# AI Refusals & Human Handoff | onmsg > See what happens when onmsg URL: https://onmsg.app/guide/what-happens-when-the-ai-cant-answer/ Last-Modified: 2026-09-08 Decision guide # When the AI Can't Answer: Refusals and Handoff See what happens when onmsg's AI can't answer: a configurable refusal message, handoff after a set number of refusals, turn limits, timeouts, and provider errors. Published September 8, 2026 · 5 min read ![A chat conversation escalating from an AI agent to a person](/images/featured/modern-3d-illustration-of-a-chat-conversation-esca.webp) ## The moment that decides whether AI chat works Every AI chat setup is judged on what happens when it does not know. Answer confidently and wrongly, and you have created a problem your team has to clean up. Say nothing useful, and you have wasted the visitor’s time. Refuse honestly and route them to a person, and you have done better than most websites manage. The Grounded AI Agent & Knowledge Base [/features/grounded-ai-agent/ →](/features/grounded-ai-agent/) treats that moment as something you configure rather than something you hope about. ![An escalation paths diagram showing refusals, turn limit, timeout and provider error routes](/images/content/escalation-paths-diagram-showing-refusals-turn-lim.webp) ## The refusal message In Strict mode, when the agent checks for relevant knowledge and finds none, it can return a configured refusal without calling the answer model at all. What that refusal says is entirely yours to write. Three things separate a good refusal from a bad one. It names the limit rather than being vague: “I don’t have confirmed pricing for that” beats “I’m not sure I can help with that”. It does not apologise repeatedly, because repeated apology reads as evasion. And it offers a concrete next step: a person, a form, a phone number, a callback. Assume visitors will see this message more often than you expect, particularly in the first weeks while your knowledge base is still thin. Write it as a real sentence from your business. ## The four escalation triggers **Refusal threshold.** You set how many refusals should trigger a handoff. Two is a sensible default, one refusal is a gap in the knowledge base, three in a row is a bad experience being repeated. **Turn limit.** Each agent has a maximum number of turns. Long conversations that are not converging usually need a person, and the turn limit is what notices. **Timeout.** If a response does not arrive in a reasonable window, the conversation escalates rather than leaving a typing indicator running. **Provider error.** When the AI provider fails, the conversation routes to a person instead of erroring at the visitor. This is also why AI error and status information exists in the workspace. You find out from your dashboard rather than from a complaint. ## What the handoff actually does ![A visitor seeing a refusal message followed by an offer to reach a person](/images/content/modern-3d-illustration-of-a-visitor-seeing-a-refus.webp) A handoff stops bot ownership. That phrase matters: from that point the conversation belongs to a person, and the bot does not interject over the top of a colleague mid-sentence. The conversation moves into the Shared Team Inbox [/features/shared-team-inbox/ →](/features/shared-team-inbox/) with its history intact, the visitor’s original question, whatever the AI said, and anything a Form or One field step collected earlier. Whoever picks it up reads the context instead of asking the visitor to start again, which is the single thing visitors most resent about being escalated. ## When nobody is there Escalation assumes someone to escalate to, and evenings exist. Handle it explicitly rather than hoping. Set business hours, a timezone and after-hours behaviour on the widget. Branch the flow with a By availability step so the offline path says plainly that nobody is around and collects details instead. Offline message intake keeps the enquiry in retained conversation history, so it is waiting in the inbox in the morning rather than lost. This is also the honest caveat about handoff generally: what happens after one depends on human availability. The product can route a conversation to your team reliably. It cannot make someone be there. ## What to configure before you go live Five things, ten minutes: 1. Grounding mode per agent, Strict anywhere prices or promises are involved 2. Refusal message, written properly 3. Refusal threshold for handoff, typically two 4. Maximum turns, so unproductive conversations reach a person 5. Business hours and an offline path that collects details With those set, the worst case for a visitor is an honest answer and a route to someone who can help. That is a considerably better floor than most website chat manages. ## Writing the refusal, with examples The difference between a refusal that keeps a visitor and one that loses them is about twenty words of care. **Weak:** “Sorry, I don’t have information about that.” Vague, apologetic, and offers nothing. The visitor is left to work out what to do next. **Better:** “I don’t have that confirmed, pricing depends on a few details I can’t check from here. I can pass this to someone who can give you an exact figure. Want me to?” It names the limit, explains why it exists, and offers a specific next step in the form of a question the visitor can answer with one tap. Three rules behind that. Name the boundary rather than apologising for it, people trust a system that knows its own edges. Do not apologise more than once. And always end with an action, because a refusal without a route is just a closed door. ## Watching the escalation rate Once you are live, the number worth tracking is how often conversations escalate and why. A high refusal rate in week one is normal and self-correcting, every refusal names a missing knowledge source. If it is still high in month two, the knowledge base has a real gap, or your retrieval strictness is set tighter than your content supports. Turn-limit escalations tell a different story: conversations are going long without resolving, which usually means the agent is answering adjacent questions rather than the one being asked. Timeouts and provider errors are infrastructure signals rather than content ones. If they cluster, check AI error and status information rather than rewriting your knowledge. Read next: how to reduce AI chatbot hallucinations [/guide/how-to-reduce-ai-chatbot-hallucinations/ →](/guide/how-to-reduce-ai-chatbot-hallucinations/) , or how the shared inbox routes and assigns conversations [/guide/how-the-shared-inbox-routes-and-assigns-conversations/ →](/guide/how-the-shared-inbox-routes-and-assigns-conversations/) . ## Learn more about Grounded AI Agent & Knowledge Base An AI agent that answers from your own knowledge base, with grounding modes, configurable refusals, and handoff to a person. Read the feature page [/features/grounded-ai-agent/ →](/features/grounded-ai-agent/) FAQ ## Questions people ask about this ### What does the visitor see on a refusal? The refusal message you configured, followed by whatever next step you set up, usually an offer to reach a person. You write the wording, so it can sound like your business rather than a system error. ### When does a conversation hand off? After the number of refusals you configure, and also on turn limits, timeouts and provider errors. A visitor can ask for a person directly too, and a flow can escalate on any branch you choose. ### What if nobody is available? Set business hours and after-hours behaviour on the widget and branch with By availability in the flow, so the offline path collects details as an offline message. The enquiry stays in retained history for your team to pick up. ## Related guides ### How to Add Your Business Knowledge: Paste, Upload and URL Import Supply knowledge by pasting text, uploading PDF/DOCX/TXT/Markdown/CSV, or importing one public web page. See processing status, reindex and edit sources. Read guide [How to Add Your Business Knowledge: Paste, Upload and URL Import →](/guide/how-to-add-your-business-knowledge/) ### How to Reduce AI Chatbot Hallucinations with Grounding Controls Why chatbots invent facts, and how strict retrieval gating, refusal thresholds, handoff and source attribution help reduce it, controls that guide behaviour. Read guide [How to Reduce AI Chatbot Hallucinations with Grounding Controls →](/guide/how-to-reduce-ai-chatbot-hallucinations/) ### Strict, Balanced and Open Grounding Modes Compared Compare onmsg's grounding modes: Strict (refuse when no relevant content), Balanced (prefer your knowledge), and Open (general conversation), plus when each fits. Read guide [Strict, Balanced and Open Grounding Modes Compared →](/guide/strict-balanced-and-open-grounding-modes-compared/) ### Website Data Privacy: What Visitors' Chat Data Means for GDPR and CCPA What chat and enquiry data is collected, how retention settings differ by plan, and general GDPR/CCPA considerations for chat intake, guidance, not legal advice. Read guide [Website Data Privacy: What Visitors' Chat Data Means for GDPR and CCPA →](/guide/website-data-privacy-gdpr-and-ccpa/) ## Want to try this on your own site? No credit card required. Start free [https://chat.onmsg.app →](https://chat.onmsg.app)