pranaykotadia.
Voice AI for BCG X, 2025

A voice assistant that answers with the owner’s own numbers

For small-business owners. They ask out loud, and it answers with their sales, reviews and bills, then gets the next step ready.

Scope
4flows, designed end to end
Designed in Figma
42screens
English, Arabic and Hindi
3languages
RoleProduct designer. I designed all four flows: growth, customer satisfaction, legal, and network and billing.
TeamThe only product designer on these flows
Timeline12 weeks in 2025
ToolsScreens designed in Figma. Interactive prototype built with Claude.
/ The problem

Owners run the business from their phone, between customers

The brief covered four small-business customers: a restaurant, a barber shop, a law firm and a factory. Their numbers already existed. Getting to them, and knowing what to do next, took time they didn’t have.

Owner question 01“How’s my business doing?”The answer is split across the POS, the bank, reviews and bills.
Owner question 02“What should I do about it?”“Run a promotion” doesn’t say which dish, for whom, or what it costs.
Owner question 03“Will it spend money without me?”Owners want help, but they don’t want it paying or messaging anyone on its own.
The briefA voice assistant that answers with the owner’s own numbers, sets up the next step, and asks before it pays for or sends anything.
/ Why voice

Why we went with voice

Owners don’t open dashboards between customers, but they would ask a question out loud.

What each option gives an owner mid-shift
/ Research findings

Four things about voice that changed the screens

Each comes from published research on voice interfaces, checked with the client’s small-business team, and led to one decision in the UI.

Finding 01People hold about three things they hear
→Say one short line. Put the numbers on a card.Reading 238 wpm, speech 140–180 (Brysbaert 2019, Cowan 2001)
Finding 02Voice AI takes about a second to answer
→Each loading state says what the assistant is doing.People 0.2 s, voice agents 0.9–1.4 s (J. of Cognition 2023, Twilio)
Finding 03Voice fails in noisy places
→Every spoken word also appears as text.Tasks done: 85% typed, 26–38% spoken (τ-Voice, 2026)
Finding 04Owners switch languages mid-sentence
→Show what it heard. Write campaigns in English, Arabic and Hindi.25–54% word errors on mixed Arabic and English (COLING, 2025)
/ Design principles

What the product is built on

Idea 01A chat instead of a dashboard
Idea 02Every choice can be spoken or tapped
Idea 03It brings things up when the owner isn’t busy
Idea 04Specialists join when the work needs them
Design idea01]

A chat instead of a dashboard

Sales, bank, reviews and bills live in different apps that owners don’t open during the day. So there’s no dashboard. The owner asks a question, and the answer comes back as cards in one conversation, where the next step happens too.

Restaurant: “How’s my business doing?”
Factory: team data usage
From our earlier iterations
✕ Earlier iterationA dashboard with four tiles
✓ What we choseOne thread: ask, answer, next step
BecauseOwners asked questions. They didn’t browse charts between customers.
Design idea02]

Every choice can be spoken or tapped

Owners talk to it between customers, but not always. Next to a customer or in a loud kitchen, speaking isn’t an option. So every choice the assistant offers works both ways: say it, or tap the card. Either way, the choice comes back as the same outlined bubble, so the thread reads the same.

Restaurant: picking a way to grow
Barber shop: choosing a camera plan
From our earlier iterations
✕ Earlier iterationVoice only
✓ What we choseSpoken or tapped, into the same thread
BecauseKitchens are loud and customers stand close, so owners need to be able to tap instead.
Design idea03]

It brings things up when the owner isn’t busy

The assistant watches the business and raises things on its own. It uses footfall and sales to tell when the shop is busy, holds insights until it’s quiet, and only interrupts straight away when money is at risk. Every notification opens into the conversation with the evidence on screen.

Barber shop: satisfaction is falling
Factory: bill due tomorrow
Design idea04]

Specialists join when the work needs them

The assistant handles everyday questions itself. When a task needs more, it brings in help: a Legal agent drafts a contract, a Recruit agent screens candidates, a Web agent builds a website, and a person takes the calls that need one. Their work comes back into the same thread.

Law firm: contract drafted after the call
Who joins, across the four flows
Legal agentDrafts contracts, like Alta’s
Recruit agentScreens candidates for a new barber
Web agentBuilds and runs the factory’s website
A human expertTakes the call when the owner is unsure
/ Decisions

A yes, then a confirmation

The assistant makes suggestions and the owner decides. Saying yes opens a confirmation that shows what will happen and what it costs. Nothing is ordered or sent until they tap Confirm.

ONE DECISION, START TO FINISH01 IT SUGGESTSThis package adds twocameras. Should I order it?YesNo02 THE OWNER SAYS YESYes, order it03 A CONFIRMATION OPENSConfirm your order2 cameras69.00 AED a monthStarts onApril billInstalledThis weekConfirmCancel04 THEN IT HAPPENSOrderedAdded to the monthly billCamera Package+69.00Total1,269.00 AED
Owners make the real decision on the confirmation screen
What needs confirming
TellsNothing to confirm. It just tells you, like “Your busiest hours are 6 to 11 AM”.
DraftsYou read the draft first, like the contract for Alta.
ActsNeeds a yes, then a confirmation with the cost.
NeverPaying or messaging anyone on its own.
Only actions need a confirmation. Answers and drafts don’t.
/ The four voice states

What the screen does while the owner talks

The orb and the words on screen always show which state the assistant is in, so the owner never has to guess whose turn it is.

State 01IdleReady, and asking one question.
The ownerOpens the app between customers.
The screen
1One plain question fills the screen
2The orb pulses slowly at the bottom
12
12
State 02ListeningEvery word shows up as it’s heard.
The ownerAsks “How’s my business doing today?”
The screen
1Words appear on the right as the owner speaks
2The orb glows and the wave follows the owner’s voice
State 03ThinkingWhile it works, it says what it’s doing.
The ownerWaits, or keeps talking.
The screen
1A named step instead of a spinner
2The orb spins faster
12
12
State 04SpeakingIt says one short line and puts the detail on a card.
The ownerListens, or glances down.
The screen
1One short spoken line
2The numbers show up on a card
/ When it mishears

When it mishears

A busy shop is loud, and owners switch between Arabic and English in one sentence, so it will mishear sometimes. When it does, it shows what it heard or asks, instead of guessing.

HEARDShowmetheمبيعاتsalesthisweekThe word it wasn’t sure of is marked.One tap swaps it.
01Two languages at onceThe word it wasn’t sure about is marked, and one tap fixes it.
HEARD“How did we do last month?”?OctoberLast 30 daysIt asks once, with both answers as buttons.
02Could mean two things“Last month” might be October or the last 30 days, so it asks.
AssistantTHE THREADLegalOrbDRAFTING✕ Stuck on one clauseTNTarik NREAL ESTATE EXPERTIt says so in the thread and offers a person.
03A specialist gets stuckIf LegalOrb can’t finish a clause, you hear about it, and Tarik is offered.
/ Outcome

Launched in 2025

Performance numbers are confidential, so here’s what I can share.

ShippedLaunched to the client’s small-business customers in 2025
The demoBCG X used the four flows to walk new clients through the product
4 of 4Flows finished by the client team in testing, by voice or by tap
/ Next

What I’d work on next

Test with owners who mix languages

Gulf Arabic, English and Hindi in one sentence, in a busy kitchen. There isn’t much research on that yet.

Show what changed on every bill

“AED 340 more than last month” is quicker to check than a full list of charges.

How this was made

Designed in Figma, prototyped with Claude, motion in Google AI Studio

Figma42 screensEvery screen in the four flows, designed as the product shipped.
ClaudeThe prototypeEach flow rebuilt in code. Timing the orb for real showed that a silent second felt broken, so every wait got a label.
Google AI StudioMicro-interactionsThe orb’s states, the named waits and the card transitions, explored in Google AI Studio before they went into the prototype.
How the prototype was built, and where we went back
/ More work