--- title: "What Questions Should an AI Concierge Answer vs Escalate?" description: "Which AI concierge use cases should you answer from cited pages, and which must escalate? A fail-closed decision tree for lean ecommerce and SaaS operators." url: "https://articles.flowcrews.com/what-questions-should-an-ai-concierge-answer-vs-escalate" category: "Concierge" date_published: 2026-09-30 reading_time_minutes: 6 word_count: 1449 --- # What Questions Should an AI Concierge Answer vs Escalate? *A fail-closed rule for cited-page concierges: answer, clarify once, or hand off.* ![What Questions Should an AI Concierge Answer vs Escalate?](https://pub-07fb5e4955ba485b822d6b388be96d9a.r2.dev/2e474d51-a072-472e-ab2f-1a43a8990566/what-questions-should-an-ai-concierge-answer-vs-escalate/hero-7233e060-199b-4c4e-8836-26cb910b2eb7.jpg) **TL;DR:** - Answer only repeatable, low-stakes questions you can ground in an approved page or connected record. - Containment is not resolution: Ada's gap is about 72% contained versus 52% actually resolved. - Only 24% of consumers in a March 2026 Ada/NewtonX study said AI fully resolved their latest service chat. - Hard-escalate refunds, disputes, legal, security, safety, exceptions, and missing knowledge — confidence does not override those topics. - Clarify once, then hand off with a case packet instead of looping the customer. ## Containment Is the Wrong Score ## Questions Worth an Answer, Not a Guess ## What Must Escalate No Matter the Confidence Score ## Clarify Once, Then Answer or Escalate ## Send the Case Packet and Log the Refusal ## Conclusion - True resolution, not containment — Score the concierge on whether the customer was actually helped, and refuse a 100% automation target before you classify intents. - Answer only the safe band — Reply to repeatable, low-stakes questions when an approved page or connected record actually says the answer, and do not guess past that contract. - Hard-escalate without a confidence override — Refunds, safety, complaints, legal or accessibility issues, emergencies, and SaaS account, billing, and security exceptions always go to a person. - Clarify once, then decide — Ask one question about an in-scope ambiguous referent, then answer or escalate; a second clarification round, a sentiment drop, or “this isn’t helping” is a handoff. - Send the case and log the gap — Hand off a five-part packet so the customer does not repeat themselves, and store every refusal with the exact question, the missing page or record, and the reason. Map your live intents against the answer, clarify-once, and hard-escalate rules before you raise the automation target. ## Frequently Asked Questions ### What should an AI concierge answer without a human? Repeatable, low-stakes lookups it can cite from an approved page or connected record: product facts, policies, hours, shipping, order status, and returns that follow a written rule. If the source does not cover the question, the safe move is a handoff, not a guess. ### Which topics must always escalate? Hard-route refunds, billing disputes, safety issues, formal complaints, legal or accessibility requests, security issues, policy exceptions, VIP accounts, and any explicit request for a person. Those topics need judgment and carry liability, so a high confidence score is not an override (When Should AI Hand Off to a Human? — Conduit). ### Why is containment a bad success metric for an AI concierge? A contained chat can still come back as a repeat contact, a call, or a complaint. Ada's own figures show about 72% containment against about 52% automated resolution, and a telecom case fell from 78% containment to about 41% true resolution once further contact was counted. ### What confidence score should trigger a handoff? A practical hospitality starting floor is to route any response the AI scores below **70–75%** confidence rather than inventing a policy (When Should AI Hand Off to a Human? — Conduit). Tighten that floor on regulated or financial topics, and calibrate it against your own transcripts. ### How many clarification attempts are too many? One. After two or more bot clarification rounds, **64%** of customers report dissatisfaction, versus **71%** satisfaction when the bot resolves on the first message (Chatbot Containment Rate Statistics 2026 — Stealth Agents). Escalate after two failed attempts or two negative turns. ### What should a good AI-to-human handoff include? A case packet: the transcript, a short summary, the escalation reason, the sources already cited, and what was already tried. Log the gap afterward so the unanswered question can become a page or a policy, not the next customer's loop. ## Sources - [https://www.getmacha.com/blog/ai-customer-service-resolution-rate-statistics](https://www.getmacha.com/blog/ai-customer-service-resolution-rate-statistics) - [https://enderturing.com/blog/ai-chatbots-in-customer-service-the-containment-rate-thats-lying-to-you](https://enderturing.com/blog/ai-chatbots-in-customer-service-the-containment-rate-thats-lying-to-you) - [https://www.ringly.io/blog/ai-voice-support-vs-human](https://www.ringly.io/blog/ai-voice-support-vs-human) - [https://stealthagents.com/research/customer-support-chatbot-containment-statistics-2026](https://stealthagents.com/research/customer-support-chatbot-containment-statistics-2026) - [https://www.ecommerce-nation.fr/service-client-ia-40-conversations-automatisees](https://www.ecommerce-nation.fr/service-client-ia-40-conversations-automatisees) - [https://www.conduit.ai/blog/the-escalation-question-when-should-ai-hand-off-to-a-human](https://www.conduit.ai/blog/the-escalation-question-when-should-ai-hand-off-to-a-human)