AI for inventory, orders and procurement
Practical controls for forecasting, inventory, orders and procurement: data, uncertainty, override, waste, suppliers and decisions about people.
AI ACT · OPERATIONS AND INVENTORY
An inventory forecast is not dangerous merely because it uses AI, but it can cause real loss when data, seasonality and uncertainty are hidden inside one number. Good control keeps operational decisions explainable and separates goods from decisions affecting people.
Operational forecasting is not automatically high-risk
High-risk status follows intended purpose and Article 6 plus Annexes I and III. Estimating flowers, packaging or products needed for orders is not automatically a high-risk category. Commission guidance remains draft and non-binding, and its examples do not replace assessment of the real system. [EU-REG] [EU-RISK]
Separate forecasting from decisions about people. The same score later used for employee assessment, work allocation or a natural-person supplier’s access to opportunities may have a different profile. The register should capture not only output but every downstream decision consuming it. [EU-REG] [EU-RISK]
Data quality starts at receipt
A good forecast cannot repair duplicate codes, mixed units, physical and recorded stock differences, missing returns or waste recorded as sales. Define each field’s owner, validation rules and periodic reconciliation. Mark missing data rather than filling it with invented precision. [EU-LIT] [EU-REG]
Keep sales, cancellation, damage, expiry, substitution and manual adjustment separate. For perishables, date and cause of waste may be more useful than a monthly total. The model needs the meaning of a record, not merely a number in a column. [EU-LIT]
Show the range, not only the recommendation
A recommendation to order 120 units appears certain although it may depend on weather, holidays, lead time, promotions and incomplete data. Show scenario, range and key factors. For rare events or new products, explicitly mark low confidence and require a human estimate. [EU-LIT] [EU-REG]
Define asymmetric costs: stockout may lose an order, surplus may create waste and substitution may affect the customer promise. Do not optimise only for average accuracy; metrics should reflect the real consequence of error by category. [EU-LIT]
Overrides should be tracked, not discouraged
A manager should be able to change a recommendation with a reason: special order, local event, supplier issue, price change or information unavailable to the model. Record proposal, final decision, reason and outcome. Repeated overrides may reveal missing data or a stale model. [EU-LIT] [EU-REG]
Do not make low override rates a performance target; people will stop correcting the system. Measure when intervention improved or worsened outcomes and update rules. Escalate when a recommendation exceeds budget, capacity, loss threshold or historical deviation. [EU-LIT]
Supplier and price decisions need boundaries
A system may compare availability, quality, lead time and price, but document criteria and check unintended effects. Do not let proximity, historical volume or complaint rates become an opaque score excluding suppliers without correction. Disputed data should be reviewed before material penalties. [EU-LIT] [EU-REG]
Where the supplier is a company, the operational decision usually differs from a decision about a person. If the platform assesses people, employees or access to work, reassess purpose and Annex coverage. Do not reuse a procurement score for HR or reputation without a new assessment. [EU-RISK] [EU-REG]
A practical control model for ProFlorist
For flower-shop management, a prudent model combines a data register, stock reconciliation, range forecasts, budget and waste thresholds, reasoned override, approval of material orders and error monitoring by category. The objective is better decisions, not replacement of management experience. [EU-LIT] [EU-REG]
Reassessment is triggered by changes to data source, model, assortment, season, supplier or downstream function. Retain version, tests, limits and approval. This model describes general controls for inventory, procurement and orders and does not expose ProFlorist’s proprietary logic. [EU-DESK] [EU-LIT]