Let a voice agent take product questions, warranty claims and returns: automate the four repeat questions, escalate every dispute
Retell AI 编辑团队（厂商博客，未署名个人）Retell AI Blog · Date not disclosed
Read originalSaved text / editorial summary
原始页面 retellai.com 本次直连可读（HTTP 200，已落盘 cache/auc_raw/retell_refunds_exchanges.html）。关于 Anker，原文只有一句：「Anker rebuilt global customer support on AI voice agents to handle product questions, warranty claims, and returns across multiple languages. The result was human-quality conversations at enterprise scale, without the staffing cost of running round-the-clock multilingual coverage.」同一页面的其余部分是一篇面向电商的通用退货自动化教程，给出八步做法：先建一个能做测试通话的 agent；拆成退款申请、换货申请、退款状态查询三条独立流程；连接电商订单接口做实时查询；把退货政策写成判定规则；加欺诈评分与高风险升级；生成预付面单并触发仓库工作流；用库存感知优先推荐换货；上线前跑 20 到 30 个场景模拟。文中提醒旺季并发要在 12 月 26 日之前扩容，并明确损坏争议、国际海关争议、B2B 部分退货与涉及拒付的案件仍属人工工作。
Business model
Consumer electronics brand Anker rebuilt its global customer support on AI voice agents handling product questions, warranty claims and returns in multiple languages - one sentence in a Retell AI blog post, with no numbers and no implementation detail. The same public tutorial gives a skeleton worth copying: build three short flows (refund, exchange, refund status) instead of one giant one, wire live order lookup, encode the returns policy as decision rules (window in days, category exceptions, return-frequency thresholds), escalate serial-return and wardrobing signals, issue a prepaid label the moment a return is approved and notify the warehouse, and offer an exchange before a refund. What this saves is peak-season repetition, not the whole support function: damage disputes, international customs, partial B2B returns and anything touching a chargeback stay with humans.
The asset is knowing which part to automate. Repeatable, rule-decidable questions go to the agent; anything adversarial or ambiguous goes to a person. The turning point is the last mile - if the agent cannot issue a label and tell the warehouse in the same call, it is a talking FAQ, not automation.
Workflow
- Pull the top ten return questions from last quarter's tickets; build only for those.
- Build three short flows - refund request, exchange request, refund status - each with greeting, lookup, policy check, branch and confirmation.
- Wire live order lookup with a timeout of at least five seconds, plus a fallback for callers who cannot find their order number.
- Encode the returns policy as decision rules: purchase date against window, category against exceptions, return count against the human-review threshold.
- Run risk signals in parallel and escalate hits to a human with full context, without telling the caller they were flagged.
- On approval, generate the prepaid label immediately and fire a webhook to the warehouse or 3PL with order number, expected items and reason code.
- Check inventory first and offer an exchange before falling through to a refund.
- Simulate 20 to 30 scenarios before launch, and raise concurrent-call capacity before the peak - the source says before December 26 - not after it.
Adaptation plan
The Anker mention carries no metrics, no volumes, no staffing numbers and no tooling detail; it is a vendor's one-line case reference. The rest of the page is a generic tutorial whose figures (conversion of refunds to exchanges, setup timelines) are the vendor's own claims and are untested here. Workflow, prompt and pitfalls are AgentFlow's proposals, not measurements.
Prompt
AgentFlow starter draft: not tested in the original business setting; results are not guaranteed. Synthetic examples do not verify the author’s revenue.
You are an ecommerce support-flow design assistant. Turn my returns policy into executable decision rules and three conversation flows. You do not make business decisions and you do not loosen my policy. [INPUT] Returns policy text: [paste] Order fields available: [order number, order date, delivery date, category, amount, return count] Top questions: [the ten most frequent from last quarter] Always escalate: damage disputes, international customs disputes, partial B2B returns, anything involving a chargeback [EXECUTION STEPS] 1) Break the policy into a condition table: condition, action on match (approve / decline / escalate), what the customer hears, any alternative to offer. Flag anything the policy does not cover; do not invent rules. 2) Design three separate flows - refund request, exchange request, refund status - each with greeting, lookup, policy check, branch and confirmation, and a target number of turns. 3) List the fields each flow must collect and the fallback order when a field is missing, including callers who cannot find an order number. 4) Produce a separate escalation list: trigger, context to attach before transfer, and what the customer hears. 5) Finish with a policy-gaps list: cases the written policy does not cover. [OUTPUT] - Condition table with columns: condition / action / customer wording / alternative / policy reference. - Three flows written step by step with required fields per step. - Escalation list with triggers and wording. - Policy-gaps list; no invented rules. [HUMAN CONFIRMATION] - The condition table must be approved before flows are designed. - Anything that loosens the policy or offers compensation needs my explicit written approval. - Chargebacks, damage, cross-border and B2B cases must be hard-coded to escalate, never auto-approved. [SYNTHETIC EXAMPLE - not real data, format only] Input: policy is 30 days from delivery, intimate categories non-returnable, more than 6 returns in 90 days requires review; order fields include order date and category. Output (extract): | Delivered 47 days ago, past the 30-day window | Decline | This arrived 47 days ago and our window is 30 days; I can offer one-time store credit instead | Alternative needs supervisor approval | Policy 2.1 | | Category is intimate apparel | Decline | This category is not returnable under our policy; I can show you similar styles | None | Policy 3.4 | | 7th return within 90 days | Escalate | Let me get a colleague to look at this order with you | Attach order number and return count | Risk threshold | Policy gaps: 1) whether bundled gifts must be returned is not stated; 2) who pays exchange shipping is not stated.
Running costs and pitfalls
The Anker mention carries no metrics, no volumes, no staffing numbers and no tooling detail; it is a vendor's one-line case reference. The rest of the page is a generic tutorial whose figures (conversion of refunds to exchanges, setup timelines) are the vendor's own claims and are untested here. Workflow, prompt and pitfalls are AgentFlow's proposals, not measurements. No independent cost benchmark is available. Check subscriptions, permissions and external actions before running a workflow.
