The phrase gets used loosely enough to have lost most of its meaning. Vendors apply it to everything from genuine machine learning to a dashboard with conditional formatting. For a manufacturer trying to decide whether any of this is relevant, the useful question isn’t what AI is — it’s what changes about how decisions get made.
Rules Versus Learning
Conventional planning software follows rules a person wrote. If stock falls below this level, order this quantity. Lead time is fourteen days because someone typed fourteen. The rules are transparent, predictable, and static — they don’t improve, and they don’t notice when the world they were written for has changed.
A learning system derives its rules from data instead. It observes that this supplier actually delivers in eleven to twenty-six days with a median of eighteen, that demand for this item spikes ahead of a customer’s own seasonal peak, that these two components fail quality inspection together for a reason nobody documented. It updates as those patterns shift.
That’s the substantive difference. Not intelligence in any interesting sense — pattern recognition at a scale and frequency people can’t match manually.
Where it Applies in a Manufacturing Supply Chain
Demand forecasting, using more inputs than sales history alone and selecting a method appropriate to each item’s demand behaviour.
Inventory optimisation, setting levels per item and location from measured variability rather than uniform rules applied across item classes.
Production scheduling, sequencing against real constraints including changeover matrices, machine capability, and material availability.
Procurement, timing purchases against price forecasts and scoring suppliers on total cost rather than quoted price.
Risk detection, spotting the trends in supplier behaviour that precede failure.
Quality and maintenance, identifying the conditions that historically preceded defects or breakdowns.
What it Doesn’t Do
It doesn’t predict genuine shocks — a fire, an export ban, a sudden insolvency. It doesn’t work on products with no meaningful history. It doesn’t fix bad data; it amplifies it. And it doesn’t remove the need for judgement, because most real decisions involve trade-offs between objectives that a model can weight but not decide.
It also doesn’t replace your ERP. Transaction records, controls, and audit trails remain exactly where they are.
What it Requires From You
Reasonably clean historical data — at least a couple of years of transactions where the records approximate what actually happened. BOMs that match reality. Stock records that agree with the floor. Consumption posted close to when it occurred rather than in weekly batches.
Manufacturers who get poor results from these tools have usually skipped this. The model learns whatever the data says, and if the data says you sold two hundred units when you actually ran out at two hundred, it learns the wrong lesson faithfully.
Where to Start
Pick the decision that currently costs you most when it goes wrong. For most manufacturers that’s either inventory level or production sequence. Start there, on one product family, and measure.
ticktick.ai applies these methods across BOM, procurement, production, and inventory using data from systems you already run.
