2025–PresentAI Product & Human-Centered AI
Lisa - AI Receptionist
An AI receptionist for small businesses. Owners miss calls because they are busy doing the work. My job was getting them to let Lisa answer.
- 0 to 1
- AI product. Lisa shipped, now bundled in every plan
- 100+
- Character assets for a character-first product
- 25+
- Tested interface states across six eras
- 6 mo
- First working prototype to shipped product
How much time do you have?
Jump to the depth that suits you.
Or just keep scrolling: everything is on this page, in order.
The 1-minute summary
Getting owners to let an AI answer their phone.
- The problem
- For LinkedPhone's 50,000+ small businesses, the owner is also the receptionist. When they can't pick up, the business stops. AI could answer for them, but owners wouldn't hand their phone to something they couldn't predict.
- What I did
- Led product design for Lisa, an AI receptionist, from a working voice prototype to launch in six months: guardrails, onboarding, information architecture, character and the launch site. Engineering owned the voice stack.
- What changed
- Owners teach Lisa by sharing a link, check what she learned, and choose how much she handles: no calls, missed calls only, or every call. Her response gap went from 8–10 seconds to under one.
- What happened
- A three-month staged beta, then general release on mobile, web and tablet, in every plan.
Shipped product · Lisa after setup
The full story · about 20 minutes
Two problems, four decisions.
Each decision opens with a short card: what we could have done, what we did, and what it cost. Read the cards for the argument, or keep going for the work behind each one.
Situation
Where AI already was
Before Lisa, AI only ever arrived after the call or next to it.
LinkedPhone gives small businesses a business phone line. Before Lisa, we had put AI where it was safest: after the call was over.
Call intelligence came first, and I owned it from the design and product side. When a call ended you got a summary, key points, follow-up actions, sentiment and lead details. Adoption was strong, and it taught us something we would need later: people accept AI most easily when it hands them back their own information.
Then we moved one step closer. AI reply suggestions, drawn from your history with that customer, for the moments you don't know what to say. The obvious next step was the one that changed the risk: AI that answers the phone itself.
My role
Design owned the product. Engineering owned the model.
I led design end to end: voice-AI research with engineering, the prompt and onboarding flow, the information architecture, the interface, Lisa's character system, the films and the product page.
Engineering owned model selection, orchestration, telephony and infrastructure. The founders held the vision. Most decisions on this page were made in that three-way conversation, and my part was translating between them: intent into something buildable, constraints into something a business owner could understand.
Lisa runs on mobile, web and tablet with near-identical interfaces, because the design system carried the layouts across without a redesign. This page shows mobile only.
Problem
Too slow, then too convincing
The owner's problem was simple. Building something they would trust with it was not.
A small-business owner is also the receptionist, the salesperson and the technician. When they're on a job, the phone rings out.
“I don't care if it's AI. I just want someone to pick up when I can't.”
Discovery 1
A caller reads a pause as nobody being there.
We built the voice layer in house: speech recognition, language model and speech synthesis, chosen by us and tuned to hold a phone call. I worked with engineering through the comparison, running models against each other and listening to the output.
Every option traded speed against accuracy against intelligence. A small model answered quickly and shallowly; a larger one reasoned well and arrived late. The first working build left 8 to 10 seconds between the end of the caller's sentence and the start of Lisa's. I argued the ceiling had to be 3 to 4 seconds, measured at the caller's ear rather than the server, because that is about how long a person waits before deciding the line is dead. Engineering got it under one second.
Discovery 2
A thin, slow prototype already felt like a person.
Five or six weeks in, the stack was running on a live phone number and I could call it.
The intelligence was thin and the latency was bad, and none of that mattered. Something was answering, and the illusion held anyway. That changed what the risk was. The danger wasn't that Lisa would fail to convince callers. She would convince them too easily, and say things on the business's behalf that nobody had approved.
So the design job had two halves
Make Lisa safe enough to trust. Then let owners hand her the phone one step at a time.
Decision 1
Make her safe first
- Could have
- Tuned her to sound more human and answer anything a caller asked.
- Chose
- Eight rules that turn every unknown into a handoff to a person.
- Cost
- Some calls end with a message instead of an answer.
The guardrails
What Lisa is not allowed to do.
If the illusion is that easy to create, the job is deciding where she has to stop.
01
She only knows what you taught her.
Lisa answers from the business's own information and nothing else. Owners can't loosen this. A model that improvises an answer about your refund policy is a liability with your name on it.
02
When she's unsure, she takes a message.
Faced with a question her knowledge doesn't cover, she collects the caller's details for a human instead of reaching for a plausible sentence.
03
She says she's an AI.
Asked directly, she answers directly. No version of this product protects the illusion.
04
Asking for a human is a command.
If a caller asks for a person, she transfers. She doesn't try to satisfy them first.
05
Frustration escalates.
An angry caller, or one who won't accept her answer, goes to a person. If nobody is free, she takes a message and ends the call cleanly. She never argues back.
06
Some fields are locked on.
Name and reason for calling can't be switched off, so every call produces something an owner can act on.
07
Some fields she avoids.
Voice is a lossy channel. A misheard email looks like a lead and isn't one, so the product steers owners away from asking for it.
08
Nothing enters her knowledge unseen.
Everything she reads is shown back to the owner before she can say it to a customer. Decision 2 is built on this one.
She never fills a gap with something invented. She fills it with a person, or with a message to one.
Decision 2
Teach by link, check in a list
- Could have
- A knowledge-base form: fields for hours, services, prices and FAQs.
- Chose
- Share a link, let Lisa read it and ask only about the gaps. Then review everything she learned as an editable list.
- Cost
- An extra verification step, and some knowledge bases start thinner than a long form would make them.
Why not a form
The business already had the answers.
Owners had a website, a Google profile, a menu, a brochure. Asking them to type it all in again is why setup never gets finished.
“It's all already written down somewhere, I just never put it in one place.”
Chat beat the form in testing because it discloses one thing at a time: you never face a field you don't yet understand. It also changed how setup felt. Teaching Lisa by talking to her is briefing a new hire, which made the character work in Decision 3 possible.
What happens underneath
Verify the source. Fill the gaps. Check coverage.
This is where design and prompt engineering stop being separate jobs. Four mechanisms, each with a cost.
Verify first
Nothing is absorbed silently.
Lisa checks a link resolves, pulls what she found and shows it back. The real failure isn't refusing to answer; it's confidently giving a customer the wrong hours. Cost: one screen.
A question budget
Ten to twelve questions, then stop.
Setup that runs indefinitely is setup nobody finishes. Cost: some gaps get filled later, from real calls.
Waiting as conversation
Slow sites are read in the background.
If a site will take more than about 15 seconds, the chat carries on while reading happens. Cost: two paths through one flow.
A hidden score
Coverage drives Lisa, not the owner.
Below 80, Lisa asks another round of questions; above it, setup completes. Telling an owner their AI is 74% ready makes them anxious about the other 26%, and anxious owners switch her off. They experience the extra questions as thoroughness.
Two surfaces
Conversation to build it. A list to audit it.
Chat is the best way to build a knowledge base and the worst way to check one. So knowledge lives in two places: a chat for teaching, and a structured list of FAQs, links, files and images for review.
Decision 3
A face, not a presence
- Could have
- An abstract voice visual, an orb or particle field, like most voice AI products.
- Chose
- A drawn human character with a name, a job title and a profile, deliberately short of photoreal.
- Cost
- The most technically involved interface work in the project was thrown away, and a feminine default that draws fair criticism.
The whole search in four headers
Vox. Vox AI. Lisa. Lisa Answers All Incoming Calls.
Over six eras the product's own header moved from naming the system, to naming the behaviour, to naming a person doing a job. Most of what I tried along the way was wrong, and the wrong versions are what made the decision.
Eras 1–2 · What failed
The orb felt like being watched.
The first interface was built straight on the architecture: settings, a dashboard, configuration. It worked, and it looked like every other voice-agent builder, which a barber or a contractor can't use.
So we tried presence without a face: an iridescent form that reacted to voice and to the phone's own motion. It was beautiful, expensive on battery, and it felt like HAL. A responsive abstract shape reads as an intelligence observing you. And the name was still Vox, Latin for voice. Two eras in, we were still naming the technology.
Era 4 · The character
Answering the owner's real question: who is going to pick up my phone?
Five families of character: cute mascots, faceless suited figures, glowing abstract forms, fuzzy monsters and a human figure. Only the human figure answered that question.
Short of human
Testers found a human figure friendlier than anything abstract, up to the uncanny valley. She stays drawn and familiar, which also keeps a hundred assets consistent.
A blazer, because she has a job
She reads as professional and as part of the interface, not pasted on top of it.
Brand blue, used twice
Blue everywhere cancelled itself out. The goggles and the shirt alone carry the brand.
She's feminine by decision. In our testing, users found the feminine character more approachable and easier to trust, which matters for a product whose whole job is earning permission. Feminine-by-default assistants are fairly criticised for reinforcing a stereotype, and I don't think that criticism is wrong. Character customisation is the direction, so the default stops being the only option.
The character library and the films were AI-generated under my direction: mood board, brand essence, variation runs, curation and the consistency rules. The direction, selection and system are the design work; the rendering is not.
Era 5 · The profile
Her knowledge. Her jobs. A transfer to the team.
With a character in place, settings became a profile. Labels turned possessive: Anna's Knowledge, Anna's Jobs. And Route Calls arrived, which changed the category. Answering questions and taking messages is a voicemail with a pulse. Knowing when to hand a call to a colleague is what makes her part of a team.
The name
Anna was my name for her. Naming the thing your customers talk to belongs with the people who carry the company's voice, so I wrote a brief instead of an answer: two syllables, friendly, timeless rather than trendy, matching the traits we'd agreed. The founders chose Lisa. Both fit the brief, which is what a good brief should do.
Six eras, side by side
What ended, what failed, what stayed.
Decision 4
One step at a time
- Could have
- One "Turn on AI" switch, or folding Auto Attendant into Lisa and moving every customer over.
- Chose
- Three modes: no calls, missed calls only, all calls. Auto Attendant stays the default, and the rollout followed the same order.
- Cost
- A routing model that has to explain itself, and the mode that asks for everything waited until the last phase of beta.
Where it broke
Two systems answered the phone, set up in two places.
By beta, real businesses were using Lisa, and the structure had stopped explaining the product. Auto Attendant, the call menu, lived in one part of the app. Lisa had her own tab, put there for visibility. Nobody could tell how they related, or which one would pick up.
Every step that got us there had been reasonable. The accumulation wasn't. So I stopped extending the structure and rebuilt it.
“Let it take the calls I'd have missed anyway first. Then we'll see.”
The inversion
Lisa first, Auto Attendant as the fallback, and nobody forced to switch.
Lisa had started as one destination inside Auto Attendant: what happened when the menu ran out. From a product standpoint that was backwards. Flipped, Lisa sits in front of every call and Auto Attendant is what happens when she's off.
We explored moving everyone into Lisa's setup in one go. It would have broken every existing customer's call menu. What shipped is a Setup tab where Auto Attendant is the default and Lisa is something you switch on, for missed calls or for all calls. Existing users keep what they built. Someone who only wants a voicemail greeting can still have exactly that.
The flip was the product decision. Making it optional was the design decision.
- 01No calls
Auto Attendant answers, as before.
- 02Missed calls only
Lisa takes the calls nobody picked up. The owner risks nothing they weren't already losing.
- 03All calls
Lisa answers first and transfers to the team.
Shipped · The mode picker
How the interface got there
The screen stops describing the mode and becomes it.
Across the studies, the routing became a visible chain (incoming call, Auto Attendant, Lisa, message), then the header turned into a sentence: Lisa Answers All Missed Calls. The handling list reorders per mode so the order on screen matches the order on the call, and the metrics change with it.
What the structure became
Built for managing an agent, not configuring one.
Call Handling
Hours, which calls she takes, and what she does on them. One question for the owner, so one place.
Knowledge
What she knows. Unchanged since the first architecture, which is the best evidence it was right.
Performance
What she gives back. Promoted from one branch among six to a top-level section.
You brief her and she reports back. After the inversion, Auto Attendant and the AI receptionist became one tab, called Setup: it had to cover both, fit the bottom navigation, and not collide with Settings, which already meant account preferences.
The rollout
The release followed the same ladder.
Beta ran for three months in widening rounds: the first cohort, then a thousand more people, then ten percent of the user base, then general release. For almost all of it, only missed-calls mode existed. All calls opened in the final ten-percent phase.
That wasn't engineering readiness. We earned the right to ask for every call by first handling the ones businesses were already losing, the same order the interface asks each new owner to follow.
Outcome
What shipped
Lisa is live on mobile, web and tablet, in every LinkedPhone plan. Here is the product in the order an owner meets it.
- Availability
- Every planNo premium tier
- Platforms
- 3Mobile, web and tablet
- To launch
- 6 monthsFrom a working voice prototype
- Staged beta
- 3 monthsWidening rounds, then general release
1 · Switching her on
From a pop-out to answering calls, in one conversation.
Lisa introduces herself in a short video inside the inbox. Setup starts from the Setup tab, where Auto Attendant is still answering. Meet Lisa says what she does, and the teaching chat does the rest. When she knows enough, she goes online and starts answering missed calls.
2 · Setup
One switch: all calls, missed calls, or none.
Once she is on, Setup is where an owner manages her. The top card says who answers first, and one sheet changes it: Lisa takes every call, only the ones the team misses, or none, and Auto Attendant answers as before. The rest of the page follows from that choice.
From top to bottom, the page holds everything she does:
3 · Knowledge and performance
What she knows, and what she did with it.
Knowledge is a library an owner can filter and add to, one FAQ, link, file or image at a time. Performance counts the work: calls answered, transfers and leads captured, with the details behind each number.
4 · Her voice and her jobs
She answers, takes details, hands calls on, in a voice the owner picks.
Owners can try her as a customer before a real caller does. They decide which details she collects and when she transfers a call. They choose her personality, or give her their own voice from about thirty seconds of reading.
The clone asks the owner to read a fresh random passage each time, which guards against replaying someone else's recording without turning setup into identity verification.
5 · Ring groups
She transfers to a team, not a person.
The departments in Call Transfer are ring groups: a named set of teammates with its own ring order. Lisa sends a call to the group that matches what the caller asked for, and the group decides who rings. Each person's profile shows the departments they answer for.
The same groups in the shared design, with a missed-call auto-reply per group:
6 · After the call
“No-one answered” became “Lisa answered”.
The call list marks every call she took, with a Lisa filter. The conversation and the call detail carry the same summary, sentiment, captured details and transcript as before, because that work never changed. Only who took the call did.
7 · Where Auto Attendant went
No answer → Lisa. Press 4 → Lisa.
With Lisa on missed calls, Setup hands the menu to Auto Attendant Answers. Open it and the call menu says so at the top, and every branch that ended in voicemail now ends in her. Inside Default Handling, her card replaces the voicemail greeting, with room for instructions.
8 · In the world
A character can introduce herself.
Because Lisa is a drawn character, making video of her cost almost nothing. She presents herself in a vertical “hire me” film and walks new owners through her own setup. I directed, produced and edited some of the eight to ten reels and tutorials, and directed the team on the rest.
I designed the LinkedPhone website, so the product page was mine too. Its FAQ is long on purpose: each answer retires one fear. Can I turn her off? Who does she transfer to? It sells the outcome and never the model.
9 · Assets
A character system, not a folder of pictures.
Over a hundred assets held together by three rules: same setting, same costume, same background. The work reached past Lisa too: her profile pattern became the pattern for contact and team-member profiles, and putting a character in the interface exposed spacing and dark-theme colours that were fixed across the whole design system.
Reflection
What I'd change
What I learned · 1 of 4
A score out of a hundred told owners nothing they could use.
The first dashboard put Response Quality and Caller Satisfaction on the home screen, out of a hundred. I expected the numbers to build trust. People found them vague: 86 out of 100 didn't say what Lisa had done, or what to change.
Every later version replaced verdicts with counts of work done: calls answered, transfers, leads captured, each with the calls behind it. The coverage score from Decision 2 moved underneath, where it changes Lisa's behaviour instead of the owner's mood.
What I learned · 2 of 4
Realism was the easy part. Knowing when to stop was the design.
A thin, slow prototype on a live number already felt like a person. I had planned to spend the project making her more convincing. The real work was the opposite: deciding where she hands over.
So the guardrails came before the personality. Owners can try her as a customer before a caller does. When she is unsure, the call goes to the team, and the call record says so.
What I learned · 3 of 4
Owners didn't want AI. They wanted to know who picks up.
The orb was meant to feel alive. Owners said it felt like being watched. The character worked because she answered the question they were actually asking: who is going to pick up my phone?
Once she had a face and a name, the rest of the product could speak in her voice: her knowledge, her jobs, Lisa answered.
What I learned · 4 of 4
Nobody hands over their phone in one go.
“Let it take the calls I'd have missed anyway first.” That line from research became the structure. Missed calls only is the first step, because the owner risks nothing they weren't already losing.
Making it optional mattered as much as the flip itself. Existing call menus kept working, and Lisa simply took the place voicemail used to have.
What I'd do differently
I'd audit the structure against the interface on a fixed cycle, from the first beta. We rebuilt it only once it had stopped moving, and by then two systems were answering the same phone. I now run that check before the structure breaks, not after.
Closing
A summary after the call, then a suggestion beside it, then a voice inside it.
More work
-
Auto Attendant
Call Routing
Call routing a small business could actually set up: one system where basic and advanced setup are the same screen at different depths.
View case study · 12 min
-
Website Transformation
Design → Publishing
A 500+ page website moved to Framer and rebuilt around a new structure and content system, so design could publish.
View case study · 12 min
-
YONO by SBI
App Redesign Concept
A first-principles redesign exploring how mobile banking could become clearer and easier to use.
View case study · 20 min
-
Cyberpunk 2077
Website Concept
A launch-website concept built around the game’s world, atmosphere and visual language.
View concept · 5 min
-
Snake
Browser Game
Snake, reimagined for the devices we hold today: a 3D space worm you steer with a thumb, a mouse, keys or a gamepad.
View case study · 10 min