Demand forecasting and suggested purchase orders

The problem

Retailers, distributors and spare-parts businesses carry hundreds or thousands of items. Buyers often decide what to reorder using memory, a spreadsheet and a feel for the season. It works until the range grows or a key buyer is away. Then two failures appear together: popular items run out, and slow items sit on the shelf tying up cash.

The decision is not only how much will sell. It also depends on how long the supplier takes, minimum order quantities, what is on hand and what is already on order. Our team has built forecasting and ordering tools, and the biggest gains come from bringing these factors into one suggestion.

How it works

1. Load the history. Sales by item and date come from your billing or ERP system. Stock-outs, returns and one-off bulk orders are flagged so they do not distort the forecast.
2. Forecast per item. Models learn trend, seasonality and the effect of promotions. Steady items and lumpy, slow items use different methods, and forecasts come with a range, not a single number.
3. Add the constraints. Lead times, minimum order quantities, pack sizes and your target service level are applied.
4. Net against stock. Stock on hand and on order are subtracted to find what is actually needed.
5. Suggest orders. The system proposes an order per supplier, with the reason for each line.
6. Buyer approves. The buyer edits or approves, and the order is exported to your ERP or sent to the supplier.
7. Learn. Actual sales are compared with forecasts every cycle and the models are tuned where they miss.

Dashboards and exception alerts

Buyers should not read thousands of rows. The dashboard shows forecast against actuals, projected stock-outs and slow movers. Exception alerts surface the few items needing attention: a sudden demand change, a supplier delay, or a forecast that keeps missing.

Inputs, AI action and output

Input	What the system does	Output
Sales history and seasonality	Forecasts demand per item with a range	Item-level forecast
Lead times, minimum quantities, pack sizes	Applies supply constraints	Feasible order quantities
Stock on hand and on order	Computes net requirement	Suggested purchase orders
Forecast against actuals	Monitors accuracy	Dashboard and alerts

Integrations

We read from Tally, Zoho, SAP or a custom ERP, usually through exports or APIs, and write approved orders back or produce them as files for suppliers. Data is modelled on Google Cloud and shown in dashboards on a laptop or phone. This sits within our custom software and BI dashboards and process automation work.

Guardrails and human review

• The system suggests and a buyer approves. It does not place orders on its own unless you opt in for specific low-risk items.
• Each line shows why it was suggested.
• Forecast accuracy is tracked and shown openly.
• New items with no history use proxies you agree on and are marked low confidence.

Data you need to start

• At least a year of sales by item and date, ideally two.
• The item master with pack sizes, minimum order quantities and main supplier.
• Supplier lead times, as actually experienced.
• Current stock and open purchase orders.
• Notes on promotions, price changes and stock-outs, even in a spreadsheet.

A hypothetical walkthrough

This is a generic illustration with made-up numbers. An item sells steadily, with a rise in one season. Lead time is three weeks and the pack size is 20. The forecast for the next month is a little higher than last month. Stock covers two weeks and nothing is on order. The system suggests 30 packs and shows the lead time, forecast and current stock behind it. The buyer changes it to 28 and approves.

What changes for your team day to day

Buyers begin the week with a short list of orders to approve and exceptions to check, instead of building the list from scratch. Sales and purchasing argue less about whose number is right. Buyers keep the last word, and their overrides are recorded to improve the model.

Rollout plan

We start by checking data quality. Then we forecast a sample of categories and back-test against what really happened. Next we add constraints and the suggestion screen and run it beside your buyers. Once suggestions are close to what good buyers would do, we extend to the full range and set up monitoring under an AMC.

Who it is for

Retailers, distributors and spare-parts businesses with a few hundred items or more, repeat ordering and at least a year of sales data. It is less useful for one-off project sales. Read AI demand forecasting for distributors, or talk to us about your data.