Section 1
What a model adds over a moving average
Most small businesses forecast with an average of recent sales, which handles a stable product and nothing else. A trained model earns its cost when three things are present. Seasonality, so demand has a repeating shape. External drivers, such as promotions, weather, paydays or a school term. And variable supplier lead times, so the reorder point moves rather than sitting fixed. If your demand is flat and your lead times are reliable, the average is fine and a model is an expensive way to reproduce it. If you sell hundreds of items with different seasons and three suppliers with different reliability, the human version has already stopped coping and you are absorbing the cost as write-offs. Return mechanics are in [The ROI of AI Automation: What Founders Need to Know](/blog/the-roi-of-ai-automation-what-founders-need-to-know).
Section 2
The forecast is not the decision
This is the distinction that decides whether a project works. A forecast is an estimate of demand. A reorder policy is a rule about what you do given that estimate, your lead time, your cash position and your tolerance for a stockout. Vendors sell forecasts. Money is made or lost on the policy. A better forecast attached to an unchanged ordering habit produces exactly the same stock position you had before, plus a subscription. So write the policy explicitly: for this class of item, hold this much safety stock, reorder at this point, and escalate to a human above this order value. The forecast then feeds a decision rather than a dashboard nobody reads.
Section 3
Where AI helps beyond the forecast
Three adjacent jobs often pay back faster than demand prediction itself. Reading supplier documents: order confirmations, delivery notes and invoices that arrive as PDFs in various formats and currently get keyed in by hand. Catalogue hygiene: matching duplicate items, standardising units and flagging the same product listed three ways, which quietly corrupts every forecast built on top of it. Exception triage: surfacing the twenty items that need attention this week out of two thousand, ranked by cash at risk, rather than producing another full report. Related workflow patterns appear in [Streamlining Customer Service with AI-Powered Chatbots](/blog/streamlining-customer-service-with-ai-powered-chatbots).
Section 4
Starting on one category
Do not start with the whole catalogue. Pick one category with enough sales history to be meaningful and enough value at stake to matter. Run the model in parallel with your current process for a full season, ordering as you would have anyway, and record both sets of decisions. You are looking for whether it would have improved fill rate without raising holding cost. A single quarter is not enough if the category is seasonal. Then expand by category rather than all at once, because the accuracy you achieved on stable items will not transfer to intermittent ones, and discovering that across the whole catalogue at once is expensive.
Section 5
When to override the model
NIST's risk headings, trustworthiness, design, evaluation and use, reduce in inventory to one governance question: who may override, and when must they. The mandatory overrides are regime changes, where history has stopped describing the future. A new competitor, a price change, a supplier failure, a product recall, a market you just entered. The model cannot know these happened. It will confidently extrapolate from a world that no longer exists. Set an approval threshold by order value so nothing large is placed without a person. Log overrides with a reason, because that log becomes the most useful training data you have. And keep a human owner for the policy itself. See [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation) for the internal-communication half of this.
Section 6
Fill rate, turns, and forecast error
Three numbers, and one plain-English definition to make them usable. Fill rate: the share of demand you met from stock on hand. It is the customer-facing measure. Inventory turns: how many times you sold and replaced your stock in a year. It is the cash measure. Forecast error: on average, by how much your prediction missed, usually expressed as a percentage of actual demand. Track it per category, not overall, because a good average hides categories where you are consistently wrong. The pair to watch is fill rate and turns together. Either one alone can be improved by making the other worse, which is how a stock programme can report success while cash disappears into a warehouse.