Cycle counting is not the problem. How most teams sample for it is. The common approach, count a fixed number of random locations each day, treats every SKU-location as equally worth counting and produces an accuracy figure that is hard to defend statistically. A Bayesian framing fixes both: it tells you where to count and how confident your accuracy number actually is.
Why uniform sampling is weak
If you count locations uniformly at random, you spend the same effort confirming a stable, low-value bin as you do on a volatile, high-value one. Worse, the headline “98% accurate” you report is a point estimate from a sample, with no statement of uncertainty. Two operations can both report 98% while one is far more confident than the other. For a number that drives planning and audits, that is thin.
The Bayesian framing
Treat each location’s accuracy as an unknown probability with a distribution, not a single number. A Beta-Bernoulli model is the natural fit: each count of a location is a Bernoulli trial (matched the book, or did not), and the Beta distribution is the conjugate prior that updates cleanly.
# illustrative: per-location accuracy belief
# prior Beta(a, b); each count updates it
# match -> a += 1
# mismatch -> b += 1
posterior_mean = a / (a + b)
posterior_variance = (a * b) / ((a + b)**2 * (a + b + 1)) # uncertaintyNow you have, per location, both a best estimate of its accuracy and how uncertain that estimate is. That uncertainty is what tells you where the next count is worth spending.
Counting where it pays
Combine the uncertainty with business value to prioritise:
- High uncertainty + high value (fast-moving A items you have not counted recently): count first.
- Low uncertainty + low value (stable C items counted many times): count rarely.
This is just ABC thinking from inventory control, made rigorous: the count schedule follows the posterior, so effort flows to where it reduces the most risk per count.
What you gain
- A defensible accuracy figure with a credible interval, not a bare percentage.
- Fewer counts for the same confidence, because effort concentrates where uncertainty is high.
- Early detection on the locations that matter, because volatile high-value bins get revisited while uncertainty stays high.
None of this replaces the mechanics of cycle counting covered in physical count vs cycle count; it replaces the sampling design. The same daily count budget, pointed by the posterior instead of a random number generator, gives you a stronger accuracy KPI and catches drift sooner.
Implementing this at your scale?
The walkthrough above comes from production work. AvanSaber’s inventory practice has implemented variations of this pattern across multiple customer engagements.
If you are building this and want expert review of your design, or would rather have the team that built this build yours, book a discovery conversation or describe your situation at [email protected].