A forecasting demo is a seductive thing. Clean charts, a confidence interval, a little animation of the reorder point sliding into place. I’ve watched more than one retail leadership team fall for it, sign on, and then spend the next two quarters wondering for Inventory Optimization why the system keeps suggesting they reorder a product sitting in the stockroom, three cases deep.
The model was fine. What it was fed wasn’t.
IHL Group’s latest inventory distortion research still puts out-of-stocks and overstocks at $1.73 trillion a year globally, and that’s after $172 billion in improvements. Here’s the detail I find more telling, though. AI and machine learning improved demand planning for 76% of retailers. Fewer than one in four had actually deployed it where distortion hurts most.
The Real Bottleneck Is Record Accuracy, Not Prediction
There’s a study I keep coming back to. Nicole DeHoratius and Ananth Raman examined nearly 370,000 inventory records from 37 stores belonging to one retailer. Sixty-five percent were inaccurate.
Yes, it’s from 2008. And yes, a lot has changed since. RFID, handheld scanners, cloud POS. But the improvement hasn’t been spread evenly. A national chain running daily cycle counts lives on a different planet from the independent shop owner who reconciles stock in Excel after closing on a Sunday.
Why does this matter so much for AI? Because a forecasting model believes what it’s told.
Say the system thinks a SKU has 12 units on hand. Really, there are 3, the other nine having walked out the door, been broken in the back room, or never arrived in the first place. Sales look sluggish, so the model trims the next order. The shelf goes empty. Sales fall more. The model trims again.
Shrink costs you once at the register. Then it keeps costing you, decision after decision.
A Four-Layer Stack For Inventory Optimization
I find it more useful to think about stock management as layers than as a software purchase. Four of them, built bottom-up. Get one wrong, and everything above it inherits the problem.
1.Count What You Actually Have
Twice-a-year full counts were never great. Now they’re simply too slow. The operators I’d point to as examples spot-count their fast movers and their theft-prone items every week, phone or scanner in hand, without closing the store.
They also write down why a number was off. Theft? Breakage? A mis-scan at the register? A distributor who shorted the delivery? An unexplained variance is just noise, while an explained one tells you where to look next.
And they know their record accuracy as a percentage. Try asking a group of store owners for theirs sometime. You’ll get shrugs.
2. Connect The Systems That Touch Stock
Checkout, receiving, online orders, purchasing. If those don’t all write to the same inventory record, errors creep in at every handoff. Somebody keys in a vendor invoice by hand. The webstore sells from its own pool of stock. A transfer between two locations lives on a clipboard.
Lots of teams put a BI layer on top of all this, and the benefits of Power BI for modern enterprises are genuine. Still, a dashboard is a mirror. Point it at messy data, and you get a very attractive picture of a mess.
3. Predict, Once The Inputs Deserve It
Only here does forecasting really start paying for itself. Promotions, local weather, a stadium event down the road, the week before Thanksgiving. Decent models now absorb all of it and adjust as fresh sales come in. At the enterprise end, you’re looking at RELEX, Blue Yonder, Oracle. Smaller retailers increasingly get some version of it baked into their POS.
Trusting the forecast also lets you hold less. That’s part of why on-demand production is replacing bulk inventory in apparel retail, and it’s a lesson that travels to other categories.
4. Automate The Decisions You’ve Already Validated
Auto-replenishment. Draft POs. AI that reads a supplier invoice and books it into stock. All useful, all dangerous if layer one is shaky.
Coruzant’s piece on automation in distribution describes systems that watch stock move in real time and kick off replenishment once quantities dip below a set level. Great, if the quantities are real.
My own habit is to let automation handle the routine stuff and require a person to approve anything over a dollar limit you’ve decided on in advance. Because automation doesn’t fix anything. It just does whatever you were already doing, faster.
What This Looks Like In A High-Value, Regulated Category Inventory Optimization ?
Want a tough test case? Liquor stores.
Bottles are small, pricey, and constantly targeted by shoplifters. Distributor deliveries arrive short or with a swapped SKU more often than owners like to admit. In control states, there’s regulatory reporting layered on top. Run the numbers on a store doing $2 million a year at a hypothetical 1.5% shrink, and you lose roughly $30,000.
SanteHQ’s guide to cutting liquor store losses makes an observation I’d apply to almost any independent retailer: most shrinkage stays invisible until a count comes up short, sometimes weeks later. That’s the reason serious inventory optimization begins at the receiving door and with frequent spot counts, long before anyone touches a forecast.
The vendors in this corner of the market know it, which is why they fight over the bottom layer. Category-specific players like SanteHQ, Bottle POS, and Korona POS go up against broad platforms such as Lightspeed and Square.
SanteHQ, to take one example, links mobile spot counts to individual employee logins and runs distributor invoices through AI so they land in inventory automatically. A short shipment shows up the same afternoon instead of in next quarter’s variance report. You can usually get the generalists to do something similar. It just tends to take add-ons and a bit of integration work.
Is there a downside to going niche? Sure. If you ever expand into adjacent categories, you’ll feel the limits. For a store that sells one thing and sells it well, I think that trade is worth making.
Where To Start In The Next 90 Days for Inventory Optimization?
None of this needs a transformation office. It needs a few dull decisions, in roughly this order.
First, find out your actual accuracy. Pick 200 SKUs at random, count them by hand, and compare against the system. The match rate is your baseline. Don’t be surprised if it stings.
Second, sort out receiving before you spend a cent on forecasting. Count each delivery against its invoice before signing. Write exceptions on the paperwork while the driver is still standing there. Chase every credit until it shows up.
Third, stop treating all SKUs equally. Something like a fifth of your catalog drives most of the revenue and most of the theft. Count those weekly. The rest can wait for the quarter.
Then put your automation rules on paper. Which reorders can go through on their own, which need a sign-off, and at what dollar figure. Written down, so nobody quietly loosens them in March.
If you’re planning to grow, think about locations now. The logistics challenges that come with business expansion usually surface first as stock that doesn’t match between stores, and a second site makes the counting problem more than twice as hard.
Finally, ask vendors the unexciting questions. What does your system do when it suspects a record is wrong? Can it flag phantom inventory, where the record says the stock exists and the shelf disagrees? Fuzzy answers there usually mean you’re being sold a forecasting engine without the foundations to support it.
Conclusion: Better Data Beats A Better Model
Retail AI talk tends to orbit around model accuracy and autonomous replenishment. Fair enough, those things matter. But anyone can license Inventory Optimization a good model in 2026. What IHL’s figures actually separate is retailers whose system matches the shelf from retailers whose system doesn’t.
So start with the count. Wire the systems together. Let forecasting prove itself, and automate only what you’d trust a new hire to do unsupervised. Slower than buying a dashboard? Definitely. It’s also the only version I’ve seen hold up.
Also Read :- 7 Logistics Challenges That Come With Business Expansion
