How data leakage can make an AI pilot look better than it is

A promising pilot score can collapse in production without any mysterious change in the market. The test may have contained information that the deployed system would not possess.

Examine the preparation pipeline

Scikit-learn’s common-pitfalls guidance warns that fitting transformations using test data can produce overly optimistic results. Evaluation separation must extend beyond the final model fit.

Check feature selection, normalisation, duplicate removal and any manual curation. Ask when the split occurred and which records informed each fitted component.

An illustrative customer-churn model should not use a field created only after cancellation to predict an earlier decision. The field may be highly predictive while being unavailable at the intended prediction time.

Look for related records across splits

Records from the same customer, document or transaction can share information. A random split may place closely related examples on both sides of the evaluation.

Choose a split that reflects the intended deployment question. Predicting future work for existing customers and generalising to entirely new customers are different tests.

Keep a record of the split method and data version so an independent reviewer can understand the boundary.

Re evaluate before making a commercial claim

If leakage is found, correct the pipeline and rerun the relevant tests. Do not keep the old score in sales material because it is more attractive.

Explain the revised result and what remains useful. Discovering a flaw early is a reason to improve the evidence, not to hide the evaluation process.

Ask a reviewer to trace one test prediction backwards through the data-preparation steps. The objective is to identify information that would not exist at prediction time or should have been fitted only on training data. A clean model-training script is insufficient if an earlier spreadsheet already introduced leakage.

GAT’s pilot-to-revenue coverage places evidence at the centre of buying. Leakage review helps ensure that the evidence describes work the system could actually perform outside its test environment.