Beyond the Prototype: LA Tech Week

LA Tech Week: Tuesday, October 13, 2026, 6:00 p.m.

Welcome and framing

Moderator - John Moore

Good evening, everyone, and welcome to Beyond the Prototype: Building Voice AI That Wins.

It has never been easier to build something impressive enough to pull up on a phone and show at a happy hour. But a compelling demo is not the same thing as a product—and a product is not automatically a business.

Tonight we’re going to focus on what happens after the demo. How do you know whether an agent solves a problem people truly care about? How do you make the economics work when every interaction has a cost? How do you build a system that performs consistently enough for real customers? And in a world where anyone can make a bespoke assistant for themselves, what makes a product worth paying for—and difficult to replace?

We’ll look at those questions through entertainment, avatars, consumer products, games, and go-to-market strategy.

Introductions

Geremie Camara is Chief Creative Officer at Live Play Mobile, where the team has built live and AI-powered hosts for social games. Geremie brings a career in entertainment and game-show production to the question of what makes an artificial host genuinely engaging.

Lina Colucci is the founder and CEO of Lemon Slice, which builds highly expressive interactive avatars, including the technology behind an interactive Theodore Roosevelt experience.

Kelly Nyland is an independent consumer hardware and AI operator, founder, advisor, and angel investor. Across more than 160 shipped products, she has helped turn unfamiliar technologies into consumer experiences, including launches for Parrot’s AR.Drone, Sphero’s Star Wars BB-8, and Snapchat Spectacles. Her current work explores action-oriented personal agents and the missing connective tissue between digital and physical AI.

Ross Ingram is part of Sixth Wall at Hasbro, bringing the perspective of a global company that has to balance emerging technology, beloved intellectual property, safety, and products that can operate at real consumer scale.

What is one product or experience you have helped put in front of real users, and what did shipping it teach you that a prototype could not?

Where agents have real product market fit

MODERATOR

Let’s start with the most important question: Where have you actually seen product-market fit for agents in production—not interest, not press, and not a great demo, but sustained user behavior?

To Kelly

You describe your work as turning strange new technology into things people actually love, and you argue that innovation and adoption often arrive inside a “Trojan horse of fun.” Across products such as AR.Drone, BB-8, and Spectacles, how could you tell when technical fascination could become a consumer behavior with the potential to endure?

Fascination and value are different things. Fascination comes from novelty, and novelty’s only job is awareness. The gimmick gets people to look, but it doesn’t make them stay.

A product has to give people a reason to stop. On TikTok, you get 3 seconds. A mall is no different. When I launched the first app-controlled drone with Brookstone, we asked the engineers to build a ‘flip’ button into the app — a trick to grab mall visitors’ attention and draw them in.

There is rarely a single moment when fascination flips into everyday behavior. Some might argue we’re sitting in the middle of a ‘big flip’ right now — watching AI go from novelty to habit. But even a big flip is one chapter in a much longer story. Big technologies are decades-long storylines. Rosie the Robot debuted with The Jetsons in 1962, and it will likely take another 10–20 years before humanoids reach some level of adoption in our homes. So the useful question is where your product sits on that storyline right now, and what you can newly show off.

Value has to show up at every stage, not just at the end of the vision. Each step needs its own product-market fit. Plenty of startups built and funded personalized AI companions and assistants early. But users had to get comfortable asking ChatGPT everyday questions first. Only then did they get curious about using AI to sort through their feelings about their relationships. “Value” means two concrete things. Is there a benefit? And what is the trade-off — what does it cost, and what does it replace? The real competitor is the alternative people already have. An in-home humanoid isn’t competing with another robot. It’s competing with a housekeeper.

Examples I have seen or had a hand in:

Tesla — value on day one

I didn’t buy a Tesla because it was cool. I needed a car, I didn’t want to pay for gas, and I bought it on an app in under thirty minutes. Then I went to pick it up and there wasn’t a human involved. As a woman — no creepy, pushy salesperson! Just peace.

When I got my first flat tire, they sent a mechanic to my home and changed it in 20 minutes. After my first accident, I filed the insurance report right inside the app in 5 minutes. That experience hooked me for life.

And now — 12 years after Elon announced the first version of Autopilot in October 2014 — it has driven 98% of my last twenty trips, parallel parking included.

I bought it to stop paying for gas. I stayed because they came to my home and changed my flat tire. I’m a forever customer because I have all but stopped driving, and I find myself more at peace in LA traffic. The futuristic part only paid off years later — but there was a plain reason to buy on day one, and a reason to stay at every step in between.

Spectacles — the counter-example

Ten years ago I launched Snapchat Spectacles — the first face camera, made for teenagers. We staged the launch as outdoor theatre: the only way to buy a pair was from a vending machine we dropped into some of the world’s most beautiful, hands-on places. Our mission was to show that seeing both of your hands in your own content was compelling.

And it still took me ten years to find a compelling reason to wear camera glasses on a regular basis. I play beach volleyball. Seeing my own passes, sets and spikes gave me a new perspective on how I play the game and where I can improve.

That was it! That’s what finally got my Metas on my face. Ten years of fascination before one moment of real value for me as a player.

Drones — value at each step

The flip button got people to stop. It isn’t what they bought.

When I launched the first consumer-grade drone with streaming video, our customers bought the gift of flight. That was the whole value. When we released 2.0, we added recording — and now they could share their footage with viewers around the world.

Version one sold flight. Version two sold sharing. Each version had to earn its own reason to buy.

Humanoids — the live case

I think billionaires buy them first, as toys — novelty items, not for the value they really create. And that’s fine. That’s the Trojan horse.

The early ones will be teleoperated. They’ll need a white-glove onboarding team and ongoing support just to function. A humanoid that puts dishes away and folds laundry is competing with a housekeeper who comes daily for just above minimum wage. And right now, the housekeeper wins on reliability, cost and trust.

As founders, we have to get really honest with ourselves about the alternatives to what we’re offering. And not just the direct ones — the adjacent categories too, anywhere the solution to our problem overlaps.

But in three to ten years, as everyday consumers begin to buy them, we start affording services we don’t afford for ourselves today — a personal chef, a housekeeper, a housecleaner. The toy pays for the road to that.

Suggested follow-up: When does fun become an effective adoption wedge, and when is it merely a novelty trap?

Fun is a wedge when it lowers the cost of trying something strange. It’s a trap when the fun is the whole benefit.

Novelty traps aren’t bad. They’re required. The first thing we do with any new technology is whatever makes our brains itch. Program a robot to dance. Use AI to make a new genre of adult film content. Use a drone to film whatever the neighbor is doing in the backyard. Adoption is really a series of novelty traps.

A trap becomes a wedge when the magic ‘trick’ is intimately tied to the value the product has to offer. How fast that happens comes down to how ready the market is, how acute the problem is, and how naturally the solution fits.

So my test is simple. When the fun wears off, what’s left? Who gets the most value out of what’s left? And how many of those people are there?

The economics from usage to margin

MODERATOR

Traditional SaaS trained many founders to expect very low marginal costs. Agents can invert that model: more engagement can mean more inference, speech, video generation, memory, moderation, and real-time delivery costs.

To all: How should an agent company align revenue with usage so its best customers do not become its least profitable customers?

To Kelly

Your career spans consumer hardware, AI, commerce, and a venture-backed startup. Hardware forces teams to account for manufacturing, inventory, channels, returns, and support; agents introduce recurring compute, model, and service costs. What economic discipline should agent founders borrow from hardware—and where does the analogy break?

Bill of Materials aka BOM (COGS or Unit Economics)

A hardware company knows the price of every screw before it builds anything. An AI company should know the cost of every piece that goes into one answer: each model call, each token, each lookup.

What’s In The Box

With hardware it’s obvious, because it comes in a box. With AI you have to choose: are you selling a finished job (one report), a month of access, or a bucket of credits? That choice decides whether you make money.

What good looks like from the buyer’s side: tell me what’s in the box. Gemini gives me “three videos a day” or Dreamina tells me up front “this will cost $40, approve?”. I know what I’m buying before I pay.

Where the analogy breaks?

The cost never ends. Once a drone leaves the factory, it costs you almost nothing more, even if the customer flies it every day. With AI, every use costs you again, so the customer who loves you most costs you the most.

There’s no way to return a bad result. A bad drone goes back to the store, and the company planned for that. A bad AI answer still uses up the customer’s credits. They can’t hand it back, so they either ask for credits or quietly leave.

Follow-up: Your current research emphasizes action-oriented systems and proof of completion. If the product takes responsibility for an outcome rather than merely offering advice, should the company price the completed job instead of the conversation?

Yes — price the job when you can prove the job.

To me, a job means the model goes A to Z, and doesn’t stop at Q unless it needs clarification. The reference to action-oriented models is a concept Ross and I worked on called HAMs (or Healing Action Models), where agents take responsibility for outcomes.

Recent personal example with voice agent

The most compelling voice AI experience I’ve had was calling a Japanese restaurant in Detroit called Hiroki-San. My girlfriend and I were running 15 minutes late. The AI told me it would text the host on my behalf, and not to worry — we were within the grace period for reservations. It situationally addressed my issue and solved it.

My guess is that’s one of the top 3 reasons people call that restaurant. So price that. Include up to 100 calls for late arrivals, reservation changes, and questions about cakes and corkage fees. Then show the return: human hours saved, or reservations won back. That kind of specialization in skills is where I get really excited.

There’s still a grey area between making and doing. Researching, writing, making — easier to price as a job. Booking, paying — harder.

And not everything is a job. I paid forty dollars for one video — I handed it a bunch of UGC and it reformatted the whole thirty seconds for me. I never thought about tokens. I bought an output. Claude Code, I pay a hundred dollars a month and I rarely find its edges. I’m not buying a job there. I just trust the model to think.

One catch. If you price the outcome, you own the failure. In my research, 93% of people required proof of completion before they’d hand a task off. That’s the returns line. Hardware founders budget for it. Agent founders have to start.

Fast round: "The unit-economics spreadsheet is lying to you if it leaves out ___."

The brain the stack and production reliability

MODERATOR

Creating a basic agent is easy. Creating a “brain” that behaves well, consistently, across thousands of interactions is a serious engineering and product challenge. And users do not separate the stack: if the voice is slow, the avatar feels wrong, the memory fails, or the model says something inappropriate, the entire product gets blamed.

Core question to all: What broke—or nearly broke—when you moved from a controlled prototype to production?

To Kelly

You have written about the invisible burden AI places on users: they must brief the system, supply its memory, verify its work, and carry output across tools. What does “production-ready” mean if the user is still doing that much of the product’s operational work? Where should the product absorb that burden?

What users go through today:

A product makes a promise, so they try it.
They learn its limits the hard way: “Oh, it can’t do that yet.” “Oh, I need to do / give it X, Y, Z first.”
Eventually they settle on what it’s good for, at a quality, price and frequency that works for them.

Where the product should take that work off them:

Help write the brief. The “enhance my prompt” button in Google Flow does this. So do the clarifying questions Claude asks before it starts or as it’s working.
Put a human in the loop where the model is weak. Every early product does manual work behind the scenes, so agent companies should find those spots too.
- Approval before acting. The agent does the work, then stops and asks me to confirm before anything irreversible — paying, booking, sending. “This will cost $40, approve?” is exactly this. I’m the human in the loop.
- Escalation when the agent is unsure. The agent handles the routine cases and hands off to a person when it hits something outside its range. That’s what happened when I called Hiroki-San: the AI took my call, then texted the human host to act on it.
- A person doing the work behind the curtain. Early on, the team quietly completes or fixes what the model can’t, and the customer just sees a finished job. I did this at full scale. My co-founder and I answered texts around the clock for a year before we built our text message commerce engine. Teleoperated humanoids are the same idea in hardware.
- Review after the fact. People check a sample of the agent’s outputs, catch the failures, and feed the corrections back so the product improves.
Set expectations up front. Tell people what the product is good for instead of letting them discover the edges.

My personal example: AI is weakest at visuals that involve a real product. The model knows how people move, but it doesn’t know how my product looks, how it moves, or how it sits in a hand. So my fix is manual. I give samples to dozens of creators with a recipe, take the best clips, and use those as the reference the model works from.

Suggested follow-up: If persistent context is what turns a generic agent into a personal one, which context should be portable and user-owned—and which information should the system decline to retain?

Beyond one to one chat

MODERATOR

Much of the agent market still looks like a person staring at a chat box. How do we create experiences that are genuinely social, participatory, and entertaining rather than another one-to-one conversation with an assistant?

To Kelly

You argue that chat is an entry point rather than the final interface, and that every new form factor—voice, ambient devices, wearables, and eventually robotics—moves the learning curve rather than eliminating it. Which form factor has the greatest potential to take agents beyond the chat box, and what new burden does it create for the user?

Don’t invent a new behavior. Copy a real one.

I usually don’t try to invent new behaviors. I observe them and take advantage of them. So if you want to build an engaging product around voice, ask how people already use voice with each other. They send voice memos to friends. They leave voice memos in the group chat. They send memes back and forth.

I’ve stopped texting my friends. They get a long voicemail from me. Some of my girlfriends call it their podcast on the group thread. That’s what a social AI product should look like. Not a chat box.

The burden is always the same: learning the edges.

Every new interface asks the user to figure out what it can and can’t do. We’ve all learned to press 00000 on a phone tree to get a human — because voice trees don’t work.

Do it by hand first.

Before you build, have the actual conversations with your customers. Before I built a conversational commerce engine, my co-founder and I took shifts — 24 hours a day, 7 days a week, for one year — answering questions and buying anything online for any customer who texted us.

There’s a lot of AI slop out there because people forget you still have to do the manual work to get a high-quality outcome. Until you don’t. So ask yourself: could your model sustain that conversation in your domain? Where are its limits?

Know where you’ll fail, and make it charming.

We’ve all seen a really cute, clever 404 page. I always advocate for knowing where you will fail, and using those moments to delight the customer. Build Easter eggs into the places where your agent falls short.

Suggested follow-up: How can play teach people a new interaction model while reducing the emotional and cognitive tax of talking to a computer?

The commercialization test

MODERATOR

Let’s close the main discussion with a rapid-fire test. Short answers—one sentence each.

What is your simplest litmus test for a great agent product?

What metric would you put on the main dashboard?

What is one thing founders should refuse to build, even if the demo looks amazing?

Which is the bigger moat: data, distribution, experience, trust, or operational execution? Pick one.

Final main-stage question to all

Imagine a founder asks you three questions: Can I build it? Should I build it? Is there a market for it? Which one do you insist they answer first, and why?

Beyond the Prototype: LA Tech Week

Welcome and framing

Introductions

Where agents have real product market fit

The economics from usage to margin

The brain the stack and production reliability

Beyond one to one chat

The commercialization test

Tell me whatyou’re working on.

Tell me what
you’re working on.