Most of our design experience came from screens, and most of what we learned there turned out to be misleading the day we started designing for a phone call.
A screen forgives you. You can show five options at once, let the eye skip what it doesn't need, lay out a hierarchy, and trust people to find their way. None of that exists on a call. There's no layout. The only interface is time: what gets said, in what order, and how long each part takes.
That's the central problem with AI voice UX today. Most voice products, including early versions of our own, still get designed like a screen with the pictures removed. They read out menus, list every option, and ask the caller to remember things a screen would have kept in view. Callers feel it, even if they can't name it, and they hang up or ask for a human.
If you run a business considering an AI phone agent, this matters to you even if you never design anything. Much of the difference between a voice agent that feels like a helpful person and one that feels like a polite machine comes from design decisions you can hear on a single call.
If you're evaluating a voice agent rather than building one, skip to the mistakes and best practices near the end; they double as a checklist for a test call.
What is voice UX design for phone calls?
Voice UX design is deciding what a voice system says, how much it says, in what order, and in what tone, so that a caller can get what they need without effort. For a phone agent, it covers greetings, questions, confirmations, error recovery, and handoffs to a person.
It's different from writing a script. A script describes the conversation you expect. Voice design is mostly about the conversations you don't expect: the caller who interrupts, changes their mind, or answers a question you haven't asked yet.
Designing without a canvas
The thing nobody warns you about, moving from visual design to voice, is that you lose the artefact. A screen design is something you can put in front of someone: a file, a frame, a link. A conversation isn't a thing until someone actually has it.
You can write a script, but a script is not the experience any more than a shot list is a film. You have to hear it, in something close to real conditions, before you know if it's any good.
Recorded calls are the real deliverable
That changed what we think a design deliverable is. Increasingly it's a set of real recorded calls, annotated, that the team argues about the way designers argue about a mockup. We've written about what that looks like week to week. It's less elegant than a design file and more useful for this job.
How a well-designed phone conversation flows
Step 1: Find out what the caller wants
Open by inviting the caller to say what they need, in their own words. Don't start with a menu. Most callers know exactly why they called, and a menu makes them translate their need into your categories. This is the core argument in our comparison of AI receptionists and IVR phone trees.
Step 2: Narrow one branch at a time
Instead of listing everything, ask the one question that cuts the options most. "Earlier or later in the week?" does more work than reciting eight slots.
Step 3: Confirm before acting
Before booking or changing anything, say back the key details in one short sentence: day, time, service. On a call, this replaces the summary screen a website would show.
Step 4: Keep an exit to a person
Every flow needs a clear, quick way out. When a caller asks for a human, they should get one, or a message taken with context, without having to argue for it.
Information architecture for the ear
Why spoken lists fail
Screen design lets you show several things at once and let the eye sort them. Say three appointment options out loud and, by the third, many people have lost the first, even when they're paying attention, because holding a spoken list in working memory is harder than scanning a written one. It's a lesson usability research into voice interaction keeps returning to, and one screen-first teams tend to relearn the hard way.
Screen-shaped vs. voice-shaped
A screen-shaped call says: "We have Tuesday at 2, Wednesday at 11, Thursday at 3, or Friday at 10. Which works for you?" It expects the caller to hold four options and pick.
A voice-shaped call says: "What usually works better for you, earlier or later in the week?" and narrows from there. Both reach the same booking. Only one feels like talking to someone who's listening.
Does offering fewer options limit the customer?
It's a fair worry. The answer is to make the narrow path the default without making it the only path. Aiva Chat leads with the option most likely to match what was asked, and expands the moment a caller says "what else do you have?" Most callers want the fastest correct path to done. The ones who don't will ask, and should get a real answer.
We test for this specifically, deliberately sounding like a caller being boxed in, and fixing it when a flow drifts that way.
Tone is a spec you can't point at
The hardest part isn't deciding what the agent should say. It's specifying how. A designer can tell an engineer a corner radius is 12 pixels and everyone agrees instantly. Nobody can specify "warm but efficient, reassuring without being saccharine" with that precision, yet that's the brief for almost every moment in a call.
Examples over adjectives
What works, imperfectly, is triangulating with examples: this recorded exchange, not that one. The spec becomes a direction you keep correcting toward. It's why good references, like Google's conversation design guidelines, lean so heavily on personas and sample dialogue rather than adjectives.
You can specify a corner radius to the pixel. You cannot specify "reassuring" to anyone's satisfaction. Most of the actual job lives in that gap.
The same words, a different language, a different design
The gap widens across languages. Aiva Chat speaks 13, and a pause that reads as attentive in English can read as hesitant in a language where responses are expected faster. Directness that sounds efficient in one language can sound curt in another, where a softer opening phrase does real social work.
Native speakers finish the design
No single designer speaks all of these fluently, so part of the work belongs to native-speaking teammates who can say when something is technically correct and quietly wrong. Treating voice design as one deliverable that gets localised afterwards is, we think, the most common mistake in this category, and we made it in our own earliest version.
Where design ends and engineering begins
How fast the agent responds, how it handles being talked over, whether a pause reads as thoughtful or broken: that's latency and turn-taking, engineering's craft. It shapes the experience as much as any wording.
Interruptions are a shared problem
Interruptions are their own design problem, because real callers don't wait politely. They backtrack, correct themselves, and talk over a response once they've heard enough. Design decides what's worth saying; engineering decides how the system yields.
We've got this wrong both ways: wording so careful it arrived a beat too slow, and pacing so fast that untrimmed sentences sounded clipped. Neither team could fix it alone. The fix both times was listening to the same call in the same review.
Screen UX vs. voice UX at a glance
| Screen UX | Voice UX on a phone call | |
|---|---|---|
| Showing options | Many at once, user scans | One or two at a time, narrowed by questions |
| Memory load | Low, everything stays visible | High, so say less and confirm key details |
| Navigation | Back buttons, menus, links | Conversation; the caller must be able to redirect |
| Tone | Mostly visual style and microcopy | Wording, pacing, pauses, and politeness |
| Errors | Inline messages, highlighted fields | Spoken clarification without blame |
| Localisation | Translate strings, adjust layout | Redesign pacing and politeness per language |
| Testing | Click-throughs, screenshots | Listening to real calls |
Screens still win for some tasks. Comparing a dozen options, reading a long policy, or reviewing a receipt are easier with something to look at. That's why Aiva Chat also runs a website chat widget and why booking confirmations go out by text after a call.
For the conversation itself, though, voice needs its own rules. The biggest shift is memory: anything the caller must hold in mind is a cost, and good voice design spends that budget carefully.
Common AI voice UX mistakes
Reading the menu aloud
Converting a website's options into a spoken list is the most common mistake. It forces callers to remember choices and pick by position. Ask what they want instead, and offer options only when narrowing is needed.
Over-explaining
Long sentences that would be fine on a page feel endless on a call. If a sentence can be cut in half without losing meaning, cut it. Callers can always ask for more detail.
Translating instead of redesigning
A directly translated English script often sounds wrong in another language: too blunt, too formal, or oddly paced. Each language needs its own review by someone who speaks it natively.
Judging calls from transcripts alone
A transcript can look perfect while the call felt slow, clipped, or robotic. Pacing and tone are audible, not readable. Listen to real calls before deciding whether a change worked.
Hiding the human
Making it hard to reach a person might keep the AI's resolution numbers high, but it destroys trust. Design the handoff as carefully as the booking: quick to reach, with the context passed along.
Voice UX best practices for AI phone agents
- Open with a question, not a menu. Let callers state their need in their own words.
- Offer one or two options at a time. Narrow with questions instead of listing.
- Confirm before committing. Repeat day, time, and service in one sentence.
- Keep sentences short. Say less, and let the caller ask for more.
- Allow interruptions. The caller should be able to redirect at any moment.
- Design per language. Review pacing and politeness with native speakers.
- Listen, don't just read. Evaluate changes by calling the line yourself.
- Make the human easy to reach. Treat handoff as a feature.
What we actually optimise for
A voice that sounds impressively human is a parlour trick on its own. What we optimise for is a conversation that doesn't ask more of a caller's attention than it has to: nothing to hold in memory that didn't need holding, nothing said that could have been shorter, no moment where the caller has to work out what's being asked of them.
Almost none of that shows up in a feature list. Almost all of it shows up in whether someone hangs up feeling understood.
Conclusion
The lesson we'd pass to any team moving from screens to calls: budget for listening. Most of what we got wrong was invisible in transcripts and obvious within ten seconds of audio, so the review process matters more than the first script.
For a business owner, the useful takeaway is narrower. When you configure an agent's greeting and policies, write them to be heard. Short sentences, one question at a time, and no lists longer than two will do more for your callers than any voice setting.
And if you serve callers in more than one language, have a native speaker listen to a few calls in each before you call the setup finished.
Ready to hear the difference on a real call?
Ring +1 (970) 972-0974 and listen for the things this post is about: how many options it offers, whether it lets you cut in, how it confirms a booking back to you.
To put the same agent on your own number, the voice platform page explains how calls are handled, and every plan from $49 a month includes it.
Start your 14-day free trial with no card, or book a demo and we'll play back a few real calls and talk through the design choices in them.
Posts from the people building Aiva Chat in Rajkot — how the product works, what we have learned building it, and guides for getting the most out of an AI receptionist.