Predictive Analytics for One Expensive Decision

The most common failure in business analytics is ambition without a target. A company decides it needs an AI strategy, funds a broad initiative, hires or contracts a team, and eighteen months later has a set of impressive demonstrations and no decision that is made differently.

The alternative is unglamorous and works: pick one expensive decision that is still being made on instinct, and build a model for that decision alone.

Choosing the decision

The right candidate has four properties.

• It is expensive. Pricing, churn, inventory, demand, credit, staffing. Decisions where being ten percent better is worth real money.
• It is repeated. A decision made weekly or monthly gives you enough history to learn from and enough future to benefit.
• It has an owner. Someone whose job gets easier if the model works and who will actually use it.
• The data exists. Not perfect data, but enough history of the decision and its outcome to validate against.

A retail business deciding weekly replenishment quantities across hundreds of products is a good candidate. A one time decision about whether to enter a new country is not, however important it may be.

What building it looks like

The work begins with the decision, not the algorithm. What exactly is being decided, by whom, how often, with what information, and what does a good outcome look like? Only then does the data work start: assembling history, cleaning it, and testing whether the signal is strong enough to be useful.

The model itself is often simpler than people expect. Interpretable methods that the decision owner can understand frequently outperform complex ones once the cost of trust is included.

Validation is against your own history. If the model had been in place for the past two years, how would it have done? That question, answered honestly, tells you whether to deploy.

Prescriptive is the next step, not the first

Predictive analytics tells you what is likely to happen. Prescriptive analytics recommends what to do about it. The second depends entirely on the first being trusted, which is why we sequence them. A churn model that the retention team believes becomes the basis for a retention playbook. A churn model they do not believe becomes a slide.

The business impact

A model for one decision typically takes four to eight weeks. The return is measurable because the decision was expensive to begin with: fewer stockouts, less discounting, customers saved before they leave. And the organisation learns something more valuable than the model itself: how to move from data to a decision it can defend. The second model is always faster than the first.