AI · Voice · Customer service · 2026

Voice AI that knows when to hand over.

Sayra answers every call in a brand’s own voice, resolves what it can, and passes the rest to a person before frustration starts, with the whole story attached. I designed the brand, the conversation, the product and the website.

Role: Brand, conversation design, product design
Scope: Brand · Conversation · Console · Agent workspace · Website
Year: 2026

Try the live demo Design files

Northwind Air and its callers are fictional.

The problem

The worst day of the year is a phone problem.

When a storm cancels flights, an airline’s phone lines collapse in minutes. Menus make callers do the routing. Bots misunderstand and loop. After every transfer, people tell their story again. And the leaders who’d buy an AI agent fear one thing above all: an agent that’s rude to a customer and ends up in the news.

“Press 1 for…”
Menus make callers do the routing work.
“Can you repeat that?”
After a transfer, people tell their story twice.
“What did the bot say?”
Leaders can’t see or trust AI answers.

The idea

Hand over before it hurts.

Most voice AI competes on how much it can do alone. Sayra competes on how well it works with people. Seconds before a call goes wrong, Sayra notices, offers a person, and that person picks up mid-sentence. Everything else in the product exists to make that moment safe.

Priya takes over with a five-second summary, and picks up mid-sentence.

Who it’s for

Four people inside the company, and one on the phone.

Maya
VP of Customer Experience · the buyer
Proof it works, and no brand incidents.
Daniel
Conversation designer
Change the agent without waiting for engineers.
Priya
Senior agent
The whole story the second a call lands.
Sam
Quality lead
Find what went wrong, and fix it.

“My flight’s cancelled, I’m at the gate with two kids, and I just need to get home tonight.”

Jordan, the caller

Conversation design

The interface you can’t see.

Most of Sayra’s users never look at a screen; they hear it. I designed the call itself following the Conversation Design Institute method: a system persona, sample dialogues written before any UI, flowcharts, repair strategies, and a Wizard-of-Oz test plan for the pilot.

The method as six steps of the Conversation Design Institute workflow: human-centricity (desk research; interviews planned with the pilot), persona, sample dialogues, flowchart design (two flows) and long-tail design are done; the Wizard-of-Oz test is planned. — The CDI workflow, step by step, with what’s done and what’s planned.

A sample call between Sayra and Jordan, whose flight was cancelled, with six annotations: Sayra says it’s AI once; acknowledges, then acts; reads the booking code back; fills the silence while it checks, then offers two flights; hands over when Jordan mentions a medical appointment; and Priya continues with what she already knows. — A real call, annotated: disclosure, read-backs, two options, and a handover the caller doesn’t have to repeat.

Confirmation strategy in four cards, by confidence and risk: high confidence, low risk confirms implicitly in the next turn; low confidence, low risk asks a quick yes or no; high confidence, high risk reads everything back before acting; low confidence, high risk asks again differently, then hands over. Each card has a sample line. — How strongly Sayra confirms depends on how sure it is and how costly a mistake would be.

Declarative or generative: fixed, approved words for the AI disclosure, read-backs, payment and security handoffs, legal statements and the handover sentence; generated speech within guardrails for acknowledging, explaining options, answering policy questions with sources and summaries; and the checks generated speech must pass before it is spoken. — The higher the risk, the more scripted the words.

Structure & flows

Structure before style.

Information architecture: six destinations in the console — Overview, Agents, Simulations, Live, Review and Settings, each with its pages — and a separate Agent workspace for the people who take over calls: incoming handovers, the call with full context, and the after-call note. — Six destinations in the console, and one focused workspace for people who take over calls.

Key flow

Flow: Jordan calls about a cancelled flight. Sayra greets him, says it’s AI and understands the reason, then checks whether he’s upset, confused or has asked twice. If not, it verifies his identity, offers two new flights, rebooks and confirms by voice and SMS. If yes, it offers a person and writes the handover; Priya’s workspace receives the call with context, and Priya continues, so nothing is repeated. — The core loop: Sayra resolves what it can and hands over early when it shouldn’t continue.

Wireframes

Live monitor: the calls that need a person first, each with its reason in words.
Handoff workspace: the packet first, in the same four lines on every call.
Agent builder: words in, rules out, tested before they’re saved.

1 / 3

Key design decisions

Say why, in words.

A flag that only says “risk 82” gets ignored.

Every flagged call says why in the caller’s own words, so a supervisor decides in seconds whether to step in: listen, whisper a hint, or take over.

The live monitor: the flag, the reasons in words, and three ways to step in.

A builder you talk to.

Daniel describes a change the way he’d brief a colleague.

Sayra shows exactly how it understood him and sorts every behaviour into three groups: does on its own, asks a person first, never does. Every change is tested on difficult callers before it ships.

Plain words in; a new rule appears under “Asks a person first”.

The one test caller that failed turns green with the fix.

Every answer can be explained.

Each thing Sayra says links to the source it came from.

A wrong answer leads straight to the outdated policy behind it, and the fix goes through the same tests as any change.

Review: the wrong sentence, its outdated source, and a fix tested on 24 calls.

When it isn’t simple

Designing the unhappy paths.

Nobody free to take the call, a handover missed, a caller who hangs up, a booking system that goes down. These states decide whether a CX leader trusts the product.

Handover edge cases: nobody is free, so Sayra offers a callback or to keep helping while the caller waits; the handover isn’t accepted, so it is re-routed to the next person with the same packet; the caller hangs up mid-handover, so it becomes a callback task and an SMS goes out; emergency words skip the queue as a priority handover. — Four of the handover’s unhappy paths, each with what the caller hears and what the team sees.

Brand

Calm at any volume.

The buyer’s biggest fear is losing control, so the brand promises calm, not power. The mark is an S drawn as two voices that never touch; the gap in the middle is the handover. Deep lagoon teal reads as a steady hand, mint is the breath of a live voice, and an editorial serif gives Sayra the warmth of a person.

The mark alone, and with the wordmark on lagoon and on mint.

Two equal circles, 225° each, and a fixed gap in the middle: the handover.

Colour proportions: 60% soft white #F8F9F7 for backgrounds, 30% lagoon #0E3B3A for text, primary buttons and structure, and 10% mint #7ED6BE for the live voice and accents on dark backgrounds. — Soft white, lagoon and mint, used 60 · 30 · 10.

LinkedIn banner, social posts, business card and pitch deck.

Website

A site that sells the moment, not the model.

Website section “Hand over before it hurts.”: what Sayra notices and three points, next to the summary card the agent receives, ending with what to say first. — The handover, shown as the summary the agent receives.

The whole website on a phone, top to bottom: the hero with the live call, the handover, how it works in three steps, the limits you set, the call to action and the footer. — The full page on a phone.

Outcome

What a pilot would measure.

These are the targets I designed for, to validate with a pilot. They are goals, not results.

Goal
6 in 10 calls resolved without a person
Goal
Caller satisfaction on par with human agents
Goal
0 stories repeated after a handover
Goal
Under 30 seconds to understand any call in review

Deliverables

Live prototype (opens in a new tab)
Product design file (Figma) (opens Figma in a new tab)
Brand & website file (Figma) (opens Figma in a new tab)
Brand identity
Conversation design (persona, sample dialogues, repair, prosody)
Console and agent workspace UI
Marketing website (desktop and mobile)