wibbu

Which customer messages should AI answer? A simple guide

AI can answer most customer messages on its own. Here’s how to tell which ones, and how to hand the rest to your team with a draft prepared.

Branko Conjic

In short

Most customer messages are the same few questions, and AI can answer them without help. Decide each message with two checks, what’s at stake and how sure the AI is. Clear answers go out automatically. Judgment calls, health questions and upset customers go to a person, with a reply drafted by AI.

Most customer messages are the same handful of questions, like when you’re open, what something costs or whether there’s a slot on Friday. AI can answer those in seconds, at any hour.

At HB-Themes, my WordPress theme business, AI answers more than 90% of our support messages. That includes refund requests, because it can check the license key and whether the purchase is still inside the refund period.

The rest still needs a person. Think of a complaint, a refund outside the policy or a customer who’s upset. AI helps there too, by passing the conversation to your team with a reply already written.

So every message comes down to one decision. Can AI send this reply by itself, or should a person see it first?

Two kinds of messages

Most messages fall into one of two groups:

  1. AI sends it. The answer is written down, or AI can look it up in your records. Opening hours, prices, shipping times, order status and refunds that clearly meet your policy.
  2. A person takes it, with a draft ready. The message needs judgment or care. AI hands it over with a suggested reply. Someone checks it, edits it and sends it, or writes their own. Complaints, custom quotes, refunds outside your policy and anyone who’s upset.
AI sendsAutomatically
  • “What time do you open on Saturday?”
  • “Where’s my order?”
  • “I bought this yesterday. Can I get a refund?”
Goes to a personWith a suggested reply
  • “Can I get a refund? It’s been three months.”
  • “How much for 40 custom mugs?”
  • “This is the second time you’ve got it wrong.”
Two ways AI can handle customer messages.

Even in the second group, AI saves time. Editing a good draft takes seconds, while writing a reply from scratch takes minutes.

Two checks for every message

The topic gets you most of the way, but not all of it. To decide each message, AI should check two things:

  • What’s at stake? Health questions and upset customers go to a person, however sure the AI is. Money can stay with AI, but only when it can check the facts itself, like a valid purchase inside your refund period.
  • How sure is the AI? If the answer is clearly in your information, AI sends it. If it’s only fairly sure, it writes a draft for your team to check instead. If it can’t find an answer, it hands the conversation over.

Some tools let you choose how sure the AI must be before it replies without a person. Start strict, and loosen it as you see which answers it gets right.

Why the checks matter

Air Canada found out in court what a wrong AI answer can cost. Its chatbot told a customer he could claim a bereavement discount after his flight, which wasn’t the airline’s policy. In 2024, British Columbia’s Civil Resolution Tribunal made the airline pay anyway. It ruled that a business is responsible for everything on its website, chatbot included. So AI should only handle money when it can check your real policy and the customer’s records.

Customers also want a person within reach. In a 2026 Gartner survey, 87% said companies using AI for customer service must offer a way to talk to a real person. Sending the harder messages to your team covers both. Customers get quick answers to simple questions and a person when it matters.

When AI should hand over

Even a simple conversation can change halfway through. AI should pass it to your team when:

  • The customer asks for a person.
  • The customer gets upset.
  • The AI can’t find the answer.
  • Two replies haven’t solved it.

When it does, it should tell the customer what happens next, so they don’t feel dropped or have to repeat themselves. For example:

  • During opening hours: “Thanks! Someone from our team will take it from here. They can see this whole chat, so you won’t need to repeat anything. You’ll get a reply by [time].”
  • After hours: “We’re closed right now, but I’ve saved your message. Someone will reply by [time] tomorrow.”

If several people share the inbox, decide who picks up handovers. Our guide to managing Instagram DMs as a team has simple ways to split that.

Tell customers it’s AI

When AI replies automatically, it should say so in its first message. In the EU, that’s the law.

The EU AI Act is the EU’s main law on AI. Its Article 50 requires chatbots to tell people they’re talking to AI, unless it’s obvious. It has applied since August 2, 2026, and breaking it can cost up to €15 million or 3% of worldwide annual turnover. Small and medium-sized businesses face the lower of the two.

This applies outside the EU too. The rule covers any chatbot that talks to people in the EU, wherever its provider is based. If you built your own chatbot, that’s you. If you use a tool, check that it does this.

Our guide to EU AI Act chatbot disclosure has a first line for each channel and a short checklist.

How Wibbu decides

Wibbu is built around the same two checks, what’s at stake and how sure the AI is. For every message, a language model writes a reply from your information. Then a separate decision model looks at the message and the draft and answers a few questions. Is the answer in your information? Is money involved? Is the customer upset? Did they ask for a person?

Your rules come first. You might send every refund to a person, or let AI handle the ones it can verify, as we do at HB-Themes. Then your confidence setting decides whether Wibbu sends the reply, saves it as a draft or hands the conversation to your team.

  1. Message Customer

    I bought this yesterday. Can I get a refund?

  2. Draft Language model

    Of course! Your purchase is inside our 14-day refund period, so I’ve started your refund.

  3. Checks Decision model

    • Answer in your information97%
    • Money involved99%
    • Customer upset3%
    • Asked for a person1%

  4. Decision Your rules, then your confidence setting

    • RuleRefunds inside your policy can go out
    • Confidence97%, above your 85% setting
    • Send (chosen)
    • Draft
    • Hand over
How Wibbu handles one message, with example numbers.

Why two models? Language models tend to be overconfident about their own answers. Wibbu uses a decision model trained to be calibrated. Across many answers, when it says 80%, it’s right about 80% of the time. That’s what makes a confidence setting mean something.

Wibbu also learns from your team. If you keep sending the AI’s draft for the same question without changing it, Wibbu suggests letting AI answer that question by itself.

Common questions

How many of my messages can AI send on its own? It depends on how often the same questions come up. Count your 20 most common questions and see what share of your messages they cover. That’s roughly how much AI can answer directly. The rest goes to your team as drafts.

Can AI handle refunds? Yes, when it can check the facts, like a valid purchase inside your refund period. Anything outside your policy goes to your team, who can start from the AI’s draft. Complaints and upset customers should always get a person.

What if AI gives a customer the wrong answer? Treat it like a mistake by a staff member. Reply personally, put it right and honor any reasonable promise the AI made. Then fix the information it used.

Do I have to tell customers they’re talking to AI? Yes, if you have customers in the EU, wherever your business is based. The EU AI Act has required it since August 2, 2026, unless it’s obvious. The tool you use should do this for you, and our EU AI Act guide covers the details. Without EU customers, it’s still the right call.

Keep reading