When Should an AI Agent Hand Off to a Human? Escalation Rules for Ecommerce Support

The short answer. An AI agent should hand off when it cannot do the job, should not do the job, must not decide, or the customer asks for a person. In practice that means nine rules you set before launch: the customer asks, the same problem repeats, the customer is upset, money goes outside policy, identity cannot be checked, the data shows a real problem, the topic touches health, safety, legal or privacy, the agent does not know, and an action cannot be undone. Then set rules for the handoff itself: what the customer is told, what travels with the conversation, who owns it and when. Test each rule with real examples before customers see it.

Disclosure: Rep AI publishes this guide and sells an AI agent for Shopify. The rules below are written for you to set in any product. They describe what a merchant should decide, not how any one product behaves. One customer comment appears near the end, labeled and linked; it comes from Rep AI's own published case study.

Last reviewed 2026-10-16. This guide describes a method. It makes no claim about how accurate any agent is, because accuracy depends on your policies, your data and your questions.

Why handoff rules matter

A wrong answer from an AI can do more damage than a slow answer from a person. A customer stuck with an agent that will not let go is a second risk. The risk is not only a wrong answer. It is an agent that keeps going when it should stop.

So write the stopping rules down. Do it before launch, in plain words your team can read, and give each rule an owner. A rule you cannot state clearly is a rule you cannot test.

Each rule below has the same shape: a trigger, an action, and a message to the customer. Where a number appears in an example, it is invented for method only. Choose your own after reading your transcripts.

The nine rules

1. The customer asks for a person

Rule. If a customer asks for a human, hand off. Do not ask them to justify it or try one more answer first.

Set a list of phrases in your customers' own words ("agent", "talk to someone", "real person", "call me"). If nobody is available, the agent says when someone will reply and collects a way to reach the customer.

2. The same problem comes back

Rule. If the customer repeats or rephrases the same question after an answer, or the conversation goes round in circles, hand off.

An example setting, invented for method only: after two unsuccessful attempts on the same issue, escalate. Tune it against your own conversations. A customer who writes the same thing twice may be telling you the answer failed.

3. The customer is upset

Rule. If the customer shows strong frustration, makes a complaint, or mentions a dispute, a chargeback, a review or a legal step, hand off.

List the signals you want caught. A calm message that says "just checking" about a delayed parcel is not a complaint, so test that your rule tells the two apart.

4. Money goes outside your policy

Rule. Refunds beyond your window or above a limit, goodwill credits, price adjustments and discount exceptions go to a person.

Put the limit in writing. An example, invented for method only: any refund above a figure you choose goes to a person even if it is inside the window. The agent may explain policy. It should not make exceptions.

5. Identity cannot be verified

Rule. If the agent cannot confirm who it is talking to, for example because the order number and email do not match, it shares no order details and changes nothing. It hands off with a note that identity was not confirmed.

This protects your customers' data. Decide in advance what counts as verified.

6. The data shows a real problem

Rule. When order or carrier data shows a problem, hand off with the facts attached.

Examples to list: a parcel marked delivered that the customer says did not arrive, a shipment with no scan for longer than your carrier's normal window, a return to sender, an address error, a payment or fraud flag on the order, or an item that turned out to be unavailable. The agent can state what the data shows. A person decides what to do about it.

7. Health, safety, legal and privacy topics

Rule. Questions about medical conditions, allergies, reactions, injury, safe use, product recalls, legal threats, accessibility needs and personal data requests go to a person. The agent gives no advice on them.

This matters more in categories where a wrong answer can hurt someone: supplements, skincare, pet products, children's items and anything ingested or applied. State the rule in the agent's instructions and test it with questions that sound innocent ("is this safe if I am pregnant?").

8. The agent does not know

Rule. If the answer is not in your knowledge sources, or two sources disagree, the agent says it is not sure and hands off. It does not guess.

Keep a log of these questions and review it weekly. Each one is either a policy answer to write or a topic that stays with people.

9. The action cannot be undone, or the customer is special

Rule. Actions that cannot be reversed, or that touch accounts you handle with care, need a person or an explicit confirmation step.

Examples to consider: canceling after fulfillment has started, changing an address after a label is printed, closing a subscription where you have a retention process, orders flagged for fraud review, wholesale accounts and high-value orders. Decide which of these the agent may do alone, which need the customer's confirmation, and which go to a person.

The rules for the handoff itself

A good trigger with a bad handoff still fails the customer. Set these as well.

• Tell the customer. The agent says it is passing the conversation to your team and what happens next. No silent switches.
• Attach the conversation. Your team should see what was said, the order number, who the customer is, whether identity was confirmed, the reason for the handoff and anything the agent has already promised. The customer should not have to start again.
• Name the owner. Say who picks it up: a queue, a role or a person. An escalation nobody owns is a dead end.
• State the hours and the window. Say when a person will reply, and promise a window only if you can meet it. Outside staffed hours, the message says so.
• Leave no dead ends. If the handoff cannot be completed, give the customer another route, such as an email address or form.
• Keep the thread with the person. Once a person has taken over, the agent should not keep replying in the same thread. Ask any vendor whether the product can do this.
• Close the loop. Tag handed-off conversations by reason and review them weekly. A repeated reason is either a new thing to automate or a policy gap.

An illustration of handoff wording, to adapt to your own hours and tools (not a customer conversation):

I am passing this to our team so they can sort it out properly. They can see our conversation. You will get a reply by email, and we aim to respond within one business day.

How to test your rules

1. Build a test set. Write at least one realistic conversation for each rule, using your own past customers' wording. Add near misses: messages that look like triggers but are not, and messages that are triggers but sound polite.
2. Test both directions. Count handoffs that should have happened and did not, and handoffs that happened and should not have. Early on, treat the first kind as the bigger risk: a missed handoff leaves a customer who needed a person with an agent.
3. Start strict, then loosen. Begin with more handoffs than you expect to need. Relax a rule only when transcripts show the agent handles that case correctly.
4. Count by reason. A total handoff rate tells you little. A low rate can mean a healthy agent or one that is overreaching. The reasons tell you which.
5. Re-test after changes. When you change a policy, a knowledge source or an agent setting, run the set again.

For how to run a wider test with success criteria and a comparison group, see How to Run an AI Sales Agent Proof of Concept on Shopify: Success Criteria, Holdouts and Mobile Rules. To connect handoff to the helpdesk you already use, see How to Add an AI Agent to Gorgias, Zendesk or Kustomer Without Replacing Your Helpdesk. For the wider question of what to automate, see AI Customer Service for Shopify Stores: The Complete Guide to What to Automate and What to Hand Off.

Rules checklist

One store's published view

VIBAe sells footwear to more than 150 countries, and the whole business is run by a team of fewer than 10 people. Rep AI's case study on VIBAe says its co-founder explained that returns, exchanges and where-is-my-order questions grew with the business. Hannu Salminen, Co-founder at VIBAe, said in the case study: "I would say our customer care team is happier today. It's not that hectic anymore. They can focus on the more difficult issues."

That is one store's account, as Rep AI's case study reports it. It does not say which of the rules above to set, and it is not a statement about how any agent decides to hand off. It points at what the rules are for: keeping people free for the cases that need them.

Where Rep AI fits

Rep AI escalates to your team with the conversation attached, and it integrates with helpdesks including Gorgias, Zendesk, Kustomer, Freshdesk and Richpanel. This guide does not describe which of the rules above can be set in Rep AI or how, and you should not assume any product supports all nine. Ask any vendor, Rep AI included, to show each rule working on your store before launch. You can try conversations in the simulator at /simulator, or start the 14-day free trial, which needs no credit card. Rep AI also has its own helpdesk, Rep Inbox, if you want the AI and your agents working in one inbox.

FAQ

When should an AI chatbot escalate to a human agent? When it cannot do the job (no data, cannot verify who it is talking to, does not know), should not do it (outside your policy), must not decide (money outside policy, health, safety, legal or privacy topics), or the customer asks for a person. Add triggers for repeated failure and for upset customers.

How do I set up human handoff for an AI agent on Shopify? Write the rules first, choose where escalations land and who owns them, decide what the customer is told and what travels with the conversation, then test each rule with real examples. How you configure the rules depends on the product, so ask your vendor to show each one working on your store.

Which AI agent can resolve tickets safely and hand off to humans? Safety comes from the rules you set and test, not from a product name. Any vendor can say it hands off. Ask to see each of your rules triggered on your own questions, and compare agents on that. Rep AI is one option and escalates to your team with the conversation attached. This guide makes no claim about accuracy.

Should the AI tell customers it is an AI? Yes. Say so, and make it easy to reach a person. Some places have rules about this, so check the ones that apply where you sell.

What if nobody is available when the agent hands off? The message should say that, say when a person will reply, and offer another route. Do not promise a window your team cannot meet.