"Do we have it?" sounds like a question with one answer. In most retail estates it has five, and they disagree with each other.
5
different numbers all called "stock"
Each correct in its own system. Only the last one is a promise you can keep.
1
of them reaches the customer
Usually the first, because it is the easiest to expose.
- Book stock. What the system believes you own.
- Physical stock. What is genuinely in the building.
- Findable stock. What a colleague can actually locate inside the pick window.
- Sellable stock. Findable, undamaged, and not already committed elsewhere.
- Promisable stock. Sellable, minus what you are holding for other channels and open orders.
Where the gap comes from
Between book and promisable, stock leaks in ways that are individually small and collectively decisive:
- Shrinkage that has happened but has not yet been counted.
- Units in the wrong location — present in the building, invisible to the picker.
- Returns received but not yet graded, so not yet sellable.
- Units allocated to a click-and-collect order that has not been picked.
- Damaged stock awaiting a decision nobody has scheduled.
- Goods in transit between sites: owned, but not sellable from either end.
Every one of these is legitimate. None of them are visible in the number a customer sees.
System says
Deduct
Customer sees
Most availability rules apply one national buffer across all of this.
Every over-promise is a customer who trusted the number. That is the real cost, and it does not appear on the stock report.
The asymmetry that matters
Under-promise and you lose a sale you could have made. Over-promise and you lose the sale, the outbound cost, the return leg, a customer service contact, and a quantity of trust that takes several good experiences to rebuild.
These two are not symmetric. Most availability rules are set as though they are — a single buffer, agreed once, applied everywhere.
What a store-level accuracy model is actually for
Stock accuracy is not uniform. It varies by site, by category, by time since the last count, and by how a store is laid out and staffed. That variation is stable enough to model, which makes it one of the more honest applications of prediction in retail.
The useful output is not a number on a dashboard. It is a per-site, per-category confidence that feeds the promise itself: high-confidence locations can sell down to the last unit, low-confidence locations hold a buffer. The buffer stops being a single national figure set in a meeting and becomes a consequence of measured reality.
Before modelling anything
- Can you measure your own accuracy? A counting programme that produces a variance figure is the prerequisite.
- Do you know your over-promise rate — orders accepted and then cancelled for stock?
- Is that figure reported anywhere a commercial decision-maker actually sees it?
If the answer to the last one is no, that is the first thing to fix — and it does not need AI, it needs a report and an owner.
Availability is one of the few areas where a retailer can improve the customer experience and the cost line with the same piece of work. It rarely gets the attention it deserves, because the number on the screen has always looked fine.
In short
- Separate book, physical, findable, sellable and promisable stock — they are not the same number.
- Over-promising and under-promising carry very different costs. Stop treating them as equal.
- Measure your own stock accuracy before modelling it.
- Make the buffer a function of measured confidence, not a single national figure.
- Put the over-promise rate in front of a commercial owner.