# What Should an AI Agent Hand Off to a Human? A good AI agent needs to know more than how to answer. It needs to know **when it should stop answering**. That sounds simple. In a real customer conversation, it can be one of the most important parts of the design. A customer may ask for a human. A request may require an exception. The system may lack the information needed. A transaction may need approval. A complaint may require investigation. The AI may understand the words perfectly but still not be the right party to handle the issue. That is where human handoff comes in. [ZazzyAgent Blog](http://blog.zazzyagent.com) already has detailed implementation documentation for assigning conversations to individual team members or roles, so this article focuses on the **decision behind the handoff**, not the configuration itself. [How to Hand Off AI Conversations to Human Agents](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) ## Human handoff is not the opposite of automation It is easy to think of the system like this: > AI = automation > Human = manual work That misses an important part. A well-designed automation system can include: **AI** → handles what it can **Automation** → handles what is predictable **Human** → handles what requires judgement The handoff is part of the automation. The question is: > **What should cause the transition?** ## There are six common handoff triggers A useful starting framework is: ### Explicit request The customer asks for a person. ### Missing capability The AI doesn't have the information or tool needed. ### Exception The customer needs something outside normal policy. ### Human authority A decision requires approval. ### Sensitive situation The conversation needs judgement, empathy or investigation. ### High-value transition The customer is ready for a person who can complete the next commercial step. ZazzyAgent's current human-handoff guidance covers these kinds of boundaries, including explicit human requests, complaints, exceptions, unavailable information and manual approval. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) # 1\. When the customer explicitly asks for a human This is the easiest case. Customer: > “Can I speak to someone?” The AI shouldn't respond with: > “Sure, but first answer these seven questions.” The handoff itself is the customer's request. A better approach is: > “Sure. I'll connect you with our team.” Then provide any useful context already available. ZazzyAgent supports assigning conversations to a specific team member or team role. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) ## Why this matters The customer has already told you what they want. Making them pass another AI screening process creates friction. # 2\. When the AI doesn't have the required information Consider: > “Can you tell me why my payment was reversed?” If the agent doesn't have: **payment data** **transaction details** **support information** then it shouldn't guess. The next step could be: > ask for a missing identifier or: > retrieve information through an API or: > hand the conversation to support. The right choice depends on whether the missing information can be obtained automatically. ZazzyAgent's HTTP API guidance specifically emphasizes using external APIs for live business information and avoiding invented results when an API fails. [HTTP API Actions](https://blog.zazzyagent.com/ai-agent-http-api-actions-zazzyagent) ## Don't hand off just because the agent doesn't know immediately There is an important middle step. Customer: > “What's my order status?” Agent doesn't know the order number. Bad: > “I don't know. I'll send you to support.” Better: > “Sure. What's your order number?” The distinction is: > **Missing information that the customer can provide isn't necessarily a human-handoff case.** # 3\. When an external system can solve the problem Suppose: > “Can you check whether my appointment is available tomorrow?” The agent doesn't need a human. It needs live availability. If the booking system is connected, the agent can retrieve it. This is where tools should come before escalation. A useful hierarchy is: ```text Customer request ↓ Can AI answer from knowledge? ├── Yes → Answer │ └── No ↓ Can the agent retrieve it? ├── Yes → Use tool/API │ └── No ↓ Can the customer clarify? ├── Yes → Ask └── No → Human ``` That keeps the human team from becoming the default fallback for every unknown. # 4\. When the customer requests an exception Imagine: > “Your return policy says 14 days, but I'm on day 19. Can you still approve it?” The AI may know the policy. It may even know why the customer is asking. But if exceptions require human approval, the AI should not invent authority. A better route: > Explain the normal policy. > Collect any relevant information. > Hand the case to the appropriate person. This pattern is particularly useful in: **refunds** **discounts** **billing exceptions** **cancellations** **contract changes** **special service requests** The exact rule should come from the business. # 5\. When the customer is angry Not every angry customer requires a human immediately. A simple complaint such as: > “The delivery is late.” may still be resolvable. But: > “I've contacted support three times and nobody has fixed this.” is different. The problem now involves: **history** **frustration** **potential failure in the process** **possible need for investigation** The AI can gather the relevant context, but a human may need to take ownership. ZazzyAgent's human-handoff guidance specifically identifies complaints requiring investigation as a useful handoff trigger. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) # 6\. When a decision needs human authority This is different from not having enough information. The AI may know everything needed. It just isn't authorized to make the decision. For example: > “Approve a 30% enterprise discount.” If discounts above a certain threshold require a manager, the AI should not bypass that rule. It can: **collect the requirements** **prepare the context** **recommend the appropriate route** then: **request human approval** The mistake would be assuming: > “The AI knows enough, therefore it is allowed to decide.” Knowledge and authority are separate. # 7\. When the customer is ready for a human sales conversation This is an interesting case because human handoff isn't caused by failure. The AI may have done its job perfectly. Customer: > “Yes, I've seen the pricing. We have 200 users and want to start next month. Can someone call me?” At this point, continuing to ask questions may create more friction. The AI has helped qualify the lead. Now sales can take over. ZazzyAgent's human-assignment capability can route the conversation to a configured sales member or role. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) ## The best handoff contains context One of the worst handoffs looks like: > “Customer wants to speak to sales.” The salesperson now asks: > “What are you looking for?” The customer repeats everything. A better handoff carries useful information. For example: **Company:** ABC Ltd **Employees:** 200 **Requirement:** WhatsApp automation **Timeline:** Next month **Intent:** Ready for sales call Now the human starts from context. ZazzyAgent's handoff documentation supports adding internal notes before the conversation is assigned. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) ## Don't interrogate before handing off There is another common mistake. Customer: > “I want to speak to a person.” AI: > What's your company? > What's your budget? > What's your use case? > How many employees? > What is your expected timeline? > What are your current tools? That may be useful in some sales journeys. It can also be exactly the wrong response. A simple rule: > **Only collect information before handoff if it materially helps the person taking over.** And don't make the customer earn access to a human. ZazzyAgent's Instagram handoff guidance makes the same distinction: useful context can be collected before transfer, but the AI should not interrogate a customer who has already explicitly requested a person. [Instagram Human Handoff](https://blog.zazzyagent.com/instagram-ai-human-handoff-zazzyagent) # 8\. What should happen after handoff? This is easy to overlook. Once the human owns the conversation, the AI should know that responsibility has changed. Otherwise: Human: > “Hi, I'm Rahul from support.” AI: > “I can also help you troubleshoot this issue.” Now both are responding. That creates confusion. ZazzyAgent's handoff configuration includes explicit post-handoff behavior so the AI can stop acting as the conversation owner once a human takes over. [Human Handoff](https://blog.zazzyagent.com/ai-agent-human-handoff-zazzyagent) ## Handoff should be a state change Think about it as: ```text AI owns conversation ↓ Handoff condition met ↓ Context prepared ↓ Human assigned ↓ Human owns conversation ``` That's better than treating handoff as simply “send a message to support.” # 9\. Human handoff and AI-to-AI transfer are different Sometimes the next step shouldn't be a human. It might be another specialist agent. For example: > General Agent → recognizes order issue → Order Agent That is AI-to-AI transfer. Another situation: > Support Agent → recognizes billing dispute → Human Billing Team That is AI-to-human handoff. ZazzyAgent supports both mechanisms. [Transfer Conversations Between AI Agents](https://blog.zazzyagent.com/ai-agent-transfer-another-agent-zazzyagent) The decision is: > **Does another configured system have the right responsibility and authority, or does this need a person?** # 10\. Don't hand off every difficult question Over-escalation creates its own problem. Suppose the customer asks: > “What time do you open tomorrow?” The agent knows the answer. Sending it to a human wastes human capacity. Likewise: > “Do you deliver to Mumbai?” If the answer is in the agent's knowledge, there is no reason to transfer. A good system should be able to say: > **I can resolve this.** and: > **This needs someone else.** Both are necessary behaviors. # 11\. A useful handoff decision tree ```text Customer request ↓ Known and within scope? ├── Yes → Answer │ └── No ↓ Can clarification solve it? ├── Yes → Ask │ └── No ↓ Can a tool/API solve it? ├── Yes → Use tool │ └── No ↓ Can another specialist AI solve it? ├── Yes → Transfer │ └── No ↓ Human required ↓ Preserve useful context ↓ Assign ↓ AI stops ``` This is far more useful than a vague instruction: > “Transfer complicated questions to a human.” ## A good handoff has a reason The business should be able to answer: > **Why did this conversation go to a human?** Possible reasons: **Customer request** **Exception** **Complaint** **Approval** **Missing capability** **Sensitive issue** **High-value sales stage** That information can also be useful for later process improvement. If 40% of handoffs happen because the AI can't answer one particular question, perhaps the knowledge base or integration should be improved. If 40% happen because customers request a discount, perhaps the sales process needs a clearer path. Handoff data can expose business problems. ## The goal isn't to eliminate human conversations That's not the right target. A better goal is: > **Make sure humans spend their time where human involvement creates value.** The AI can handle: **routine questions** **information retrieval** **initial qualification** **basic routing** **structured data collection** Humans can focus on: **exceptions** **judgement** **complex support** **negotiation** **sensitive conversations** **relationship-building** That is a more realistic model for customer automation. ## The takeaway The right question isn't: > “Can AI handle this conversation?” Ask: > **“Who is the right owner for the next step?”** Sometimes that's AI. Sometimes it's a tool. Sometimes it's another specialist agent. Sometimes it's a human. A well-designed system knows the difference.