An AI assistant that pushes on when it should step aside costs more than it brings in. The five situations where it should hand off, what the customer should see at that moment and what your team needs to receive so the customer doesn't have to repeat anything.
Isaac Poirier
Founder, Groupe Echo Inc.
An AI assistant should hand off to a human as soon as the conversation calls for a decision, a listening ear or information it doesn't have: an angry customer, a refund or an exception, a question its documents don't answer, a health, money or safety topic, and every time the customer asks to talk to someone. A good handoff announces what happens next, gives an honest timeframe and passes the whole conversation to your team.
On a Sunday evening, a customer asks on your website whether you do tune-ups on e-bikes: the assistant can answer that. But faced with "This is the third time I've brought my bike back, I want a refund," that same assistant becomes a risk if it carries on as if nothing were wrong.
To your customer, whatever the assistant writes on your website is your business speaking. A "more or less right" answer about a return policy or a warranty will be read as a commitment.
In 2024, in Moffatt v. Air Canada (2024 BCCRT 149, link checked on September 29, 2026), the British Columbia Civil Resolution Tribunal held Air Canada liable for incorrect information about its bereavement fare, given by the chatbot on its own website. A small business doesn't need to end up before a tribunal to pay the price: a misinformed customer means a lost sale, a negative review or an hour on the phone clearing up the misunderstanding.
A good handoff, on the other hand, costs your team a few minutes, at the right time, on the right conversation.
An upset customer isn't looking for information first and foremost: they want to be heard by someone who can act. The signs are often easy to spot: all caps, "again," "unacceptable," a threat to take their business elsewhere. Offering them a help article, even a relevant one, adds fuel to the fire.
Issuing a refund, making a goodwill gesture, admitting a mistake: these are decisions, not information. A landscaper can let their assistant explain their plant warranty as written. They shouldn't let it decide whether a customer's dead hedge is covered. As soon as a request falls outside the general rule, a human makes the call.
A reliable assistant answers based on what you've given it: FAQs, policies, service sheets. When the answer isn't there, it should say so plainly. "I don't have that information, but I can put you in touch with the team" is a good answer. A made-up answer that sounds confident is the worst kind.
Keep track of the questions that lead to this kind of handoff. If the same one comes up often, you're missing a document.
A physiotherapy clinic can let an assistant handle its opening hours, parking information and what documents to bring. But "My knee has hurt since I fell, should I go to the ER?" calls for a professional. The same goes for financing or taxes, and for safety: a smell of gas, a product used incorrectly. The assistant shouldn't give advice; it directs the person to someone, or to emergency services if the situation calls for it.
This is the simplest rule and the one most often broken. If someone writes "I want to talk to a real person," you don't ask them three more questions. You transfer them.
Then the main thing an assistant is likely to do is add a step between your customers and you. A notary's office or a contractor that mostly receives requests for custom quotes will often be better served by a well-designed form, a visible phone number and a clear response time. To sort this out before choosing, our article AI assistant or FAQ describes the cases where an assistant isn't useful.
At the moment of handoff, the customer should find four things.
Here's what a good handoff message looks like, whatever tool you use:
I'm passing your refund request for the September 3 repair on to our team. Someone from the shop will get back to you by noon tomorrow. You won't have to explain the situation again.
Only include a timeframe if someone can actually meet it.
So the customer doesn't have to start over, the person taking over needs:
Before turning on an assistant, write down on a single page who on your team receives handoffs, during what hours, and what response time you can promise. Without that page, even the best tool hands off into a void.
We build Echo Support. Here is what our tool does, limitations included, as checked in its code on September 29, 2026. The settings are detailed in the escalation documentation.
Before their first message goes out, the visitor provides their name and email address. The assistant answers based on the documents you upload and the published articles in your help centre. It is instructed to say it doesn't have the information when its search turns up nothing, and to offer a human when in doubt rather than guess. It transfers the conversation when the visitor asks to speak to someone or accepts the offer of a human, seems frustrated, asks a complex technical question or raises a billing issue. Visitors can also click "Talk to a person," below the assistant's latest answer: the conversation then goes to your team without the assistant replying.
Your team can also take over at any time: its first message in a conversation transfers it, and the assistant goes quiet.
After that, the assistant no longer replies in that conversation.
Test it from the widget: write "I want to talk to someone," or click "Talk to a person," and check that at least one person receives the email. Neither manually setting a conversation to "Needs human" nor a team member's reply sends it.
For a store connected through the Shopify connector, still a pilot by invitation, the assistant never carries out a request to cancel an order, change an address or make a return: it changes nothing in Shopify. When it finds the order from its number and its email, it transfers the conversation with an internal note that summarizes the order and links to it in Shopify. Otherwise, it asks the customer to check both, then offers a human if the order still can't be found.
In both cases, the assistant writes no transition message: the widget tells the visitor, in their language, that someone on the team will reply. The escalation email goes out, with the reason for the handoff.
There is no instruction specific to these topics, and no setting lets you add one. You can publish a help article explaining what to do, for example calling Info-Santé 811 in Quebec (page checked on September 29, 2026) or 911 in an emergency; the assistant may find it, with no guarantee that it will. Each team member can also create a keyword alert (Configuration → Notifications) to get an email when a visitor writes "emergency," "pain" or "lawyer." The alert warns your team; it doesn't stop the assistant.
Think about the data, too. The text of conversations handled by the AI goes to OpenRouter, in the United States, which routes it to DeepInfra (United States or Canada) or, failing that, Microsoft Azure (United States), and sensitive information is not masked before it is sent. If these questions come up often in your business, the "Enable AI" switch turns the assistant off, and your team answers every conversation itself. Our page on Quebec's Law 25 details what each subprocessor receives, and what is still sent when the assistant is off.
Echo Support's plans include a free plan to try the widget, without the AI assistant: your team answers every conversation itself.
An AI assistant only knows what your documents teach it. Returns, hours, warranties: what to include, how to write it so it's clearly understood, the traps to avoid, a way to test before going live and the habits that keep the knowledge base up to date.
Before installing an AI assistant, look at the questions your customers actually ask. Concrete criteria, a comparison table, hidden costs and four situations where an AI assistant isn't the right choice.