An empty shelf is the most expensive item in the store. The customer who did not find what they came for does not postpone the purchase - they buy it from your competitor, and often they stay there. The mirror image costs just as much: surplus stock ties up capital, eats warehouse space, and in fresh categories goes straight into the bin. I spent twenty years running operational systems, and I know how much of the morning meeting goes on arguing about those two extremes. AI-based demand forecasting is the first tool I have seen that shrinks both of them at once.
The numbers, before the stories
The global research in this field, led by McKinsey's work, measures reductions of up to 50% in on-shelf stockouts where AI-based demand forecasting has been deployed, alongside a 10-20% cut in inventory held. Those two numbers look like they should cancel each other out: less stock, and yet fewer empty shelves. That is exactly what good forecasting buys you - inventory placed where demand actually shows up, instead of expensive safety stock spread evenly over everything.
Two more figures from the same research deserve a board meeting. AI-based personalization now drives 10-35% of online revenue at leading retailers. At most Israeli retailers, personalization still means one newsletter for everyone and a blanket coupon, which makes the gap managerial rather than technological. And recent sector analyses put retailers who have adopted AI deeply at 2.3 times the sales growth of their competitors.
Now set the research aside and do the arithmetic yourself. A chain turning over NIS 1 billion that loses 2-3% of sales to empty shelves - a conservative range in this industry - is leaving NIS 20-30 million a year on the floor. Recovering half of that, which is what the benchmarks describe, is worth more than most of the growth initiatives on the table. The same sum works on the other side of the balance sheet: every percentage point off inventory held releases working capital that is currently sitting on a warehouse shelf.
What the system actually does at six in the morning
There is no magic in it, and there is a great deal of work on data you already own. Branch by branch and item by item, the system learns point-of-sale history, seasonality, promotions, weather and the holiday calendar. That last one is where imported off-the-shelf systems fail in Israel: Passover and Rosh Hashanah move with the Hebrew calendar, and a model that knows only the Gregorian one gets the strongest trading weeks of the year wrong.
The objection at this point is always the same: our data is not clean enough. In my experience, point-of-sale and inventory data are among the cleanest data in the average Israeli organization - they live in operational systems that demand accuracy because money moves through them. You do not need a year-long data program before you begin. You need one year of history and the willingness to connect it.
What comes out is not a forty-page report. It is a purchase recommendation per branch, every morning, with a short reason attached. The category manager approves, adjusts or rejects, and the system learns from the corrections as well. This is the line I hold in every system I build: the AI recommends, a person decides. That is how you get past the trust barrier inside the organization, and it is also how you meet the governance requirements regulation is starting to demand.
Why this barely happens in Israel
The State Comptroller report of June 2026 put numbers to it: only 18% of organizations in Israel have an AI strategy, 41% operate with no data governance, and 68% of AI projects stall at pilot and never reach deployment. Banks and insurers announce AI centers; large Israeli retail is almost entirely silent in public. Not for lack of pain - shrinkage, stockouts and overstock are the daily conversation in every chain.
This is the Israeli paradox at full strength. Your employees use AI tools at home more than employees anywhere else in the world, and the organization itself still manages inventory the way it did a decade ago. In a category where no competitor has moved yet, the first chain to run one working, measured system is not buying an operational improvement. It is buying years of lead.
How to start without betting the operation
You do not replace the supply chain, and you do not rush to build an in-house development team: an MIT study found that external tools reach deployment twice as often as internal builds, so the real question is which system to pick and how to fit it to Israeli conditions, not how to build one from nothing. Start narrow and measurable: one category, ten branches, 8-12 weeks. Agree the metrics before anything is switched on - on-shelf stockout rate, days of inventory, shrinkage, and category sales. The system runs alongside the existing process, its recommendations set against the orders actually placed, and management sees the comparison every week.
If the numbers do not move, you stop, and the cost is one small pilot. If they move the way the benchmarks say they will, you have a business case proven on your own data rather than on a vendor's slide. Over my career I have learned that the gap between a pilot that stalls and a project that ships is almost never the technology. It is the discipline of measurement, and one person in management who owns the thing personally.
Do you have a year of point-of-sale and inventory data? That is the whole entry requirement. Send me a message and I would be glad to show you a forecasting demo on one of your own categories.
