Demand forecasting and supplier orders the buyer only has to confirm

At a glance

Client

A Ukrainian brand of veterinary and pharmacy productsmade by contract manufacturers under its own label

Service

emilus. Procurementdemand forecasting

Status

In production at the clienta monthly order to the manufacturers

By the numbers

Up to 200 items10 years of sales history in the accounting system, 2 contract manufacturers

The “before” numbers and the result are per client data; the payback below is an estimate.

Where the client started

The company owns a brand of veterinary and pharmacy products: contract manufacturers produce them under its label. The price list has up to 200 items, and sales go through the company's own online store, wholesale, retail chains and marketplaces. Production takes about three months, so every month the company orders three to four months ahead and keeps roughly half a year of stock in the warehouse.

The buyer pulled sales from the accounting system and did the math by hand. There was almost no forecasting, and seasons and promotions sometimes slipped out of the calculation, so some items were overstocked while others ran short. The range keeps changing, with new items coming in and old ones being phased out. The accuracy the client asked for at the start: 90% or better.

OKThe main rule

The model calculates and a person decides: an order goes to the manufacturer only after the buyer confirms it.

What the system does

10capabilities in one systema human confirms

• Forecasts demand for every item 3–4 months ahead, taking seasons, promotions and past stockouts into account, so orders follow real sales.
• Works out the ready order for the manufacturer itself: goods in transit, batch sizes, safety stock and shelf life. The buyer checks it in hours instead of spending days on manual math.
• Keeps the order within the cash the company is willing to tie up in stock and shows how much money it frees.
• Flags the risk of a stockout or overstock on any item early, while the order can still be changed.
• Explains why it recommends exactly this quantity. The buyer can change any item and note the reason, which stays in the log.
• Counts all channels together and treats sales to retail chains and marketplaces as shipments, not shopper demand, so a one-off bulk order from a chain doesn't inflate the stock.
• Sets up a new item with no history from a similar one and stops ordering items that are being phased out, so a changing range doesn't skew the estimate.
• Sends the order to the manufacturer only after the buyer confirms it, so the final word stays with the person who knows the market.
• Retrains on fresh sales with one button and puts the new version to work only if it tests no worse than the previous one.
• Holds accuracy at 90% or better on fast-moving items and proves it on a blind period the model never saw, so the number can be trusted.

Item A · forecastillustration

History as a solid line, forecast as a dashed one

Order to manufacturer · draftwaiting for OK

Item Aorder

Item B · enough in stockdon't order

Item C · shelf lifeorder less

ApproveEdit

Decision logillustration

1. Buyer changed item A“promo next month”
2. New model version in testingno worse
3. Version in useprevious one kept

Every edit and every model version, with a reason and a date

The architecture in plain terms

The forecast is built from the accounting system's data and tested on a blind period before it goes into an order. In the app, the buyer edits items with a reason and confirms the order.

Data

sales by channelstockgoods in transitstockout flagspromo calendar

Roles

buyer: calculation and editsmanager: retraining

How it works

The monthly cycle: from data to an order for the manufacturer

modelhumansystemOKa human confirms

1. 1
   accounting systemDataSales, stock, goods in transit, stockouts
2. 2
   modelForecastDemand for every item, with seasons and promotions
3. 3
   systemOrder mathBatch sizes, safety stock, shelf life, cash limit
4. 4
   buyerEditsChanges items and notes the reason
5. OK
   buyerA human confirmsThe order goes to the manufacturer only after the OK
6. 6
   managerRetrainingFresh data and one button
7. 7
   systemVersion checkOnly a version no worse than the last one goes live

To be straight with you: high accuracy is achievable on fast-moving items. On the long tail of rarely sold items it isn't, and we won't promise it there.

Results and payback

−22%

cash tied up in stock

per client data

over 90%

forecast accuracy: the client required 90%, and the target is met

per client data

+5–8%

sales of seasonal items, with no stockouts in season

per client data

Project payback: ≈3.5 months estimateCash freed from the warehouse and seasonal sales without stockouts bring ≈78k UAH a month on a turnover of 4M UAH (an assumption). So the $6k System package on our pricing page pays back in ≈3.5 months from the arrival of the first calculated order.

What's live

1. live

   A forecast for every item and a monthly order to the manufacturers

2. live

   Buyer edits with a reason, and confirmation before anything is sent

3. live

   One-button retraining for the manager

More case studies

• Forecast-based purchasing for restaurants and catering
• A custom CRM and an AI sales agent for an electronics chain
• Cold outreach for a merch maker

All case studies