> For the complete documentation index, see [llms.txt](https://docs.layerlens.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.layerlens.ai/more-in-this-section-2/scenarios-10.md). # Retail and e-commerce scenarios Retail and e-commerce scenarios — product Q\&A, recommendations, pricing, customer service, returns/fraud. Industry patterns, not customer case studies. Five canonical scenarios where retailers and marketplaces use Stratix evaluation to gate AI features into production. ## 1. Product discovery and Q\&A **Setup:** Customer asks "Will this jacket fit a 6'2" 220lb frame?" "Is this dishwasher compatible with my hardwired connection?" AI synthesizes the product spec sheet, prior reviews, manufacturer Q\&A; produces an answer with source citations. **Why it's hard:** * Product attributes may be missing or inconsistent across the catalog * Reviews can contradict spec sheets * Customers act on the answer — wrong fit info drives returns; wrong compatibility drives serious damage **Stakes:** * Returns directly hit margin; some categories see 30%+ return rates * Compatibility errors on regulated/installed products (appliances, automotive parts) can trigger product-liability exposure * FTC and state AGs flag misleading product claims even when AI-generated **Stratix evaluation criteria:** | Aspect | What you measure | Pattern | | ----------------------- | -------------------------------------------------------------------- | ------------------ | | Spec grounding | Output cites the spec sheet | Citation existence | | Review-spec conflict | Output flags conflicts vs. picking arbitrarily | Code assertion | | Compatibility hard rule | High-risk categories require explicit compatibility-check disclaimer | Hard rule | | Hallucination rate | Confident wrong claims about specs | Faithfulness judge | | Reading level | Comparable to category baseline | Code scorer | ## 2. Personalized recommendations **Setup:** Recommendation engine surfaces products. AI-augmented version uses LLM to expand "similar products," generate copy, or rerank. **Why it's hard:** * Stale catalog — recommendation surfaces a discontinued SKU * Sensitive-category guardrails — no recommending alcohol/firearms/adult on a kid's account * Personalization fairness — same product should not vary in price/availability based on protected-class proxies **Stakes:** * Bait-and-switch perception when out-of-stock items dominate recommendations * FTC scrutiny on dark patterns and personalized pricing * State AG actions on sensitive categories (alcohol, tobacco, firearms) reaching wrong audience **Stratix evaluation criteria:** * **Catalog freshness rule** — Hard rule: every recommended SKU must be in-stock and live in the catalog as of trace time * **Sensitive-category gate** — Hard rule by account / age band * **Personalization fairness** — Group fairness on availability and surfaced price * **Diversity scorer** — Recommendation-set diversity to avoid filter-bubble effects * **Conversion-correlation watch** — Soft drift signal; recommend regenerating evals if conversion drops ## 3. Customer service chatbot **Setup:** Order status, returns, exchanges, escalation. Order data + policy data + customer history. **Why it's hard:** * Authoritative policy must be reflected (return windows, restocking fees) * Refund-promise accuracy — bot promises a refund the policy doesn't allow * Multi-channel consistency — same answer in chat, email, voice **Stakes:** * Refund-promise mismatch → regulatory and reputational exposure * CCPA / GDPR data subject requests must route correctly * FTC focuses heavily on dark-pattern returns/cancellation flows **Stratix evaluation criteria:** * **Policy citation rule** — Every refund/return decision cites the relevant policy section * **Refund-promise gate** — Hard rule: bot cannot promise refunds outside policy parameters * **DSR routing** — Privacy requests recognized and routed to the right intake * **Tone and de-escalation** — Judge for customer-experience quality * **Channel parity** — Same query produces same answer across chat/email/voice ## 4. Dynamic pricing and promotion **Setup:** AI surfaces pricing or promotion suggestions. Human merchandiser approves; AI does not autonomously change prices. **Why it's hard:** * Personalized-pricing legal exposure — if it correlates with protected class, it's discriminatory * Bait pricing — promotion that's never actually realized * Cross-region price-parity expectations **Stakes:** * FTC and state AG actions on personalized pricing have multiplied in 2025-2026 * Class-action exposure under state UDAP statutes * Marketplace platform-policy violations carry delisting risk **Stratix evaluation criteria:** * **Personalized-price fairness** — Group fairness on offered price across protected proxies * **Realizability rule** — Hard rule: any "promotion" must be actually achievable, not bait * **Cross-region parity** — Code scorer flagging unjustified regional spread * **Audit trail** — Every price suggestion logged with inputs, model version, approver ## 5. Returns and fraud detection **Setup:** AI scores returns for fraud markers (wardrobing, empty-box returns, serial returners, organized retail crime patterns). Outputs go to fraud team for action. **Why it's hard:** * False positives wrongly accuse legitimate customers and drive churn * Disparate impact on legitimate but high-return categories (apparel, gifts) * Privacy boundaries on what the score can use as features **Stakes:** * Retail loss to returns fraud reaches multi-billions annually across the industry * Wrongful denials drive class-action exposure * State consumer-protection AGs investigate algorithmic denial of returns **Stratix evaluation criteria:** * **Top-decile precision** — Tracked against confirmed-fraud labels * **Fairness scorer** — Group-level denial rate parity * **Reason-code coverage** — Hard rule: every flagged return cites contributing factors * **Customer-explanation rule** — Hard rule: customer-facing denial reason is plain-language * **Override rate watch** — Drift signal if humans override AI more than threshold ## See also * [Retail eval patterns](/more-in-this-section-2/eval-patterns-10.md) * [Retail compliance](/more-in-this-section-2/compliance-10.md) * [Trace evaluations (SDK)](/more-in-this-section-9/trace-evaluations-1.md)