Turn operational data
into the next decision.

It is a governed data product that helps a named person make a recurring operational or commercial decision, such as how much to buy, which stock needs intervention, where margin leaked or how work should be planned. It combines fit-for-purpose data, an interpretable method, a workflow action and an outcome feedback loop. A dashboard or model alone is not enough if the result arrives outside the decision window or nobody owns what happens next.

Useful inputs may include order history at the right product, customer, site and unit grain; availability and lost-sales context; seasonality; holidays; promotions; price; lead time; waste; substitutions; open orders and relevant external factors. More fields are not automatically better. We assess point-in-time availability, coverage, changing products and customers, missingness and whether historical demand represents the future decision before selecting a method.

It can support the decision, but performance and uncertainty should be assessed by relevant segment rather than hidden in one average. Seasonal, intermittent, new or promotion-led products may need different baselines, features or review rules. Sparse products may be better handled by a transparent policy than a complex model. The workflow should expose confidence and comparable history so a buyer can apply context the data does not contain.

We start with the decision and document the measure’s grain, timing, included and excluded states, source fields, transformation, owner and known limitations. Finance, commercial and operations may legitimately need different margin views, but those differences should be named rather than hidden behind one label. Automated tests and record-level lineage allow users to reproduce a number and identify whether a disagreement comes from source data, timing or definition.

No. A governed measure, threshold, seasonal baseline, optimisation method or scenario tool may deliver the needed value with less data and risk. Machine learning is appropriate when the pattern, data, performance and operating case justify it. ORBN compares the simplest credible baseline before adding complexity and evaluates methods against business error costs, relevant segments, interpretability and production operation—not novelty.

The workflow shows the recommendation, evidence, uncertainty and permitted action. Where risk requires it, users can approve, adjust or reject and record a reason. That is not failure: legitimate overrides can contain information missing from the data. We monitor outcome and override patterns to improve the product, while guarding against automation bias and avoiding performance measures that pressure people to accept unsafe advice.

A bounded decision and offline proof can often produce evidence in weeks. Production delivery takes longer when source access, definitions, historical quality, workflow integration, security and outcome capture need work. ORBN separates decision discovery, data and method proof, a shadow or assisted production slice and measured expansion. The schedule should reflect the time needed to observe the real outcome horizon, not only the time required to train a model.

Use technical and business measures together. Forecast error may be assessed by product and horizon, but the real outcome may be waste, availability, emergency buying, working capital, planner time or service. Compare with the current baseline, include the cost of different error types and monitor adoption, override and relevant segments. A lower average error is not automatically useful if it worsens the products or customers where mistakes are most expensive.