A tier-one component plant counts its store zone by zone, all year, without ever stopping the line. Two counters, five days a week, a printed sheet and a scanner. The count is never behind schedule and the discrepancy log is always current. At the year-end audit the same plant writes off a chunk of stock value as unexplained variance, and the assembly line has already stopped twice that quarter waiting for a part the system said was on the shelf.
Nothing in that process was done badly. The counting was disciplined, the paperwork was clean, the counters were careful. The problem is structural. A count tells you what was true in one aisle at one moment, and the stock started moving again the second the counter walked away. This guide covers what cycle counting is, how the standard methods work, how to measure accuracy honestly, and where the count event itself is now being replaced.
Cycle counting is an inventory auditing method that counts a small subset of stock on a rotating schedule, so every SKU gets verified several times a year without shutting the operation down. It replaced the annual wall-to-wall physical count in most warehouses and plants. It is now being displaced in turn: Gartner predicts that by 2027, 50% of companies with warehouse operations will use AI-enabled vision systems in place of traditional scanning-based cycle-counting processes.
What is cycle counting?
Cycle counting is a perpetual inventory audit. Instead of counting everything once a year, you count a small slice of locations or SKUs every day or every week, on a rotation, until the whole store has been covered and the rotation starts again. Operations keep running while you count.
The method exists to solve a specific problem with the annual count: by the time you find a discrepancy in an annual wall-to-wall count, it may be eleven months old. Nobody remembers the receipt that was posted twice or the pallet that was moved to the overflow rack. Cycle counting shortens that feedback loop, so a variance is found close enough to its cause that someone can still fix the cause and not just the number.
That is the real deliverable of a cycle count program, and it is worth being precise about it. A cycle count does not create accuracy. It measures accuracy, and it surfaces the process defect that destroyed accuracy, early enough to correct it. If the count finds the same variance in the same aisle every rotation and nothing upstream changes, the program is producing a very tidy record of a problem it is not solving.
What is a cycle count, and how is it different from a physical inventory count?
A cycle count is a partial count of selected items while the operation runs. A physical inventory count, also called a wall-to-wall count, is a complete count of everything, usually with production and shipping halted so nothing moves mid-count.
| Dimension | Cycle counting | Wall-to-wall physical count |
|---|---|---|
| Scope per event | A subset of SKUs or locations | Every SKU in every location |
| Frequency | Daily or weekly, on rotation | Typically once or twice a year |
| Operational impact | Runs alongside production | Line and dispatch usually stopped |
| Time to detect an error | Days to weeks | Up to a full year |
| Labour pattern | Small, continuous, specialised | Large, one-off, often borrowed staff |
| Main weakness | Only sees what is on the schedule | Finds errors long after their cause |
Most regulated and audited operations still run at least one full physical count for statutory reasons, and a mature cycle count program is often what lets an auditor accept a reduced-scope year-end count. The two are not rivals. Cycle counting is the operational control, the physical count is the statutory backstop.
What is ABC cycle counting, and how do I decide what to count?
ABC cycle counting ranks every SKU by annual consumption value and counts the high-value ones more often. A items are the roughly 20% of SKUs that carry about 80% of the value and get counted most frequently. B items are counted less often. C items, the long tail of low-value parts, are counted a few times a year.
ABC is the most common method, not the only one, and the alternatives suit different failure patterns:
- Control group counting. Count the same small set of items repeatedly over a short period. You are not auditing stock, you are auditing the counting process itself, to find whether your procedure is the thing generating the errors.
- Random sample counting. Count a random selection each day. Nobody can predict or prepare for the count, which matters when you suspect the schedule is being gamed.
- Opportunity-based counting. Count an item at a natural trigger point: when a bin hits zero, when the last unit is picked, when a location is replenished. An empty bin takes seconds to verify and is the cheapest possible count.
- Usage-based counting. Count by movement frequency rather than value. A cheap fastener touched two hundred times a day accumulates transaction error faster than an expensive casting touched twice a month.
Value and movement are different risk axes, and ABC only covers the first one. A high-velocity, low-value part is exactly the item an ABC-only schedule pushes to the back of the queue, and exactly the item most likely to stop a line.
How often should I cycle count inventory?
Frequency comes from the class and the number of counting days you have, not from a rule of thumb. A common starting pattern is A items counted monthly or quarterly, B items twice a year, and C items once a year. From there the arithmetic is simple: total counts required per year, divided by working days, gives you counts per day, which tells you whether the schedule is actually staffable.
A worked example. Take 6,000 SKUs split 1,200 A, 1,800 B, 3,000 C. Counting A quarterly, B twice a year and C once a year gives 4,800 plus 3,600 plus 3,000, so 11,400 counts a year. Across 250 working days that is 46 counts a day, every day, forever, before anyone takes leave. That number is the honest test of whether a cycle count policy is a policy or an aspiration, and it is the point where most programs quietly start skipping the C rotation.
Two adjustments are worth making regardless of class. Raise the frequency for any SKU with a live discrepancy history, because past variance is the best available predictor of future variance. Raise it again for items where a stockout stops production, since the cost of being wrong is not the value of the part, it is the value of the downtime.
How do I calculate inventory accuracy?
Inventory record accuracy is the percentage of counted items whose physical quantity matches the system record:
Inventory Record Accuracy (%) = (Number of items counted that match the record / Total number of items counted) x 100
Count 500 SKUs, find 465 that match the system exactly, and accuracy is 93%. The detail that decides whether this number means anything is the matching rule. A match has to mean the right SKU, in the right location, at the right quantity. If you only compare total on-hand quantity across the whole site, two errors that cancel out will read as a clean count, and the stock sitting in the wrong aisle will stay invisible until someone needs it.
Measure by location, not just by total, and count a variance as a variance even when it nets to zero.
What is a good inventory accuracy percentage?
Benchmarks vary by source and by industry, so the useful thing is the range and the gap between target and reality. NetSuite puts 90% as a reasonable benchmark to aspire to and 95% as world class, while some operations, particularly regulated and high-value ones, hold themselves to 99% and above (NetSuite, Inventory Accuracy: What It Is and How to Improve It).
Actual performance sits well below those targets. CAPS Research data from 2024 put average inventory accuracy at roughly 83%, with only about 69% of organisations tracking the metric at all. In physical retail environments the average is worse, around 65%. Auburn University's RFID Lab found average retail inventory accuracy of 63% before intervention, rising to about 95% with EPC-enabled RFID, alongside a 96% reduction in cycle count time (ASCM, on the Auburn University RFID Lab findings).
| Reference point | Inventory accuracy | Source |
|---|---|---|
| World-class target | 95% and above | NetSuite |
| Common aspirational benchmark | 90% | NetSuite |
| Cross-industry average | ~83% | CAPS Research, 2024 |
| Physical retail average | ~65% | Industry benchmark data, 2026 |
| Retail average before RFID | 63% | Auburn University RFID Lab |
| Retail average after RFID | ~95% | Auburn University RFID Lab |
The distance between a 95% target and an 83% average is the entire business case, and it has not closed despite decades of scanners, barcodes and warehouse management systems. That gap is the reason the counting method itself is now under review.
Why do my cycle counts still not match?
Because a count is a sample taken at a moment, and the error it is looking for was created between moments. Every mechanism below survives a perfectly executed cycle count.
What happens if stock keeps moving while we are counting?
The count and the transaction race each other, and the record loses. A picker pulls from the location you are counting, a forklift replenishes it from overflow, a receipt is posted late from the dock. The counter records a true number, the system records a true number, and the two are true at different instants. Reconciling them creates an adjustment that looks like a stock error but is really a timing error.
Freezing transactions during the count is the textbook fix, and it is the one thing cycle counting was invented to avoid. Most operations settle for a cut-off convention and absorb the noise. That noise is a floor on how accurate a count-based program can ever be.
What are the results of poor inventory accuracy?
The failure runs in both directions at once, which is why the symptoms look contradictory. Records that read high cause stockouts: the system says the part is there, production plans around it, and the line stops when it is not. Records that read low cause overstocking: the system understates what you hold, so purchasing buys cover you already own, and working capital goes into a rack instead of the business.
Then there is the location dimension, which is the one most often left out. The stock exists, in the right quantity, in a bin nobody has looked in. To the planner that is indistinguishable from a stockout, so the part gets bought again. Now the same SKU sits in two places, and the duplicate is invisible until it turns up as excess, obsolescence or a write-off years later.
Stockouts, overstocking and duplicate stock in the wrong location are not three separate problems. They are three outputs of the same defect, which is that the record and the shelf disagree and nobody knows it yet.
How do I improve warehouse inventory accuracy without hiring more people?
By reducing the number of counts you need rather than by counting faster, because counting is linear in labour and accuracy targets are not. Four moves do most of the work:
- Fix the causes the counts keep finding. Recurring variance in one zone is a process defect with an address: a mislabelled location, a two-step putaway that gets done in one, a receiving cut-off nobody follows.
- Count by exception. Trigger a count on an event that already suggests something is wrong, such as a negative on-hand, a pick shortage or a bin that hit zero early, instead of counting on a calendar.
- Make location accuracy a separate metric. Quantity-only measurement hides the duplicate-stock failure mode entirely.
- Remove the manual observation step. This is the only one of the four that changes the labour curve rather than optimising it, and it is what the technology shift below is about.
How do I improve inventory accuracy with automation?
By capturing stock movement continuously as it happens, so the record is updated by the event rather than reconstructed afterwards by a person with a scanner. The count stops being a scheduled activity and becomes a byproduct of watching the operation.
This is the shift Gartner is describing. By 2027, half of companies with warehouse operations are expected to use AI-enabled vision systems to replace traditional scanning-based cycle-counting processes, and 20% of the 506 supply chain professionals Gartner surveyed in December 2023 had already adopted them (Gartner press release, June 2024). The driver is cost and capability: cameras and pattern-recognition models improved to the point where continuous observation became cheaper than periodic human verification.
Vision-based counting works on what is already in the building. Cameras watch the store and the transfer points, a model identifies SKUs and counts units, and every movement in and out of a location updates the record against the ERP or WMS in real time. Jidoka's NAGARE does this on existing CCTV infrastructure, with no new hardware on the racks, no tags on the parts and no scanner in anyone's hand, and it writes back to SAP and WMS platforms so the system of record stays the system of record. Because the underlying evidence is video, every adjustment carries a timestamped clip showing what actually happened, which turns a variance investigation from an argument into a playback.
See how continuous counting works on a live store: Inventory Record Accuracy with Vision AI
Does this mean cycle counting goes away?
No, and any vendor who says so is overselling. Two things change and one thing stays.
What changes is the labour-driven count event and the detection lag. If the record updates on every movement, the rotation is no longer how you find out about a discrepancy, and 46 counts a day stops being the price of knowing where your stock is.
What stays is the audit. You still need an independent physical verification to validate that the automated record is telling the truth, and statutory and customer audits will still ask for one. The count shrinks to a control function: smaller, less frequent, aimed at verifying the system rather than at policing the shelves.
Cycle counting, scanning and continuous vision compared
| Capability | Manual cycle counting | Barcode or RFID scanning | Continuous vision counting |
|---|---|---|---|
| When the record updates | On the count date | On the scan event | On every movement |
| Detection lag | Days to months | Hours to days | Real time |
| Counting labour | Continuous and linear | Reduced, still manual | None |
| Location-level truth | Only on scheduled zones | Where scans occur | Everywhere in view |
| New hardware on stock | None | Tags or labels per item | None |
| Evidence behind an adjustment | A counter's note | A scan log | Timestamped video |
| Catches duplicate stock in another location | Only if that zone is scheduled | Only if that stock is scanned | Yes |
| Reaction to a discrepancy | Hope it is found next rotation | Investigate after the fact | Alert at the moment it happens |
Scanning improved the count. It did not remove it, which is why the industry average sat at 83% through the entire barcode era. Continuous observation removes it, and that is a different kind of change.
Where to start
If your accuracy is below 90%, the cycle count program is not the problem to fix first. Find what the counts keep telling you and fix that: the location that always drifts, the receipt posted a shift late, the transfer done in the aisle and entered the next morning. A count program cannot outrun the process that is feeding it errors.
If your accuracy is already respectable and the cost of holding it there is two people counting every working day forever, then the question is no longer how to count better. It is whether the count needs to be a human activity at all. Right SKU, right location, right quantity, verified continuously, is a different operating model from counting a rotation and hoping the gaps hold.
Jidoka's NAGARE runs inventory verification on the cameras already installed in your store, integrates with your existing SAP or WMS, and goes live on a real line rather than in a six-month pilot. If you want to see what continuous counting would surface in your own operation, talk to our team, or read how the approach works on the inventory record accuracy page. Related reading: AI sorting and counting, kitting verification, Vision AI for warehousing and logistics, and what inventory management is.
