterwards carries the image behind it, so you can sample the list at any time and check it yourself.
What is the first pass read rate?
Typically 99.x% for valid barcodes, measured before anyone reviews an image. A damaged or obscured barcode is the same obstacle for Sentispec as it is for a handheld scanner, so a site with a label printing or wrapping problem sees it in both methods. What changes is that an unread position arrives as a finding with the image attached rather than as a gap somebody has to walk back and check.
What happens if a label is unreadable?
The portal reports the position as unread rather than assuming it is empty or correct, so an unread label is never counted as a match. The image is attached, so an operator resolves it from a screen.
Does accuracy drop in a cold store or at height?
Neither environment is a special case for a camera on a forklift, and the highest figures in the benchmark come from a 7,000 location pharma site with frozen zones and from high bay retail sites. The unit reads the full racking height from the floor, one shelf level per pass, which also removes the scissor lift work a manual count needs. See cold chain stocktaking.
The causes of inventory error, the ways of measuring it and the trade offs between them are covered in the inventory accuracy guide. Every number above comes from the ten site benchmark, which states what the measurement does not control for.
Next step
Test the accuracy claim on your own racking.
A 30 minute demo on recordings from a site with your racking and label format, ending with an open discrepancy list you can interrogate line by line.