De Rococo was losing 3.1% of revenue to wrong orders and orders placed too late. Six months after Sizo went in, 77% of that leak was closed. This is how.
Fashion brands — especially mid-level ones, up to $100M in revenue — are fighting with data and decision-making. They don't know how to process all of it, and the reports they have don't lead them to a conclusion.
De Rococo Israel is one of the fastest-growing fashion brands in the country. We started working with them six months ago, once we understood the pain they were carrying.
At De Rococo, we found that producing a report for a single product took 30 minutes or more of fetching and calculating data — and even then, information was missing. Day to day, the work ran on Excel sheets and gut feeling. Even with an engineer as co-founder, they struggled to find a solution — including when they went looking in the market.
To make it concrete with one running style: the Mid Rise Denim Shorts (Ivory Blue) were losing 14.3% of the demand on their size ladder to missing sizes — roughly one of every seven shoppers who wanted this style asked for a size that wasn't there.
In our overall look at the industry, De Rococo is actually performing really well. Industry research puts inventory distortion — stockouts plus overstock — at 6.2% of retail sales worldwide (IHL Group, 2026), and fashion is among the hardest-hit categories: only about 60% of fashion inventory sells at full price, and unsold merchandise runs at 17–20% of stock. Measured against that, De Rococo's 3.1% is roughly half the market average — in the industry's hardest category. That is what a well-run brand looks like before Sizo: better than the market, and still leaving real money on the table.
One of the things we recognized really fast: they are not analysts, and they are not going to be. Another: an AI agent will not solve it alone — it needs statistical models to reduce cost and be more accurate.
So we decided to move from reports into decisions.
We wanted each shop to have its own training, so we built a brain for the shops. It has main categories, and each SKU is a node living inside a cluster, connected by reason graphs. The idea is to let the agent conclude — and update its own data for its shop.
The statistical model was split into different categories: supply chain, size distribution, and quantity.
The implementation also includes an alert system, and direct orders to manage them with the manufacturer.
De Rococo sells from multiple locations, each with its own demand profile. Sizo decides the total quantity first, then splits it between locations by what each store actually sells, per SKU — so a size isn't sitting in one store while selling out in another.
Another thing we understood: they want short and direct answers. Every answer follows one shape:
Every number below comes from Shopify's own daily inventory record — 7.3 million variant-days measured identically before and after — never from Sizo's own event tables, which would flatter the comparison.
Before Sizo, when a size that was selling ran out, it was back within a month about half the time. With Sizo it is 69% — and that figure holds whether "before" means the previous 6 months, 12, 18, or the weeks immediately before the install.
| Definition of “before” | Episodes | Fixed ≤30d |
|---|---|---|
| All pre-install (from Jan 2024) | 5,705 | 54.9% |
| 18 months before | 4,258 | 49.1% |
| 12 months before | 3,021 | 49.2% |
| 6 months before | 1,911 | 49.8% |
| The weeks just before install | 560 | 51.4% |
| With Sizo | 1,531 | 68.9% |
| Period | Episodes | Fixed ≤30d |
|---|---|---|
| 2025 H1 | 1,043 | 47.1% |
| 2025 H2 | 1,699 | 50.4% |
| 2026 pre-install | 560 | 51.4% |
| With Sizo | 1,531 | 68.9% |
A best-seller here is a size that proved itself before the season — at least 10 units sold in the 60 days before March. In the first season with Sizo those sizes were out of stock 30% of days, down from 43% the season before, on nearly double the number of best-sellers and with units sold tripling — more sell-pressure, better availability. The sizes that sat out for most of the season — the ones that actually cost orders — fell by 54%, from 38.1% of best-sellers to 17.5%.
| Month | 2025 | 2026 with Sizo | Gap |
|---|---|---|---|
| March | 37.8% | 32.9% | −4.9 |
| April | 41.5% | 31.8% | −9.7 |
| May | 43.5% | 27.1% | −16.4 |
| June | 47.1% | 30.2% | −16.9 |
Size-mix loss is the share of a style's demand that asks for a size that isn't there. On the group of running styles where Sizo's restocks landed, it fell from 14.9% to 5.2% — 65% of that loss gone. Before, roughly one shopper in seven met a missing size; now it is one in nineteen. The flagship is the Limited Edition Collared Satin Dress: since a Sizo-guided restock landed in June 2026, all five sizes have held with zero estimated loss in June, July and August.
Sizo is not only for restocking what already exists. In the six months together, De Rococo launched 500 new products, with first orders sized on Sizo's recommendations from the shop's own demand analysis. Compared with the same launch season a year earlier — each product judged on its first 90 days only, so the comparison is fair — a similar number of launches produced 30% more units, and the share of launches becoming real sellers (100+ units in 90 days) rose from 13% to 20%. The table compares the February–May cohorts — the launches that have a year-earlier twin; the remaining launches of the six months bring the total to 500.
| New products, launch season vs launch season | Last year | With Sizo | Change |
|---|---|---|---|
| Products launched (Feb–May) | 307 | 330 | +7% |
| Units sold in the first 90 days | 18,051 | 23,524 | +30% |
| Average units per new product | 59 | 71 | +21% |
| Became real sellers (100+ units in 90 days) | 13.0% | 20.3% | +56% |
Putting money on the improved products — conservatively. A unit only counts as "captured" if the product's availability math implies it and the sales growth actually happened; we never credit a unit that wasn't really sold.
De Rococo came in leaking 3.1% of revenue to wrong orders and orders placed too late. That leak is the thing being measured against — not the whole business. Here is how much of it closed, and the calculation, step by step. For every product sold in both seasons, we measure the share of days it stood out of stock — last season against this one. Where availability improved, the extra selling days earn the product's own measured selling rate: a product that sold its units while on the shelf 70% of days, against 20% a year ago, has sales it simply could not have made at last year's availability. Those captured units are priced at the product's own realised prices from this season's actual orders, summed, and divided by the season's total revenue. Computed two ways: counting only captured units matched by sales growth that actually happened gives 1.7%; the standard lost-sales method — the selling rate applied to all recovered in-stock days — gives more, and we cap it conservatively at 2.4%. The fixed size ladders, where lost demand fell from one shopper in seven to one in nineteen, are included the same way: the old loss rate applied to this period's real demand, minus what was actually lost. Against the 3.1% leak, the 2.4% we report is 77% of everything there was to recover.
The report that took the team 30 minutes or more per product — fetched and calculated by hand, and still incomplete — now takes seconds in the agent. Across the ordering workflow, that cut up to 95% of the working time this problem used to consume.
“Sizo does in seconds what used to take me and my employee a long time and pull us away from other work — saving us money on every single order.”
De Rococo came in with the pain every mid-size fashion brand knows: plenty of data, no decisions. Six months later — stockouts on selling sizes fixed within a month 69% of the time instead of half, best-sellers held through the season, lost demand on fixed ladders down from one shopper in seven to one in nineteen, and new products selling 30% more in their first 90 days — together closing 77% of that 3.1% leak, worth 2.4% of a season's revenue.
And it happened while the shop scaled. Growth is supposed to create mess — recovery sat stuck around 50% through the whole year before the install. With Sizo, selling more than triple the units of two years ago, the shop runs at its best level in the data. No one had to become an analyst: the brain holds the shop's knowledge, the models do the maths, and every decision leaves the app as an alert or a purchase order — not another report.