How to design an AI phone or chat agent customers actually like

Most AI agents fail on conversation design, not technology. Here's the framework we use to design agents that book, answer and hand off well.

By ··3 min read

Key takeaways

  • Design around the five or six reasons people actually call, not around everything your business does.
  • Say it's an AI, get to the point fast, and ask one question at a time.
  • Give the agent tools and approved facts instead of letting it guess.
  • Define the hand-off: when the agent stops, who it passes to, and what context goes with the call.

An AI agent that sounds clever but can't book an appointment is worse than voicemail. One that's a little plain but books accurately, answers the common questions and hands off gracefully will win customers over. Here's the framework we use.

1. Start from the calls you actually get

Before writing a word, list why people contact you. Listen to recorded calls or read through recent messages and group them. For most appointment-based businesses, five or six reasons cover the large majority:

  • book an appointment,
  • change or cancel one,
  • ask about hours, location or prices,
  • check on the status of a job,
  • a question that needs a person,
  • a complaint.

Design for those first. Everything else gets a polite hand-off.

2. The greeting

Keep it short, say who you are, say it's an AI, and invite the caller to say what they need:

"Thanks for calling Harbour Lane Motors service. I'm the AI assistant, how can I help you today?"

Avoid menus ("press one for..."), long disclaimers and marketing messages.

3. One question at a time

People on the phone can't scroll back. Ask for one thing, confirm it naturally, then ask the next:

"Sure, what's the vehicle?" → "And what day works best?" → "I have Thursday at 9 or 2:30. Which suits you?"

Don't ask for anything the caller already said. Good agents listen for details given early ("I need an oil change on my RAV4 Thursday") and skip those questions.

4. Give it tools and facts, not freedom

The most important rule: the agent should never guess about prices, availability or policies. Give it:

  • Tools to check the schedule, book, look up a job, and send a text.
  • An approved knowledge base for hours, prices, policies and directions.
  • Clear limits on what it won't discuss.

When the answer isn't in its tools or knowledge base, it says so and offers a callback.

5. Confirm and close the loop

After booking, confirm the key details in one sentence, send a text confirmation, and tell the team (a note in the right person's queue). Then ask if there's anything else.

6. Design the hand-off

Decide in advance when the agent stops: the caller asks for a person, is upset, has a complaint or warranty dispute, or asks something sensitive. A good hand-off passes the call with a summary so the customer doesn't repeat themselves. After hours, it takes a detailed message and promises a callback time you'll actually meet.

7. Review real conversations

For the first few weeks, read every summary. Look for questions it couldn't answer, places callers got confused, and bookings that needed fixing. Update the knowledge base and rules weekly; that's where most of the improvement comes from.

Hear the framework in action

Our live demo is built exactly this way: a short greeting, one question at a time, real tools for schedule, booking, texts and advisor notes, and a hand-off when needed. Call it from your browser and try to trip it up.

Frequently asked questions

Should an AI agent pretend to be human?

No. Say it's an AI. Customers are fine with that when it's fast and helpful, and pretending damages trust when people find out.

How long should replies be on a phone call?

Usually one or two short sentences. Long answers are hard to follow by ear; offer to text details instead.

What's the most common mistake?

Letting the agent answer from general knowledge instead of your approved information. It sounds confident and is sometimes wrong. Tools and a knowledge base fix that.