Week 1: connect and check the data
The first week is about the ERP, not the model. We connect IRIS to Sage, Odoo, Dynamics 365 or SAP, or set up a scheduled CSV / SFTP export, and pull four things: the item master with brand and category, stock by warehouse and transit warehouse, invoiced sales by SKU and store, and open supplier orders.
Then we check. Duplicate item codes, stores with no sales, negative stock, testers coded as sellable items: every network has some of it. Fixing the data now is the difference between a model that works and a model that is blamed for the data.
Week 2: calibrate on your history
With twelve to twenty-four months of history, the forecasting engine trains and compares several models per SKU and store and keeps the most accurate. We review accuracy by category with the planner: where the error is high, we look for a reason, usually a promotion, a launch or a stock-out that hid real demand.
This is also the week to set the rules: safety stock by ABC class, service-level targets by brand tier, supplier lead times from real deliveries, case packs and minimums.
Week 3: run in parallel
IRIS produces order proposals every day; the planner keeps ordering the old way. Each morning the two are compared. Where they differ, the planner decides and tells us why. Three outcomes are common: the proposal is better and the planner adopts it; the planner knows something the data does not, and we add it as a rule or an event; or the proposal is wrong because of a data issue we missed in week one.
By the end of the week the planner trusts the proposals for most categories and knows exactly where to look for the rest.
Week 4: go live, with thresholds
Go-live does not mean automatic ordering for everything. It means proposals are validated in IRIS and created in the ERP with the IRIS reference, and the status comes back. Approval thresholds start conservative: everything is validated by hand, then category by category, orders below a quantity or a value are approved automatically.
On the pilot we ran, 93 % of orders were auto-approved after eight weeks, and nobody had retyped an order into Sage since week four.
What makes it fail
Three things, in our experience: a data check skipped in week one, a planner who was not involved in week three, and thresholds opened too fast in week four. None of them is about the algorithm.
Key takeaway
Four weeks: connect and check, calibrate, run in parallel, go live with thresholds. The planner's judgement is the input, not the obstacle.