E-commerce & Digital

Returns prediction

Score orders by likelihood of return based on product category, customer history, order composition, and fulfilment signals. Enables proactive intervention.

Decision supportRepetitive judgmentCost reductionCustomer experience

10–20%

reduction in return rate for targeted orders

Opportunity assessment

Business Impact
4

Significant impact. Material effect on profitability, revenue, or cost base — visible at business level.

Feasibility
2

Challenging. Significant technical work required. Likely needs an AI engineer or specialist vendor. Not a quick win.

Data Readiness
2

Demanding data requirements. Relies on clean, integrated data across multiple systems. Significant data preparation work typically needed.

Risk Exposure
4

Low risk. Limited external exposure. A human reviews output before it reaches anyone outside the team.

Change Complexity
3

Moderate people impact. Part of someone's working day changes. Requires training and some adjustment time, but roles remain broadly the same.

Tooling required

Fine-tuned modelWorkflow automation

Things to consider

  • Zalando's AI-powered virtual try-on and GPT-powered assistant reduced return rates by up to 7% among users. Prior AI size tools had already reduced wrong-size returns by 21% (Reruption).

  • Requires a unified view of order history, customer history, and return history. If these sit in separate systems, the data integration work precedes the AI work.

  • Model training needs a substantial labelled dataset of past orders with return outcomes — typically 12-24 months of history minimum.

  • Define the intervention: what does the team do with a high-return-risk score? The prediction is only valuable if there's an action attached to it.

  • Design the intervention before building the model. What does the business do with a high-return-risk prediction? Options include a proactive fit recommendation, a customer service flag, or a targeted email. The action is where the value is.

  • Experiment starter: Test the intervention hypothesis first, without any ML. Have a merchandiser manually flag 100 orders they would consider high return risk based on experience. Trigger a proactive communication or intervention for those customers. If the return rate on that group is materially lower than the baseline, the intervention works — now build the model to automate the identification.

AI Transformation Playbook

Ready to assess your own opportunities?

The playbook gives you the full 150-opportunity directory, scoring tools, and 100+ templates for every stage of an AI transformation programme.