---
title: "Voice UX: designing conversations people enjoy"
description: "Voice is linear and invisible On a screen, people scan, skip and go back. In voice, they can only hear one thing at a time, in order, and cannot see…"
url: https://optimizeall.com/learn/voice-ai-agents/voice-ux-design
updated: 2026-10-05
---

Voice AI & Conversational Agents · Designing voice conversations · lesson 7 of 17 · 15 min

# Voice UX: designing conversations people enjoy

## Voice is linear and invisible

On a screen, people scan, skip and go back. In voice, they can only hear one thing at a time, in order, and cannot see options. Memory is short. That changes everything about design:

- **Shorter turns**: one or two sentences, then hand the turn back.
- **Fewer options per turn**: three at most; better, ask an open question and let the agent interpret.
- **Explicit confirmation of critical data**: dates, times, amounts, names, numbers.
- **Clear signposting**: "Two quick questions, then I'll book it."

## The anatomy of a good voice conversation

1. **Opening**: identity + AI disclosure + purpose + invitation. "Hi, I'm Aria, Nova Dental's AI assistant. I can book, move or cancel appointments. What can I do for you?"
2. **Intent capture**: open question, then clarify only what is missing.
3. **Slot filling**: gather required details one at a time, in a natural order.
4. **Action**: use tools; narrate briefly while waiting.
5. **Confirmation**: read back the key details and get a clear yes.
6. **Closing**: summary, next step, offer of anything else, polite goodbye; end the call cleanly.

## Confirmation strategies

| Strategy | Example | Use for |
|---|---|---|
| Explicit | "That's Thursday the 14th at 2:30 pm at Dubai Marina. Shall I book it?" | Irreversible actions, money, appointments |
| Implicit | "Thursday afternoon, got it. I have 2:30 or 4." | Low-risk details, to keep flow |
| Chunked readback | "Your number is 0300... 555... 1234. Is that right?" | Phone numbers, IDs, card-like numbers |
| Spelling | "Could you spell the surname?" + NATO-style confirmation for letters when needed | Names, emails |

## Repair: when things go wrong

- **No match**: rephrase, do not repeat verbatim. "Sorry, I didn't catch that. Are you looking to book, move or cancel?"
- **Escalating help**: after two failures, offer alternatives: keypad input, SMS link, or a human.
- **Out of scope**: say what you can do and offer the right channel. Do not guess.
- **Correction handling**: "No, Tuesday" should update the slot immediately, with no argument.

## Personality and tone

Define a persona in a few lines: name (optional), role, tone (warm, efficient), formality level per market, what it never does. For Gulf and South Asian markets, courtesy formulas matter (greetings, "inshallah" when a customer uses it, respectful address). Do not overdo humor or fake emotion; callers want competence.

## Accessibility and inclusion

- Speak at a moderate pace; offer to repeat or slow down.
- Support keypad input for people who find speaking difficult.
- Avoid time pressure; configure generous silence timeouts for older callers.
- Offer a human route clearly, especially for complaints, vulnerable situations and emergencies.

## Worked example: rewriting a clunky flow

**Before** (screen thinking): "Please tell me your full name, date of birth, phone number, the service you require from cleaning, whitening, check-up, braces consultation or emergency, your preferred branch, and preferred date."

**After** (voice thinking):

> Agent: "Sure, let's get you booked. What's it for?"
> Caller: "My teeth need a clean."
> Agent: "A cleaning, great. Which branch suits you, Dubai Marina or Downtown?"
> Caller: "Marina."
> Agent: "And roughly when? This week or next?"
> Caller: "Thursday afternoon."
> Agent: "Let me check Thursday afternoon at Marina... I have 2:30 or 4. Which is better?"
> Caller: "2:30."
> Agent: "Lovely. Can I take your name and mobile number for the booking?"
> ...
> Agent: "So that's a cleaning, Thursday the 14th at 2:30, Dubai Marina, for Omar Khalid. Shall I confirm?"

## Hands-on: a conversation design spec template

```markdown
Use case: Book a cleaning appointment
Entry: inbound call, caller ID known/unknown
Required slots: service, branch, date/time, name, mobile (if unknown)
Optional: new patient?, insurance provider
Opening line: "Hi, I'm Aria, Nova Dental's AI assistant..."
Happy path: [sample dialogue, 8-12 turns]
Confirmations: explicit before booking; chunked readback of mobile number
Repair: 2 no-matches -> offer keypad for numbers / human transfer
Out of scope: clinical questions -> "I can't give medical advice; for urgent pain call ..."
Closing: summary + SMS confirmation + "Anything else?" + goodbye + end call
Metrics: booking completion rate, avg turns to booking, transfer rate, CSAT/thumbs
```

Write three sample dialogues (happy path, a correction, a failure) before writing any prompt. They become test cases later.

## Pitfalls

- Reading out long lists and menus.
- Confirming everything explicitly, which makes calls tedious; reserve explicit confirmation for what matters.
- Designing only the happy path.

## Video lecture: Voice UX: designing conversations people enjoy

Lecture coming soon · 14 chapters · about 9 minutes. Read the full transcript below.

1. Voice UX
2. Why voice UX
3. Four rules
4. Analogy: phone directions to a driver
5. Six-part anatomy
6. Confirm wisely
7. Repair
8. Persona and inclusion
9. Worked example: rewrite
10. Example 2: Riyadh card activation
11. Common mistakes
12. Watch me do it: rewrite the design spec
13. Recap and next step
14. Try this now

## Lecture transcript

### Voice UX

Here's the fundamental truth of voice design: voice is linear and invisible. On a screen, people scan, skip ahead and go back. On a call, they hear one thing at a time, in order, and they can't see any options. Their short-term memory does all the work. If you design a voice agent like a web form, callers get lost. In this lesson, you'll learn the anatomy of a good voice conversation, how to confirm things without being tedious, how to repair misunderstandings, and how to make agents accessible and inclusive.

### Why voice UX

Why does voice UX matter so much? Because in voice, the design is invisible, so every flaw lands directly on the caller's memory and patience. A confusing menu on a website is annoying. A confusing sequence on a phone call feels endless. Good voice design shortens calls, reduces transfers and repeat calls, and makes the agent feel competent, which is exactly what customers judge you on.

### Four rules

Four design rules follow. Keep turns short: one or two sentences, then give the turn back. Offer few options per turn, three at most, or better, ask an open question and let the agent interpret the answer. Confirm critical data explicitly: dates, times, amounts, names and numbers. And signpost: two quick questions, then I'll book it. People relax when they know how long something will take.

### Analogy: phone directions to a driver

An analogy: designing for voice is like giving directions over the phone to someone driving. You don't read them a map. You give one instruction at a time, confirm the tricky turn, tell them roughly how far it is, and if they miss a turn, you calmly reroute instead of repeating the same sentence louder. That's short turns, confirmations, signposting and repair, exactly what good voice agents do.

### Six-part anatomy

A good conversation has six parts. The opening: identity, AI disclosure, purpose and an invitation, like: hi, I'm Aria, Nova Dental's AI assistant, I can book, move or cancel appointments, what can I do for you? Then intent capture with an open question. Then slot filling, gathering what's needed one detail at a time. Then action, using tools and briefly narrating the wait. Then confirmation, reading back key details. And finally a closing: summary, next step, anything else, goodbye, and ending the call cleanly.

### Confirm wisely

Confirmation is a balance. Use explicit confirmation for irreversible actions, money and bookings: that's Thursday the fourteenth at two thirty at Dubai Marina, shall I book it? Use implicit confirmation for low-risk details, to keep flow: Thursday afternoon, got it. Use chunked readback for numbers, reading a mobile number in groups. And ask people to spell names when needed. If you confirm everything explicitly, the call becomes tedious. Save it for what matters.

### Repair

Things will go wrong, so design repair. If the agent doesn't understand, it should rephrase, not repeat word for word: sorry, I didn't catch that, are you looking to book, move or cancel? After two failures, offer alternatives: the keypad, a text message link, or a human. For out-of-scope requests, say what you can do and point to the right channel instead of guessing. And when someone corrects you, no, Tuesday, update immediately and never argue.

### Persona and inclusion

Define a persona in a few lines: a name if you want one, a role, a tone like warm and efficient, a formality level for each market, and what it never does. In Gulf and South Asian markets, courtesy matters: proper greetings and respectful address. Don't overdo humor or fake emotion. Callers want competence. And design for inclusion: a moderate pace, offers to repeat or slow down, keypad input for people who find speaking hard, generous silence timeouts for older callers, and a clear human route for complaints, vulnerable situations and emergencies.

### Worked example: rewrite

Let's rewrite a clunky flow. The screen-thinking version asks for full name, date of birth, phone number, one of five services, branch and preferred date, all in one breath. The voice-thinking version goes: sure, let's get you booked, what's it for? A cleaning, great. Which branch, Dubai Marina or Downtown? And roughly when, this week or next? Then it checks availability, offers two times, takes a name and mobile, and reads back the booking before confirming. Same data, a fraction of the effort for the caller.

### Example 2: Riyadh card activation

A second example. A bank's card-activation line in Riyadh must confirm identity, activate a card and set a PIN reminder, all in Arabic or English. The designers use chunked readback for the last four digits, explicit confirmation before activation, and keypad entry for the date of birth, because speech recognition on dates can be error-prone. After two failed identity attempts, the agent offers a human without judgment. And every turn is one short question. Calls are shorter, and fewer people need a human.

### Common mistakes

Common voice UX mistakes. Reading out long lists and menus, which callers can't hold in memory. Confirming everything explicitly, which makes calls slow and tedious. Designing only the happy path, when real callers change their minds, mishear and ask off-topic questions. And a persona that jokes or fakes emotion, when callers mostly want competence and speed.

### Watch me do it: rewrite the design spec

Watch me do it. I open the conversation design spec for booking a cleaning and rewrite it live. Opening line: hi, I'm Aria, Nova Dental's AI assistant; I can book, move or cancel appointments; what can I do for you? Required details: service, branch, date and time, name and mobile. I reorder them the way people naturally talk: what it's for, which branch, roughly when, then name and number. Confirmation plan: implicit for service and branch, like a cleaning, great; explicit before booking, reading back service, date, time, branch and name; and chunked readback for the mobile number. Repair: after the first miss, rephrase with the three options; after the second, offer the keypad for numbers or a transfer. Out of scope: any clinical question gets the approved wording and the earliest appointment offer. Closing: summary, a text confirmation, anything else, goodbye, end call. Now I write the correction dialogue: the caller says Thursday, then, no sorry, Tuesday. The agent simply says Tuesday, got it, and continues. Finally, I read all three sample dialogues aloud with a colleague and time them. The happy path takes about ninety seconds out loud. Anything that felt long I cut, including a sentence where Aria explained the booking system. Callers don't need that.

### Recap and next step

Recap. Design for the ear: short turns, few options, smart confirmations, signposting, graceful repair, a clear persona and inclusive defaults. Your next step: copy the conversation design spec from the lesson text and write three sample dialogues for your use case: a happy path, one with a correction, and one where things fail and the caller reaches a human. Do this before writing any prompt. Those dialogues become your test cases later.

### Try this now

Try this now. Take your use case and write the opening line with identity, AI disclosure, purpose and an invitation. Then list the required details and order them the way a caller would naturally give them. Decide which ones need explicit confirmation and which can be implicit. Write the repair path for two failures. Finally, read your three sample dialogues out loud with a colleague playing the caller, and time them. Anything that feels long out loud is too long.

## Key takeaways

- Voice is linear and invisible: keep turns short, limit options, signpost, and confirm critical data.
- A good conversation has an opening with AI disclosure, intent capture, slot filling, action, confirmation and closing.
- Use explicit confirmation for irreversible actions, implicit for low-risk details, chunked readback for numbers.
- Design repair paths (rephrase, keypad, SMS, human) and inclusive defaults before writing prompts.

## Try it

Fill in the conversation design spec for your use case and write three sample dialogues: happy path, correction, and failure leading to a human.

- [Previous: Telephony: phone numbers, SIP trunks and Twilio](https://optimizeall.com/learn/voice-ai-agents/telephony-sip-and-twilio)
- [Next: Prompt design for voice agents](https://optimizeall.com/learn/voice-ai-agents/prompting-for-voice-agents)
- [All lessons of Voice AI & Conversational Agents](https://optimizeall.com/learn/voice-ai-agents)