The direct answer: do not begin by predicting which customer will return or adding friction to a high-risk order. Connect order, line item, variant, reason, exchange, service interaction, and cost. Choose one frequent cause that can still be corrected before purchase, then test a low-friction intervention at the relevant moment. AI can find patterns and select an action; rules should protect eligibility, transparency, and customer rights. The primary metric is contribution after returns, exchanges, reverse shipping, and service—with conversion and experience as counter-metrics.
An overall return rate combines different failures. Wrong size, technical incompatibility, incomplete description, color mismatch, misunderstood delivery, damage, and a changed mind require different responses. One risk score can penalize a customer while leaving catalog, content, or operations unchanged. A first project should explain which failure is avoidable, when the business can still act, and which action adds the least friction.
Write the Avoidable Return Contract
- Unit — order, line, SKU, variant, or customer–product pair being evaluated.
- Cause — observed reason and evidence that it is correctable before shipment.
- Moment — product page, comparison, cart, confirmation, pick-and-pack, or post-delivery.
- Signal — dimensions, compatibility, exchange history, viewed content, region, delivery promise, or permitted service text.
- Action — clarify, compare, recommend size, confirm compatibility, reset expectation, route to help, or do nothing.
- Boundary — never conceal policy, obstruct a legitimate request, shame a customer, or fabricate a specification.
- Outcome — kept order, exchange, return by cause, net contribution, contact, and satisfaction.
- Comparator — current experience, a simple rule, or a no-intervention group.
- Owner — accountable people for catalog, experience, model, logistics, policy, and scale.
The Contract separates prevention from obstruction. If the problem is a missing dimension, improve the product detail page and guide the choice; do not hide the return route. If compatibility is known, a deterministic rule may be safer than a model. If free-text service notes reveal poorly coded reasons, AI can organize evidence while taxonomy and decisions remain with the business.
Start with the cause, not the algorithm
- Objective compatibility — rules and structured attributes usually beat an opaque score.
- Size and fit — combine guidance, attributes, permitted history, and confidence; make exchange easy when certainty is low.
- Missing content — analyze questions and reasons to prioritize approved images, measurements, materials, and instructions.
- Delivery expectation — confirm timing and availability from current operational data, not generated prose.
- Damage or defect — fix quality, packaging, and suppliers; predicting a customer does not remove the source.
- Fraud — keep it in a separate workflow with appropriate controls, review, and legal basis.
Google Analytics recommends sending item_id and quantity with the refund event to obtain item-level metrics. That linkage is the technical minimum. Without it, the team can see a refund but cannot reconcile the variant, the guidance shown, an exchange, and the reason. Preserve raw reason text for audit while mapping it into a small operational taxonomy.
Choose technology proportional to the problem
- Rule — for dimensions, specifications, restrictions, and deterministic combinations.
- Native capability — for policy, requests, exchanges, and return operations already connected to the commerce platform.
- Text classification — for grouping reasons, conversations, and reviews into auditable causes.
- Predictive model — when labeled volume is sufficient, behavior is stable, and a different pre-purchase action exists.
- Generative AI — for approved explanations grounded in source data, never invented compatibility or policy.
The Early Bird Catches the Worm paper shows that pre-purchase return prediction is technically possible and describes live fashion experiments. It does not mean every retailer needs a custom model. Without an actionable cause, a score only ranks risk. With sparse events or unstable labels, rules and cause analysis are more honest, less expensive comparators.
An eight-week implementation pilot
- Weeks 1–2 — choose one category and one avoidable cause; establish return, exchange, contribution, and contact baselines.
- Week 3 — reconcile order, item, variant, reason, refund, exchange, logistics, and experience exposure.
- Week 4 — audit labels and false positives; define action, fallback, boundaries, and human review.
- Week 5 — run in shadow mode without changing the journey; compare rule, model, and operational judgment.
- Weeks 6–7 — release one low-friction intervention to limited traffic while preserving a comparator.
- Week 8 — decide using net contribution, cause-level return, conversion, exchange, complaint, cost, and operating capacity.
Measure protected contribution, not blocked returns
- Cause — returns and exchanges by reason, category, variant, and cohort.
- Economics — contribution after discount, outbound and reverse shipping, refurbishment, loss, and service.
- Sales — conversion, revenue, order value, and abandonment at the intervention.
- Experience — complaint, contact, help usage, satisfaction, and resolution time.
- Model — coverage, cause-level precision, false alerts, confidence, latency, and drift.
- Operations — executed intervention, fallback, human override, and update time.
- Incrementality — change against a comparator, not an aggregate decline alone.
Returns may decline because assortment, seasonality, inventory, or customer willingness changed. Conversion and contribution therefore belong in the same decision. An exchange retains revenue but has cost. A warning can prevent a mistake and also discourage a correct purchase. Scale only after accounting for those effects.
Cost, accountability, and limits
Cost rises with categories, variants, countries, order and logistics systems, reason quality, real-time requirements, personalization, and experimentation. Ecommerce owns catalog and policy; product and content correct information; CRM governs identity and pressure; data reconciles events; technology integrates; operations measures logistics; service reviews causes; finance defines contribution; legal and privacy validate local use and rights. NIST's Core emphasizes governing, mapping, measuring, and managing risk throughout the lifecycle.
Bring one real category to the diagnostic
Connect the work to product recommendations at https://makinai.co/insights/en/ai-product-recommendations-ecommerce-conversion-margin, incentive incrementality at https://makinai.co/insights/en/crm-discount-incrementality-ai-protect-margin, the minimum data loop at https://makinai.co/insights/en/cdp-ai-marketing-minimum-customer-data-loop-revenue, and first-value onboarding at https://makinai.co/insights/en/ai-customer-onboarding-first-value-crm. MAKINAI implements CRM and commerce at https://makinai.co/services/en/crm-ecommerce-commerce-transformation. Bring one category, item-level history, and the top five reasons; we can define the smallest measurable intervention.
The pilot does not guarantee lower returns or higher contribution. Poor labels, assortment shifts, logistics, and product quality may dominate the outcome. Return policies and consumer rights vary across jurisdictions and require local review. The sources support instrumentation, operations, technical feasibility, and risk management; the Contract and pilot design are MAKINAI editorial recommendations to validate in the real operation.