AI7 min

Designing chatbot conversations people don't hate

Designing chatbot conversations people don't hate: openers, turn-taking, errors, handoffs and voice, with lessons from building Sage, my own AI assistant.

Saurabh YadavFounder, Design Studio by Saurabh
AIConversation designChatbot UX designAI assistant UX
Share

Most people have a chatbot horror story. The bank bot that answers every question with the same three buttons. The support bot that can't understand "I want to talk to a person". The one that says "I'm sorry, I didn't get that" four times in a row.

Better AI models have fixed some of this, but not most of it. The bad experiences usually aren't caused by the bot being dumb. They're caused by nobody designing the conversation: what the bot says first, how it handles not knowing, when it gives up and hands over, and how it sounds.

I learned a lot of this building Sage, the AI assistant on my own site. It chats, and it talks in real time in English and Hindi, with eight voices to switch between and support for interrupting it mid-sentence. Here's what I'd tell anyone designing a chatbot, whether it's for support, onboarding or a product feature.

Start with the job, not the personality

The first instinct on most teams is to name the bot, pick an avatar and write a cheerful tagline. I'd do that last.

Start by listing the five to ten things people will actually come to the bot for. For Sage, that list was simple: who is Saurabh, what has he worked on, what tools does he use, is he available for work, and how do I contact him. Everything else is secondary.

Then, for each job, write:

  • What a successful answer looks like
  • What the bot needs to know to give it
  • What it should do when it doesn't know

That last line matters most. Sage is told to answer only from a fixed set of facts about me, and to say it doesn't know and suggest emailing me when something isn't covered. That one rule prevents most of the embarrassing failures, like inventing a project I never did. There are other designers with my name online, so this mattered more than I first expected.

The first message does most of the work

People decide in the first exchange whether a bot is worth their time. A good opener does three things:

  1. Says what it is. Be open that it's an AI assistant. Pretending to be a human backfires the moment it makes a mistake.
  2. Says what it can help with. Two or three concrete examples, not "Ask me anything!"
  3. Makes the first step easy. Suggested prompts or chips the person can tap.

A weak opener: "Hi! How can I help you today?"

A better one: "Hi, I'm Sage, Saurabh's assistant. I can tell you about his projects, his process or how to work with him. Want to start with his Pukaar.ai case study?"

The second one teaches the user what to ask. That's the job of an opener.

Watch out

"Ask me anything!" sounds generous but puts all the work on the user. Most people freeze, type something vague, and get a vague reply back.

On mobile in India, think about the person typing on a budget Android phone, maybe switching between Hindi and English mid-sentence. Suggested replies save them typing, and accepting Hinglish input without complaint is basic respect.

Write for turn-taking, not paragraphs

A chatbot reply isn't a help article. It's one turn in a conversation. Rules I follow:

  • Answer first, then explain. Lead with the direct answer in one sentence. Details after.
  • Keep replies short. If it needs scrolling, split it, or offer "want the details?"
  • End with a useful next step. A follow-up question, a link or a button, so the person never hits a dead end.
  • Don't ask for things you already have. If the user said "I need a mobile app for my clinic", don't then ask "What kind of project is it?"
  • Use cards for structured things. When someone asks for my resume, Sage shows a resume card rather than describing it in a paragraph.

This is the same plain-language thinking I used on Pukaar.ai, where the parent sees a simple sentence before any medical label. People under stress or in a hurry read the first line and skip the rest. I go deeper into this in plain language before labels.

Design the failures on purpose

Every bot fails. The question is how gracefully. Here's the plan I use, in short:

Situation What the bot should do Example line
First misunderstanding Rephrase and offer two likely options "Did you mean his banking project or his baby care app?"
Second misunderstanding Show the main topics as buttons Tappable chips for each core job
Third misunderstanding Offer a human or another channel "You can email Saurabh directly."
Doesn't know the answer Say so, then point somewhere useful "I don't have that, but he reads every email."
Sensitive topic Hand over with context Complaints, payments, legal or medical

When it doesn't understand

Never repeat the same error message. Escalate instead:

  • First miss: rephrase and offer two likely options. "Did you mean his banking project or his baby care app?"
  • Second miss: show the main topics as buttons.
  • Third miss: offer a human or another channel. For Sage, that's my email.

When it doesn't know

Say so clearly, then point somewhere useful. "I don't have that information, but Saurabh reads every email" is a far better answer than a confident guess.

When it should hand over

Decide upfront which topics always go to a human. In support, that's usually complaints, payments and anything legal or medical. On Pukaar.ai, the AI companion QARA runs a red-flag check on every reply, and escalation to a paediatrician consult is always visible. That's the gold standard: the exit is designed in, not bolted on.

Handover should carry context. Nobody wants to explain their problem twice.

Note

Treat the handover as a real screen, not a fallback. Design what the human agent sees, so they can pick up mid-conversation instead of starting with "How can I help?"

Voice changes everything

Adding real-time voice to Sage taught me that a voice conversation is a different design problem, not a chat with audio on top.

  • There's no screen to scan. People can't skim a spoken answer, so replies need to be even shorter, with one idea per sentence.
  • Lists don't work aloud. "Here are five things" makes people stop listening after the second. Give the top one or two and offer more.
  • Interruptions are normal. People talk over each other all the time. If the user starts speaking, the assistant should stop and listen. Supporting interruptions made Sage feel far more natural than making people wait for it to finish.
  • Silence needs feedback. In chat, a typing indicator shows it's working. In voice, you need a sound or visual cue so people know they were heard.
  • Language switching should just work. Many Indian users mix Hindi and English naturally. The assistant should follow them, not force a choice.
  • Voice is personality. Offering eight voices let visitors pick one that felt comfortable to them, which matters more than I expected.

Tip

Before shipping a voice reply, read it out loud at normal speed. If you run out of breath or lose the thread, it's too long for voice.

I wrote the longer version of these lessons in what I learned designing a voice AI assistant.

Personality last, then test it

Once the jobs, openers and failures are solid, then give it a voice. Keep it consistent and appropriate to the context. A banking bot should be calm and precise. A playful brand can be lighter. A healthcare bot should never be jokey.

A quick test: read three replies aloud. If they sound like a person you'd trust in that situation, the tone is right. If they sound like an ad, cut the exclamation marks.

How to test a chatbot before launch

  • Write twenty real questions from support tickets, sales calls or interviews, not ones the team invents.
  • Include the awkward ones: rude messages, off-topic questions, typos, Hinglish, one-word replies.
  • Read transcripts, not dashboards. The insights live in real conversations.
  • Watch three people use it. You'll see where they hesitate, rephrase or give up.

My guide to designing AI features people trust has more on showing reasoning so people know when to rely on the answer.

Work with us

Building an AI feature? Our AI product design agency designs copilots, agents and AI workflows that users trust.

Key takeaways

  • Design the jobs and the failures first. Name and personality come last.
  • Openers should say what the bot is, what it can do, and offer an easy first step.
  • Answer first, keep turns short, and always end with a next step.
  • Never repeat the same error. Escalate, and design the human handover in from day one.
  • Voice needs shorter replies, real interruption handling and natural language switching.

Try it yourself

The fastest way to understand these ideas is to talk to a bot that tries to follow them. Open Sage, ask it something it can't answer, interrupt it in voice mode, or switch to Hindi halfway through. Then look at how the Pukaar.ai case study handled escalation for anxious parents. If you're planning a chatbot or AI assistant for your product, tell me about your project.

Saurabh Yadav's Memoji

Written by

Saurabh Yadav

Founder of Design Studio by Saurabh, a UI/UX design studio in India. We design apps and websites that feel effortless, from research to shipped pixels. Most recently: Pukaar.ai, as seen on Shark Tank India, and the Demerara Bank app.

Keep reading

All articles →