Insights · Operational optimization

Can AI help us plan when our operational data is incomplete?

By Michael Joseph · 10 September 2026

Yes, in some cases. You can improve an operational decision before every record is complete. The first task is to establish which parts of the decision the available evidence can support. AI can help organize fragmented information; a forecast still needs a credible way to test whether it predicts what happens next.

Start with the decision you need to make

A business owner asks how much stock to order. An operations director needs to allocate staff. A ministry must decide where additional capacity would relieve a backlog. Each needs a different answer, at a different level of detail and over a different time horizon.

I start by defining that decision and the cost of getting it wrong. A weekly capacity decision may be supportable even when individual transactions are incomplete. A precise prediction for each customer may not be. That distinction determines what we build first.

Separate missing information from misleading information

Before selecting a model, examine why records are missing. An empty field can mean that nothing happened, that nobody recorded it, or that a system failed to transfer it. Those explanations lead to different conclusions. Automatically filling the gap can conceal the very problem the analysis needs to understand. Hyndman and Athanasopoulos discuss why missing observations and unusual values require attention to the underlying process in Forecasting: Principles and Practice.

In a fragmented operation, I would first connect each usable observation to its source and date, separate estimates from recorded facts, and identify gaps that could change the decision. This can produce useful decision support without pretending the evidence is complete.

Use scenarios where prediction is not yet justified

A scenario answers, “What would this choice imply if these assumptions hold?” A forecast estimates what is likely to happen. Both can help a decision, but they serve different purposes. A range assembled from assumptions should be described as a scenario range, not a statistically calibrated prediction interval.

My implemented decision-support work includes evidence-processing workflows and scenario comparison. These capabilities help organize incomplete information and examine alternatives. Their existence alone does not establish predictive accuracy. That requires evaluation against observed outcomes in the intended setting.

Make a forecast earn its place

Where there is enough reliable history, compare the proposed forecast with a simple baseline using later observations that were not used to fit it. Judge the horizon that matters to the decision. A good fit to old records is insufficient evidence of future accuracy; the distinction is explained in the textbook’s forecast evaluation guidance.

Also ask what an error costs. Underestimating demand and overestimating it may have very different consequences. A model that marginally improves an average error score may still be unsuitable for the operational decision.

What should the first engagement deliver?

• A defined operational decision, owner, horizon, and success measure.
• An account of usable data, missing evidence, and consequential assumptions.
• A justified choice between a baseline forecast, scenario analysis, or further data collection.
• A proposed test and a practical route into the existing workflow.

This is the work I bring together through operational optimization and AI system development for businesses, governments, and institutions across the US and EMEA. The starting point is the decision you need to improve.

The linked references support the general practices discussed. They do not endorse Epirroi or validate the implementation examples.