The best AI customer service doesn't try to handle everything. It handles what it should, asks before what it shouldn't do alone, and hands the customer to a person at the right moment without making them start over. Get the handoff wrong and customers feel trapped with a bot. Get it right and they barely notice the switch.
This guide covers the escalation process for an AI agent: the triggers that should bring in a person, the approval rules that keep consequential actions safe, the wording for a warm handoff, what to do when nobody is available, and the record you should keep of every step.
Escalation, handoff and approval are different things
- Escalation: The decision that a conversation needs more than the agent can give. See escalation.
- Handoff: Moving the conversation to a person, with the context they need. See human handoff.
- Approval: The agent keeps the conversation but asks for a yes before a specific action, like a refund. See approval workflow.
Mixing them up causes most handoff problems. If every refund triggers a full handoff, your team drowns in conversations that only needed a yes. If nothing triggers a handoff except a request for a human, upset customers get stuck.
Escalation triggers
Write your triggers down, in plain words, the same way you would brief a new hire. Most stores need some version of these:
| Trigger | Example | Why a person |
|---|---|---|
| The customer asks for a person | "Can I talk to a human?" | Always honor it. Refusing is the fastest way to lose trust. |
| Strong frustration | Repeated complaints, all caps, "this is the third time" | Empathy and judgment matter more than speed here. |
| Outside the rules | A refund past the return window for a loyal customer | Bending policy is a business decision. |
| Legal, safety or health | An allergic reaction, an injury, a lawyer, a chargeback threat | Risk is too high for any automated reply. |
| High value | A wholesale order, a large account, press | Relationships worth a person's time. |
| The agent is stuck | It tried twice and couldn't complete the request | Looping wastes the customer's time. |
| Missing information | The answer depends on something not written down anywhere | Someone has to decide, then write it down for next time. |
Good AI agents also escalate on their own judgment, not just on keywords. A customer who never says "human" but describes a damaged product that hurt their child needs a person immediately.
Approval rules: ask first, then act
Approvals let an agent keep helping while a person stays in control of what matters. Three questions define your rules.
- Which actions need a yes?: Anything that moves money, ends a relationship or changes the store: refunds, cancellations, credits, discounts, changes to settings. Looking things up never needs one.
- Who says yes?: Often the person who asked. If a customer asks to cancel, the customer confirms the cancellation. If a teammate asks the agent to refund ten orders, that teammate confirms. For exceptions to policy, a manager.
- What can it never do?: Some actions shouldn't be possible at all. Turning a capability off is safer than relying on an approval every time.
When the agent asks, it should say in plain words what it is about to do, and what the alternatives are: "I can cancel your subscription now, or skip your next two deliveries instead. Which would you like?" That is an approval and a retention offer in one sentence.
Warm handoff wording
A warm handoff means the person taking over knows the story, and the customer knows what is happening. Here is wording you can adapt.
When the agent brings in a person (chat or text)
"I want to make sure this is handled properly, so I've asked our team to step in. They can see everything we've talked about, so you won't need to repeat anything. You can keep messaging here."
When nobody is available
"Our team is offline right now and back at 9 a.m. Eastern. I've passed this along with the details, and someone will follow up. In the meantime, I can still help with your order."
When a teammate joins a call
Agent: "Sam from Northfield Coffee has joined the call." Teammate: "Hi Maya, I've got the details on the missing bag. Let me sort out the replacement for you." Notice the teammate opens with what they already know, not "how can I help?"
When the teammate hands back
"You're all set on the replacement. If you need anything else about your order, our assistant can help any time, and you can always ask for one of us."
What the teammate needs to see
A handoff is only warm if the person taking over can catch up in seconds. Make sure they get:
- Why the agent asked for help, in one line.
- The customer's profile, orders and subscriptions.
- The full conversation, on every channel, not just this one.
- What the agent already did, and what it offered.
- Any approval that is still waiting.
Who owns the conversation now
The most common handoff bug is two voices answering at once: the AI replies while a teammate is typing. Make ownership explicit. When a person takes over, the agent stops replying. When the person is done, they hand it back on purpose. And if a request for a person goes unanswered for too long, the customer shouldn't sit in silence; the agent should pick the conversation back up and say so.
The audit trail
Keep a record of every escalation, approval and action, with who did what and when. You will need it to settle disputes, coach the team and improve the agent. A good audit trail answers:
- What did the customer ask for, and what did the agent say?
- Which actions were taken, and who approved each one?
- When was a person asked to step in, who accepted, and how long did it take?
- Who had the conversation at each point, the agent or a teammate?
- How did it end?
Review a sample of handoffs each week. Each one is a lesson: a policy to write down, a trigger to add, or a case the agent can now handle alone.
Metrics to watch
- Handoff rate: The share of conversations that needed a person. Watch the trend, not a target.
- Time to accept: How long a customer waited between the request and a teammate taking over.
- Time to first human reply: Accepting isn't answering. Measure both. See first response time.
- Repeat contact: Customers who came back about the same problem after a handoff.
- Approval outcomes: How often approvals are granted, denied or changed. Lots of denials mean the agent is offering the wrong thing.
How Chargezen helps
Concierge agents are built to hand off well. Here is how it works today:
- Knows when to ask for a person: A customer asking for a human brings in your team, and the agent can also request a teammate on its own judgment, with a reason. Teammates you choose get alerted, including a push notification in the Finn app that shows the reason. See human handoff.
- Honest when nobody is around: If your team is off hours or nobody is available, the customer is told, with when the team is back if known, and the agent keeps helping in the meantime. If a request goes unanswered, it times out and the conversation returns to the agent.
- Clear ownership: When a teammate accepts the handoff or replies, the agent stops replying. Tap Hand back to return the conversation to it.
- Warm call handoffs: On calls, Concierge can ask a teammate to join. They listen in with the microphone off, tap Join the call, and the agent tells the caller who joined. Teammates can transfer to another teammate with notes from the Finn app and return the call to the agent. See phone calls.
- Approvals in the conversation: Consequential actions, such as cancelling a subscription, refunding an order, adjusting reward points or publishing a store change, always ask first. The person who made the request answers in their own words, on the same channel. You can also turn a capability off so an agent can't do it at all. See approvals.
- A full record: Every request, approval and action is saved in the conversation. A team-only line in the transcript notes who took over from the agent and when it went back.
Concierge Starter is $20 a month with 20,000 credits. Pro is $199 a month with 199,000 credits, or $1,990 a year. Phone and text charges are billed as used, and you can connect your own OpenAI or Anthropic account. See pricing.
Related terms: human handoff, human in the loop, approval workflow and audit trail. Related guides: conversational AI for ecommerce, how to share one business number with your team and AI receptionist vs answering service vs voicemail.
Frequently asked questions
When should an AI agent hand off to a human?
When the customer asks for a person, is clearly upset, needs an exception to your policy, raises a legal or safety issue, or when the agent has tried and can't complete the request.
What is a warm handoff?
A handoff where the person taking over already has the conversation history and context, and the customer is told what is happening, so nobody repeats anything.
Should every refund go to a person?
Not as a handoff. Use an approval: the agent keeps the conversation and asks for a yes before refunding. Save full handoffs for cases that need judgment.
What happens if nobody answers a handoff?
Tell the customer honestly, say when someone will follow up, and let the agent keep helping with what it can. Never leave the customer waiting in silence.
