Voice UX: designing conversations people enjoy

0:00 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.

0:40 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.

1:09 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.

1:38 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.

2:08 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.

2:45 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.

3:17 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.

3:50 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.

4:30 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.

5:07 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.

5:43 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.

6:07 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.

7:38 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.

8:09 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.