AI forecasting tools produce short-range and mid-range forecasts that cover the same weeks. Companies that plan production against them need a written rule for which number drives the order and who can override it.
Impact Analytics, a company that sells forecasting software to retailers, recently published a guide to AI demand forecasting that splits the work into three timeframes. In its description, demand sensing reads real-time data for the next one to four weeks, operational forecasts cover one to thirteen weeks and drive replenishment, and strategic forecasts look a year or more ahead for planning. The guide is aimed at retailers. The same structure applies to any company that builds or buys inventory against expected demand, including medical device makers, food producers and distributors. What caught my attention is the overlap in those ranges. For the next four weeks, a company running this model has two forecasts covering the same period, built from different data at different speeds, and they will sometimes disagree.
Suppose the demand-sensing number for week three comes in well above the replenishment forecast. One of them has to drive the purchase order or the production schedule, and there are only a few ways that gets settled. The software can apply a configured rule, a planner can make the call, or nobody makes an explicit choice and whichever number last reached the ERP wins. A company gets the third outcome by default when it installs forecasting software without deciding in advance, and it is the hardest one to explain afterward.
The Impact Analytics guide says stronger forecasts lower inventory risk and free up working capital. That benefit depends on the forecast reaching the order at the right moment, through a decision someone can account for. An accurate sensing signal that arrives after the replenishment order has been released has no effect on that order. A signal that overrides the order without anyone noticing can create the very shortage or excess it was meant to prevent.
In a retail setting, a poor replenishment call costs margin. In a medical device or food plant, the production run a forecast triggers also commits material lots and creates the batch and traceability records that follow the product. When a quality reviewer or an auditor asks why the schedule changed, the answer has to be more specific than "the system updated." If a planner overrode the forecast, there should be a record of who did it and why. If the software switched automatically from the operational forecast to the sensing signal, the rule that governs that switch should be written down. Changes to that rule should be controlled like any other change to a system that affects production.
That requirement creates real friction. If your ERP or planning system sits inside a validated environment, adjusting planning parameters may have to go through your change control process, which makes the rules slow to tune once the software is live. The planning team matters too. Experienced planners often know things that are not in any dataset, such as a customer's scheduled plant shutdown or a supplier's capacity problem, and they will keep overriding a number they do not trust whether or not the system gives them a place to do it. Their overrides are useful information. The aim is to capture them where they can be reviewed, so that the spreadsheet next to the planning system does not become the real record.
Most of this work is decision design, and it can be done before any software is selected. For each horizon, list the decisions it is allowed to drive. Following the source's own ranges, I think a reasonable split is that demand sensing adjusts allocation and shipments from existing inventory, the operational forecast drives replenishment orders and production schedules, and the strategic forecast informs capacity, supplier agreements and site decisions. Then define what happens in the overlapping weeks. Set a tolerance, stated as a percentage or a unit difference, beyond which a disagreement is routed to a named person for a decision.
Laminar's work for Ventura Foods is a useful reference point here, although it had nothing to do with forecasting. The company's new product development process ran on a legacy RPG mainframe and manual approvals, and the web workflow application we built to replace it across U.S., EU and Mexico facilities included logic for choosing the manufacturing site. Logic like that can only be built once someone states the criteria explicitly and agrees who is allowed to change them. Forecast handoffs need the same treatment, and the conversation is much easier to have before a vendor's default settings are already running your orders.
Practical steps for the next planning cycle:
Dealing with a similar challenge?
We work with mid-market companies in regulated industries to build AI workflows that actually hold up.
Let's TalkSean Cummings
Founder of Laminar Consulting Services. Specializes in AI workflow automation for regulated industries — medical device, financial services, and complex logistics operations.