Why AI raw material inventory management breaks on real data
Rare demand spikes break the model first. Trained on averaged history, it treats such a peak as noise and smooths it out — procurement enters the plan before the algorithm has time to rebuild seasonality for a specific item.
The second breaking point is manual reorder point rules. The reorder threshold hardcoded in SAP long ago and without revision knows neither the actual lead time for each supplier, nor the batch size, nor the packaging multiple. The rule triggers by calendar, not by warehouse state.
The third reason is fragmented data on the movement of raw materials and components. Receiving lives in the accounting system, defects and returns — in the shop floor log, actual lead times — in the correspondence of procurement specialists.
Until events are consolidated by batch, date, and warehouse, any model sees a distorted picture and learns from it.
The cost of these gaps is measured in downtime. Component shortages stop the line for days, excess raw materials in the warehouse freeze working capital and go to write-off by shelf life.
Demand forecast accuracy falls to the “shop average,” and missed delivery deadlines cease to be an exception.
Next, we break down how this is closed: which data is consolidated into a single raw material movement loop, how the reorder point is configured for each supplier, and what is checked before launch into production.
What changes after implementing a forecasting system
For the first two to three weeks, the system calculates a forecast by item groups and reconciles it with actual consumption without touching procurement.
For high-turnover items with stable demand, forecast accuracy reaches 85–90% over the lead time horizon; sparse and new items we immediately mark as a manual control zone so as not to hand them to the algorithm blindly.
After that, procurement proceeds from the forecast, taking into account lead time and packaging multiple.
Safety stock ceases to be a fixed number in the item card and is calculated from demand variability and supplier reliability: for stable groups it is reduced by about a third, for unstable ones it remains consciously high.
The share of urgent purchases declines earlier than other metrics. While the request originated from the reorder point, the shortage was discovered after the fact, and the item was shipped on an emergency basis.
With the forecast, the shortage is visible one lead time ahead, so emergency mode ceases to be a regular scenario for high-turnover items.
Replanning changes the time scale. Previously, rebuilding the procurement plan in spreadsheets took days and was done once every one to two weeks, which is why a supply failure or demand spike arrived in the plan with a delay.
Recalculation across all item groups takes minutes: the planner triggers it by event and sees which items have shifted, instead of going through the entire list.
By the third month, inventory turnover changes. Stock ceases to be evenly spread across the warehouse and concentrates where demand is confirmed by data, not where it has historically been customary to keep it.
Separately, a class of failures that was previously considered the norm disappears — downtime due to the absence of a high-turnover component. It remains possible only as a consequence of an external failure, at the supplier or in logistics, but ceases to be regular.
Which planning mode to choose: comparison of approaches
The planning mode is not a matter of preference but a function of demand predictability and the volume of accumulated history. While the item range is narrow and consumption is even, manual min-max rules work without errors and do not require data.
As soon as seasonality, promo, or supplier disruptions appear, the same rules begin either to freeze working capital or to cause line downtime.
| Option | When it fits | Limitations |
|---|---|---|
| Manual min-max rules | Homogeneous item range, even consumption, history shorter than one seasonality cycle | Does not hold seasonality and trend, edited manually, reacts to shortage after the fact |
| Statistical forecast | History for one to two cycles, pronounced trend, dozens to hundreds of items | Breaks on promo, supplier change, and new items without history |
| Machine learning model for forecasting | Thousands of items, heterogeneous factors: lead times, batch size, raw material prices | Requires drift control and regular retraining, cold start for new items |
| Hybrid mode with expert correction | High cost of error, procurement specialists have accumulated experience with specific groups | Corrections must be recorded, otherwise feedback is lost and the model degrades |
We make the transition by item groups, not all at once across the entire system: the reorder point is recalculated from safety stock, lead time, and demand variability, so for some items statistics remain more accurate than the model.
We enable the hybrid planning mode where a forecast error is expensive: the model provides a base value, the procurement specialist makes a correction, and both versions are saved — discrepancies show which factors are missing from the data.
Implementation stages: from data audit to system handover
We fix the sequence of work before the start: for raw material warehouse and for components it is the same, only the set of sources changes. Each stage ends with an artifact that can be verified, so the result does not depend on who leads the project.
- Inventory data audit. We collect the history of movements, balances, purchases, and actual lead times, reconcile units of measure and packaging multiples. The output is a list of items where data is sufficient for calculation, and separately those where accounting must first be corrected.
- Demand and feature calculation scheme. We agree on the planning horizon, target service level, safety stock rules, and rounding to batch. We introduce seasonality, supplier schedule, production duration, and failure history into the calculation.
- Configuration and training on prepared data. We determine how the system reacts to sharp demand spikes and where the planner can override the recommendation manually.
- Validation on historical periods. We run the calculation over closed intervals and compare with actual purchases: how much shortage occurred, how much settled as excess, how often the system's decision diverged from the procurement decision.
- Pilot on a narrow item group. We put one category into operation to evaluate the system's behavior on a live flow and limit risk for the rest of the warehouse.
- System handover and team training. We hand over recalculation regulations, manual correction rules, access rights, and support procedures — after that, the customer's team runs the calculation itself, and we remain on change support.
Set of configurations, regulations, and documentation
A demand forecasting and supply planning system does not live in a presentation but in a set of artifacts that procurement uses every day.
We hand them over to your repository and review them together with specialists — after delivery, there are no solutions that only the contractor can support.
- Calculation settings: planning horizons, target service level by item groups, seasonal profiles, rules for smoothing outliers and handling zero demand.
- Safety stock policy: formulas and coefficients by classes of importance and consumption stability, separate conditions for scarce, long-cycle, and imported items.
- Inventory replenishment regulation: reorder points, packaging multiple and batch size, minimum balance, procurement cycle calendar, and the procedure for approving unplanned requests.
- Exception rules: what to do in case of a supply failure, demand spike, or supplier departure, how a manual forecast correction is documented, and who confirms it.
- Inventory monitoring dashboard: turnover, shortage share, forecast accuracy, stuck batches — with thresholds at which an alert triggers.
- Instructions for the buyer: step-by-step scenarios for working with system recommendations, analysis of “forecast vs. experience” discrepancies, and escalation procedure.
- Planning system documentation: data flow diagram, exchange with the accounting system, data models, role and access matrix.
- Support regulation: model versioning, retraining procedure, input data quality control, change log.
We hand over each artifact in a working version — with an owner on the customer side, an update point, and a criterion by which it is clear that a setting has become outdated.
Case study: reduced shortages and frozen inventory
Manufacturing site: several thousand raw material and component items, lead times from a week to a quarter. The order was assembled once a week in spreadsheets, and up to 40% of lines were edited manually by procurement specialists after a call from the shop floor.
The system did not distinguish a seasonal spike from a one-off deviation — shortages were covered by urgent purchases at the upper end of the price list.
What was changed
The forecast was built for each item separately: the calculation included seasonality, lead time, supplier batch multiple, and target demand coverage share.
Items were divided into classes — by cost and consumption predictability — and each was assigned its own planning horizon and safety stock level.
Actual consumption and receipts were pulled from the accounting system 1C:ERP once a day, so the order was recalculated on fresh data, not month-old data.
Metrics before and after
| Metric | Before | After |
|---|---|---|
| Share of items in shortage | 9% of item range per week | 2% |
| Frozen inventory | baseline | −37% |
| Urgent purchases | 27 per month | 6 |
| Manual edits to order lines | about 40% | 8% |
Manual edits were not disabled but turned into a signal: each is recorded with a reason — recipe change, defective supplier batch, unplanned order — and goes into the next recalculation.
Without this loop, the model loses the context that the procurement specialist knows and after a month begins to make mistakes again.
How to embed AI forecasting if you already have your own purchasing rules
The forecast does not replace purchasing rules but is embedded on top of them. The rules — minimum balance, safety stock, lead time, packaging multiple — remain in the accounting system as constraints. The forecast changes only the recommended quantity and order timing.
At the start, we unload item movements, receipts, write-offs, reservations, and cancellations from the accounting system. Data from 1C and related systems is consolidated into one table by item code and warehouse.
Discrepancies in master data — duplicates, different units of measure, outdated categories — are cleaned before launch; otherwise the forecast learns from garbage.
For the first months, the forecast runs in parallel with the buyer's plan and does not replace anything. We regularly generate a discrepancy report: where the forecast and the specialist diverge, by how much, and why — promo, seasonality, supplier change. The buyer makes the decision; the forecast only shows its number.
The transition is gradual, by product groups. We start with items with a long history and stable demand; critical and new ones remain under manual management.
We increase the share of automatic recommendations as accuracy is confirmed in practice — for each group separately, with the option to return to manual mode.
Feedback is mandatory: when rejecting a recommendation, the buyer specifies the reason, and it goes into training. This is how knowledge that is not in the data enters the model — agreements with a supplier, assortment change. Rollback is disabling recommendations for a group; purchasing rules continue to work.
Common objections about model accuracy and reliability
Before the start, one set of questions almost always comes up: is there enough history on raw materials, what happens if a supplier fails, who is responsible for the number in the procurement plan.
What to do if raw material movement data is incomplete?
First, we calculate the share of gaps and discrepancies between the warehouse and accounting. If the gaps are local, we train the model on weekly balances — noise is smoothed, and the forecast remains workable.
If the history is shorter than the supply cycle, we add orders and confirmed supplier lead times to the calculation.
How does the system react to one-off demand spikes?
We maintain a separate layer of rules: a sharp increase for one item does not rewrite the base forecast for the entire item range. Such a spike is marked as one-off, and after verification the model returns to its previous trajectory. Sustained growth, on the contrary, immediately changes the procurement plan.
Who is responsible for the forecast if a supplier missed the deadline?
We are responsible for the quality of the forecast and its transparency, not for the counterparty's behavior. The model has a supplier reliability assessment block based on shipment history, so the risk of failure is visible in advance.
When the deadline changes, the plan is recalculated automatically, and the decision to replace the supplier remains with you.
How is model drift controlled after launch?
We calculate the error metric daily on actual consumption and compare it with the threshold agreed at the start. If it exceeds the threshold, the model is retrained, and you receive a report: whether consumption or the quality of raw material movement data has changed.
Let's discuss the task and come back with an implementation plan
To make the conversation substantive rather than “about AI in general,” we need inputs on your inventory loop: item range, movement history, and the rules by which purchasing decisions are currently made.
This is where discussing the inventory management task begins — which items actually need a forecast, where there is currently overstock, and where there is regular shortage.
We usually ask you to prepare the following:
- Item range and its structure — raw material and component groups, analogues and interchangeability, units of measure, linkage to suppliers and warehouses.
- Source data for calculation — exports of receipts, consumption, write-offs, and returns over a sufficient horizon, with marking of areas where there were manual adjustments or gaps.
- Supplier constraints — minimum batch, packaging multiple, lead time, frequency of deliveries, deferral terms, and shipment reliability.
- Seasonality and external factors — promo, planned shutdowns, repairs, large one-off orders that cannot be considered a trend.
- Item criticality — what stops production and what allows waiting; acceptable shortage level for each group.
- Working group composition — who owns the data, who confirms the recalculation rules, who accepts the result and will use it daily.
- Targets — planning horizon and step, desired turnover, points where the decision must remain with a human.
After reviewing these inputs, we come back with an estimate of the scope of work and a forecast implementation plan: what data we train the model on, how we validate it on historical periods, how the result enters the current procurement process, and how the effect is calculated.
