Seven steps for putting AI on your support inbox and chat while keeping a person one click away, from the handoff rules to the numbers worth tracking.
Give the bot one fair attempt, then offer a person. That rule, plus a handoff that carries the conversation across, does most of the work of keeping customers on side. Customers start out wary: in a Gartner survey of 5,728 customers run in December 2023, 64% said they'd prefer companies didn't use AI in customer service at all, and 53% said they'd consider switching to a competitor over it. Their top worry was that AI would make it harder to reach a person.
The seven steps below are built around that worry, and none of them depend on a particular vendor.
1. Sort a month of tickets before the bot answers anything
Export last month's tickets and tag each one by what the customer wanted. You're looking for questions with one right answer that lives in a document or a system: order status, return windows, opening hours, password resets.
As an illustration, take a shop that gets 1,000 tickets a month. 410 ask where an order is, 160 are about the returns policy, 90 are password resets, and the other 340 are complaints, damaged goods, billing disputes and one-offs. Those first three groups, 660 tickets, are the bot's job. The 340 aren't, and the bot should recognise them and step aside.
The split also caps what success can look like. Even a bot that handles all 660 perfectly covers 66% of this made-up inbox. Treat any figure quoted before someone has seen your tickets as a guess. And if support isn't the biggest drain on your team's week, other business processes you can automate with AI may pay back sooner.
2. Write the handoff rule before the answers
The US Consumer Financial Protection Bureau has a name for what happens when this step gets skipped. Its June 2023 report on chatbots in consumer finance calls them doom loops: the bot repeats unhelpful replies and offers no way through to a human. By the bureau's count, over 98 million Americans, about 37% of the population, used a bank's chatbot in 2022.
Hand over to a person when any of these happen:
- The customer asks for a human, in any wording. "Agent", "real person" and "can I speak to someone" all count.
- The customer says the answer was wrong, or asks the same thing again in different words.
- Money has already left their account: refunds, double charges, disputes.
- The message mentions a death, an illness, a legal threat or anything about safety.
- It's their second contact about the same order this week.
Put the "talk to a person" button on screen from the first message. Hiding it until the bot gives up is how you make people angry.
3. Hand over the conversation, not just the customer
People mind telling the story twice. When the handoff fires, the agent should open the ticket and see the whole chat, the order the bot already looked up, and one line saying why it escalated.
Then tell the customer what happens next, with a real time in it: "Someone on our team will reply here, usually within two hours on weekdays." If nobody's working, say when they'll be back.
4. Let it answer only from your own documents
In February 2024 a Canadian tribunal ruled against Air Canada over its website chatbot. The bot told a grieving customer they could apply for a bereavement fare after travelling, within 90 days of the ticket being issued. The airline's own policy page said the opposite. Air Canada argued, in effect, that the chatbot was responsible for its own words. In Moffatt v. Air Canada, 2024 BCCRT 149, the tribunal called that a remarkable submission, found the airline "did not take reasonable care to ensure its chatbot was accurate", and ordered it to pay $812.02, of which $650.88 was damages.
Small money, clear lesson. Whatever your bot says, you said. So it should answer from your help articles, policies and order system, and when the answer isn't there, say so and hand over. That's the brief for AI customer support automation done properly: trained on your docs and policies, with the hard cases escalated to your team.
Two habits help. Keep policy answers close to the policy's own wording, and re-test the bot whenever a policy changes.
5. Tell people they're talking to a bot
Say it in the first message: "I'm an automated assistant. I can check orders and returns, or get you a person." A friendly name is fine. Bank of America's Erica had been used by nearly 32 million customers by October 2022, according to the same CFPB report. Letting people believe they're chatting with a human is where it goes wrong.
In the EU it's now a legal duty too. Article 50 of the AI Act says systems that talk directly to people must tell them they're dealing with AI, unless that's already obvious, and under the regulation as published that article applies from 2 August 2026.
6. Set the tone: short replies, no scripted sympathy
"I completely understand how frustrating this must be" sounds like mockery when the same bot then fails to help. Cut it. Write the instructions so replies run two or three sentences, lead with the answer and use the customer's own name for the product.
- Answer first. "Your order shipped Tuesday and should arrive Friday" beats a paragraph about tracking.
- Ask for one piece of information at a time.
- No exclamation marks on bad news.
- Reply in the language the customer wrote in.
Then read 50 real transcripts in the first week. Tone problems show up there first.
7. Measure confirmed resolutions, not deflection
Deflection counts tickets that never reached a person. It rises when the bot works, and it also rises when customers give up and phone instead, or leave. The number can't tell those apart.
Vendor billing often can't either. Intercom charges $0.99 per outcome for its Fin agent, and its help article on Fin pricing counts a resolution when no further help is requested after the last AI answer. Someone who closes the tab in disgust meets that definition. So the vendor's resolution rate isn't your customers' verdict.
Track these instead:
- Confirmed resolutions, where the customer clicked "yes, that solved it".
- Repeat contacts: the same customer back about the same issue within seven days, on any channel.
- How often the bot escalates, and how long people then wait.
- Satisfaction scores on bot conversations next to human ones.
What deflection rate is worth targeting?
Set it from your own ticket sort. For the illustrative shop in step 1, a sensible first target is a confirmed "yes" on about half of the 660 routine tickets. That's roughly 33% of all volume, with repeat contacts holding flat. Raise the target once a month of transcripts shows the bot gets things right.
Intercom's own pricing calculator assumes a 76% resolution rate, by its own definition. Read that as a ceiling, not a starting point. If your deflection figure climbs while repeat contacts climb with it, the bot is moving work around instead of doing it. Go back to the escalation rules in step 2.
Quick answers
Should an AI support bot issue refunds?
It can explain the refund policy and check where a refund is. Sending money back should sit behind a human approval. Refunds and disputes are also on the handoff list in step 2, because a wrong answer there costs real money and trust.
Do I have to tell customers they're talking to AI?
In the EU, yes: Article 50 of the AI Act requires it unless it's obvious, and it applies from 2 August 2026. Rules elsewhere vary. Saying so in the first message costs nothing and heads off the complaint either way.
How long does it take to set up?
The first automation is usually live in two to four weeks, after an audit of your workflows in week one.
