Dylan Fox

1. Key Themes

Voice AI Has Crossed a Reliability Threshold That Makes It a New Data Capture Primitive

Dylan Fox argues that voice has crossed a fundamental reliability threshold for the first time in history, making it a genuine data capture layer — not just an interface.

"Voice is a reliable form of data capture now. Like, for the first time probably ever in the past year. It's like a reliable form of data capture... that means like you can actually talk to your computer now. You can actually have it listen in." 00:18:46

This is the enabling condition for everything else in the episode — healthcare scribes, robotics, consumer electronics, field service tech.

Coding Agents Have Expanded AssemblyAI's TAM by 100x

Fox identifies coding agents as the third and newest macro trend accelerating voice adoption — one that has structurally changed who AssemblyAI's customer is. The platform is no longer just for engineering teams at product companies; it is now accessible to any business.

"We're seeing this huge inflection now of like you have enterprises, you have small businesses that can build with API infrastructure now. So I think about this as like our TAM has just increased by 100x because we're not just selling to engineering teams within product companies. It's now like anyone." 00:00:27

He gave a concrete example of a lawn care chain building on the API using tools like Lovable, Claude Code, Cursor, and Replit.

Voice AI Infrastructure Is Half Model, Half Scaling Engineering

Fox draws a sharp line between creating model weights and building the full infrastructure stack. He argues this is misunderstood by most observers and is where real moats are built.

"When we say infrastructure, I think a lot of times people think like, oh, you're just creating the model weights, right? Like you're just creating models. That's a part of it. But those other parts are equally important to us, the inference, the orchestration, the developer and agent experience." 00:11:22

"It's probably like half our work is spent just making our infrastructure more and more scalable." 00:13:43

Data Alignment to Vertical Use Cases Is the Core Competitive Moat, Not Benchmarks

Fox repeatedly returns to the idea that public benchmarks are easy to game, and the real differentiation is alignment of training data to specific real-world use cases.

"It's very easy to optimize for like a public open source benchmark. It's very hard to optimize across all these real world applications. And so we spend a ton of time on evals. We have like a whole team internally that works on evals, constantly evaling our models across like a million different metrics, million different data sets, cross all these languages." 00:33:39

He illustrates this with the contrast between police body cam footage (capture every speaker) vs. McDonald's drive-through (ignore background speakers) — same model capability, completely different alignment requirement.

Voice Is an Additive Modality, Not a Replacement for Existing Interfaces

Fox explicitly rejects the "keyboard and mouse is over" hot take and frames voice as a new dimension layered on top of existing interfaces, not a replacement.

"Touchscreens didn't replace keyboards. It's like you have both... I think voice is going to be another dimension. That's probably the way I think about it. Voice is going to be a new dimension. But it's going to be an additional dimension." 00:20:35

He cites the liberating quality of voice — enabling passive, screen-free computing — as the real prize, not interface substitution.

On-Device Voice Models Are the Next Frontier for Consumer Hardware Expansion

Fox signals that AssemblyAI is actively building models that run on extremely low-powered hardware — phones, TV remotes — which will unlock a wave of ambient consumer electronics applications.

"We're, for example, working on on-device models. So models that can run like on a phone or on a, you know, really low powered piece of hardware, you know, like a remote control for a TV or something. And that's going to expand a lot of, a lot of these applications." 00:37:54

The Voice Agent Deception Problem Is an Unresolved Industry-Level UX Crisis

Fox is unusually candid that the current default for voice agents — trying to pass as human — is a design pattern the industry must move past, and that it creates genuinely unsettling user experiences.

"There's this interesting difference where every text-based agent that you talk to... is always like, 'Hey, I'm the AI SDR, I'm the AI support agent.' Voice is different where like today, when you're building a voice agent... for the most part, you're trying to trick the human into believing that it's also a human." 00:38:16

"If you ever talk to a voice agent and then like two minutes in, you realize you're talking to AI, it's just like a weird experience. And I think that's something that we still have to figure out." 00:41:14

Speaker Disambiguation Is the Blocking Problem for Humanoid Robotics and Voice Agents

Fox identifies a very specific and underappreciated technical bottleneck that is currently limiting humanoid robots and voice agents: the inability to identify who among multiple speakers should be listened to.

"If you have three people standing next to the robot, it doesn't know who to listen to. And it has a hard time disambiguating who's saying what. And so it all just gets kind of jumbled and merged. This is a big problem." 00:23:00

2. Contrarian Perspectives

Voice Is Not Commoditized — It Has More Differentiation Gaps Than Most Technologies

The conventional wisdom is that speech-to-text is a solved, commodity problem. Fox directly refutes this.

"A lot of people will think, like, oh, voice, like, what people have been telling me forever is, like, oh, voice, you know, isn't that solved? Like, isn't that just, like, commoditized technology? But, like, it's definitely not because there's so many gaps across languages, across different verticals and applications and domains that there's big rooms for improvement in still." 00:27:31

He backs this up with the language localization challenge — noting that local voice AI vendors in specific countries typically outperform global ones because linguistic and cultural policy alignment is a deep, non-obvious expertise.

Self-Hosting and Sovereign AI Is Less Critical in Voice Than in Text — For Now

While sovereign AI is a hot topic, Fox pushes back on its urgency in voice specifically, arguing the technology is moving so fast that companies that branch off to fine-tune their own models find themselves rapidly falling behind.

"We've seen some customers maybe six months ago or a year ago, they'll take an open source model and fine tune it or something. And then, you know, they're now like, okay, this is, like, shit. This is, like, really behind. And it's, like, a nightmare to maintain." 00:17:44

His implicit argument: in a fast-moving foundational layer, staying on the best-maintained API beats owning your own model.

The Right Product Vision for Voice Is "AWS for Voice," Not Application Layer

In a world where most AI companies are racing toward vertical applications, Fox is deliberately staying infrastructure-only — and argues this is the correct strategy, not a limitation.

"We're 100% focused on voice AI infrastructure. So we don't do anything at the application layer... We're like the AWS for voice capabilities. That's the like product vision that we have." 00:10:26

The contrarian bet here is that being the foundational layer — and not competing with customers — compounds in value as the ecosystem grows.

Disclosing AI Identity Actually Destroys Conversion — The Industry Is Hiding This

Fox reveals a data point that the voice agent industry generally does not discuss publicly: disclosing AI identity upfront causes users to hang up, so companies are architecturally incentivized to deceive.

"We see this, if our customers, if they're building a voice agent and you disclose up front that you're an AI, people just hang up versus if you don't, people continue." 00:39:40

This creates a structural tension between user experience and ethics that the industry has not resolved.

75% of AI Model Quality Is Determined by Data, Not Algorithms

Fox deflates the mystique around AI architecture breakthroughs, arguing data curation is the dominant factor — a less exciting but more operational truth.

"My experience has been, it's really probably like 75% of it is like the data that you're training on. Like I would say for any AI model, it's like there's always like, you know, like these like step functions and like algorithms and architectures and stuff. But the data is just so important." 00:31:41

3. Companies Identified

AssemblyAI

Voice AI infrastructure platform. Described as the "AWS for voice" — providing speech-to-text, voice agent models, and speech understanding APIs. Serves ~1 million developers, handles 120 million voice conversations per week (4x YouTube's daily volume), with 800% growth in weekly conversations over three years and 40% of developers joining in the last year alone.

"On a given week, on a peak week, there'll be something like over 120 million conversations, voice conversations going through our platform, over 2 million hours of voice, which as of December of this past year was 4x or a little over 4x the amount of daily volume that's going to YouTube." 00:02:32

Granola

AI note-taking product for professionals and companies, built on AssemblyAI's infrastructure. Cited as a flagship customer demonstrating the platform's real-world deployment at scale.

"We see, you know, our customers like Granola, right? Granola building these like amazing products that are helping companies and professionals like run their businesses." 00:05:16

Tollands (Replika-type AI companion)

AI companion platform with hundreds of thousands of users engaging for hours daily. Cited for its clearly-labeled AI avatar model — presented as a contrast to deceptive voice agent design patterns.

"We have customers like Tollands that are building, we were talking about it before, like AI companions that hundreds of thousands of people are talking to for like hours every day." 00:05:16

Athelas

Health tech company building AI scribes for doctors, cited as a key customer in the healthcare vertical.

"In the healthcare space, Athelas, Camur, which is a big health tech company building AI scribes for doctors." 00:05:16

Camur

Health tech company building AI scribes for physicians, mentioned alongside Athelas as a healthcare infrastructure customer.

"Athelas, Camur, which is a big health tech company building AI scribes for doctors." 00:05:16

Ciro AI

Startup building AI for field service technicians (plumbers, HVAC). Uses AssemblyAI to listen in on service visits and deliver post-visit sales coaching, reportedly resulting in ~20% higher take-home pay for field technicians.

"They're creating an AI for field service tech. So if you're a plumber or HVAC technician, they have an app you can run, and it will listen in on your service visit and then give you feedback after for, like, sales coaching... They're earning, like, 20% more take-home pay or something." 00:19:17

Haygen

AI video/avatar company, mentioned as an AssemblyAI infrastructure customer in the platform overview.

"Assembly AI is a voice AI infrastructure layer millions of developers build on... companies like Granola, Haygen, Ashby, and ClickUp." 00:28:29

Ashby

Recruiting software company, cited as an AssemblyAI infrastructure customer.

"Assembly AI is a voice AI infrastructure layer millions of developers build on... companies like Granola, Haygen, Ashby, and ClickUp." 00:28:29

ClickUp

Project management platform, cited as an AssemblyAI infrastructure customer.

"Assembly AI is a voice AI infrastructure layer millions of developers build on... companies like Granola, Haygen, Ashby, and ClickUp." 00:28:29

Matic (robotics)

Consumer robot vacuum company. Fox mentions his children interact with it, used as an anecdote illustrating how humans (and children) naturally engage with autonomous machines.

"We have a Matic robot. And they, like, love to mess with the Matic robot. They broke one." 00:22:33

Lovable

No-code / AI coding agent platform. Cited as one of the tools small businesses are using to build on AssemblyAI's API without traditional engineering teams.

"A lot of these small businesses are automating parts of their back office using Lovable or Claude Code or Cursor or whatever, Replit." 00:08:51

Cursor

AI coding environment. Cited alongside Lovable, Claude Code, and Replit as tools democratizing API development.

"A lot of these small businesses are automating parts of their back office using Lovable or Claude Code or Cursor or whatever, Replit." 00:08:51

Replit

Online coding platform. Cited as another tool enabling non-engineers to build on AssemblyAI infrastructure.

"A lot of these small businesses are automating parts of their back office using Lovable or Claude Code or Cursor or whatever, Replit." 00:08:51

Vercel

Web deployment platform. AssemblyAI used Claude to rebuild their entire website off Webflow and deployed it on Vercel, enabling non-engineers to make website changes via Slack.

"We had Claude rebuild our whole website off of Webflow and just deploy it on Vercel. And now anyone, you know, people in Slack can just make changes to the website." 00:16:35

Brex

Intelligent finance platform (sponsor). Cited as being used by Vercel, OpenAI, Anthropic, Granola, and Deepgram, as well as Sourcery itself.

"The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex." 00:13:50

Deepgram

Mentioned as a Brex customer — a voice AI competitor to AssemblyAI, notable for appearing in a sponsor read alongside AssemblyAI's own customers.

"The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex." 00:13:50

MongoDB

Database platform (sponsor). Cited as used by 75% of the Fortune 100 and leading AI-native startups. Their Voyage AI acquisition provides vector search and embeddings.

"It's why 75% of the Fortune 100 and leading AI-native startups run on MongoDB." 00:28:01

Smith Point

VC fund founded by Keith Block (former Salesforce co-CEO). Invested in AssemblyAI's Series C. Fox singles them out as particularly valuable operator-investors.

"Keith Block and Smith Point that invested in our company at our series C. They're, you know, operators from Salesforce. They've started their own VC fund." 00:44:18

Excl (Excel/Accel)

VC firm. Steve Laughlin cited as an investor who pushes AssemblyAI in meaningful ways.

"Steve from Excel, Steve Laughlin, Rebecca from Insight. Like they're just always like pushing the company and pushing me and in, in great ways." 00:44:18

Insight Partners

VC firm. Rebecca (last name not stated) cited as a key investor alongside Excl.

"Steve from Excel, Steve Laughlin, Rebecca from Insight. Like they're just always like pushing the company and pushing me and in, in great ways." 00:44:18

Cognition

Scott Wu's company, mentioned in the outro as a future podcast guest in the Raise series.

Cerebras

Andrew Feldman's AI chip company, mentioned in the outro as a future podcast guest.

Lumentum

Michael Hurlston's photonics/optical components company, mentioned as a future guest.

Salmonova

Rodrigo Yang's company, mentioned as a future podcast guest.

BlackRock

Asset management giant. Tony Kim cited as a future podcast guest from the Raise series.

4. People Identified

CEO and co-founder of AssemblyAI. Self-taught programmer turned AI infrastructure founder. Went through YC in 2017 in the very first AI batch. Has spent nearly a decade focused on voice AI infrastructure, staying deliberately at the infrastructure layer rather than building applications.

"For me, it was like this technology is going to get better. It's going to enable like so many new applications and voice to be something you can really build with... And it just felt like something really cool to work on." 00:06:09

Daniel Gross

Started the AI batch at YC in 2017; now at Meta. Cited as the person who created the first YC AI batch that AssemblyAI went through — a historically significant moment in the AI ecosystem.

"It was Daniel Gross, who, if you know of Daniel now at Meta, he started the AI batch at YC back when we went through in 2017. And it was me and like five other companies and we got, you know, $100,000 in GPU credits." 00:03:48

Keith Block

Co-founder of Smith Point VC, former Salesforce co-CEO. Investor in AssemblyAI's Series C. Fox singles him out as an exceptionally valuable operator-investor.

"Keith Block and Smith Point that invested in our company at our series C. They're, you know, operators from Salesforce. They've started their own VC fund and like, they're just, you know, so..." 00:44:18

Steve Laughlin

Partner at Accel. Cited by Fox as one of the investors who pushes AssemblyAI in meaningful, constructive ways.

"Steve from Excel, Steve Laughlin, Rebecca from Insight. Like they're just always like pushing the company and pushing me and in, in great ways." 00:44:18

Rebecca (from Insight Partners)

Investor at Insight Partners (last name not given). Cited alongside Keith Block and Steve Laughlin as a key investor who actively pushes the company forward.

"Steve from Excel, Steve Laughlin, Rebecca from Insight. Like they're just always like pushing the company and pushing me and in, in great ways." 00:44:18

Max Cook

Sector head at Coatue. Cited for the hot take that the keyboard and mouse era is over — the thesis that voice will replace traditional human-computer interfaces.

"Max Cook, who's a sector head at Coatue, said. He thinks the keyboard and mouse is over." 00:18:25

Ben (AssemblyAI engineer)

Senior engineer at AssemblyAI, singled out by name as a key person behind the efficiency and scalability of the platform.

"People like Ben at our company. We have an amazing team of engineers, of researchers that just operate so closely to customers that they really understand like how this stuff is being deployed." 00:12:47

5. Operating Insights

Build a Personal AI Agent With Full Context Access as a CEO Forcing Function

Fox describes building his own personal agent ("Dylan Claw") that has access to his meeting notes (powered by AssemblyAI transcription), his slides, and company data — and uses it to iterate on internal materials in real time. This isn't aspirational; he describes doing it live during an onboarding session.

"I was talking through a part of the deck that was, like, a bit clunky. And then because my agent has access to the slides and can make new slides and was, you know, has the notes from that meeting powered by Assembly, I was, like, hey, go create a new version of the deck that is a bit easier to walk through and, like, look at the transcript from that most recent onboarding to make some changes... And then, like, 20 minutes later, it's done, and it's perfect." 00:15:43

The operational principle: deploy AI agents with full-context access (meeting transcripts + documents + code) and use them as real-time execution tools, not just query tools.

Deploy a Company-Wide Agent With Code Commit Access to Remove Bottlenecks

AssemblyAI runs a single agent that has access to all company metrics and information, and can submit pull requests across all their products — accessible to the whole company, not just engineers.

"We have an agent that has access to all the company information. So our metrics and the whole company uses it. It can submit PRs in GitHub across our products." 00:16:09

Rebuild Your Website on a Stack Where Any Employee Can Edit It via Slack

AssemblyAI used Claude to migrate their entire website off Webflow and onto Vercel, then wired it so any employee can make copy or content changes directly from Slack — eliminating the engineering bottleneck for marketing and product communication updates.

"We had Claude rebuild our whole website off of Webflow and just deploy it on Vercel. And now anyone, you know, people in Slack can just make changes to the website. So you, like, get off a customer call, you're like, oh, this part needs updating. And it's like, boom, you can just change it." 00:16:35

Stay Radically Close to Customers as a Research Organization — Not Just as Sales

Fox attributes AssemblyAI's engineering efficiency directly to researchers being embedded with customers rather than optimizing for academic benchmarks. He frames this as the core difference between a product research team and a lab.

"You have to know, all right, who are the customers? How are they using this? What do they care about? What errors are like break their application and what errors like don't matter? And then how do you optimize both your models and your infrastructure for that?" 00:13:16

Maintain a Continuous Release Cadence (Every Few Weeks) Instead of Big Launches

AssemblyAI ships model updates every couple of weeks, not every six to twelve months. Fox frames this as a core customer retention mechanic — the platform is continuously improving under customers rather than requiring them to migrate.

"We release model updates like every couple of weeks. And that's the big benefit that our customers have when they're building on our platform is like, it's literally like constantly getting better. It's not like every six months or every year there's an update." 00:32:09

6. Overlooked Insights

Local Voice AI Vendors in Individual Countries Are Systematically Outperforming Global Players — and Are Acquisition Targets

Fox drops a very brief but highly significant observation about the global voice AI market: local vendors in specific countries are consistently beating global models because they understand the linguistic and cultural nuance needed to properly align training data. He raises the acquisition question and then moves past it quickly.

"If you look at, like, local voice AI vendors in, you know, certain countries, like, they actually are typically the best because it's, you know, they speak the language. They understand it. And they can more quickly identify, like, which data is good and stuff to train on." 00:26:37

This is a non-obvious investment theme: there are likely undervalued, defensible regional voice AI companies in non-English markets — particularly in languages with complex tonal, dialectal, or script-based challenges — that are building genuine moats through language expertise that global players cannot easily replicate at scale. These companies are natural acquisition candidates for platforms like AssemblyAI as they try to win globally, and may be overlooked by Western investors who don't have visibility into those markets.

Humanoid Robot Companies Are Already Paying API Customers — and Their Core Problem Is Software, Not Hardware

Fox mentions almost in passing that "a lot of" humanoid robot companies are currently using AssemblyAI's APIs, and that their primary blocking problem is not mechanical — it is voice disambiguation (understanding who among multiple nearby speakers to listen to). This is a software problem, not a hardware problem, and it is unsolved today.

"This is where what we do is so important is because I'll tell you one of the main problems the humanoid robots face today, because a lot of them are using our APIs. If you have three people standing next to the robot, it doesn't know who to listen to." 00:23:00

The overlooked insight: the bottleneck constraining humanoid robot deployment in real-world multi-person environments is not locomotion or manipulation — it is voice AI speaker disambiguation. Companies solving this specific problem (multi-speaker attribution in real-time, noisy, dynamic environments) are critical picks-and-shovels infrastructure for the entire humanoid robotics wave. AssemblyAI is already embedded here, and any company that solves speaker disambiguation at scale will likely be acquired by or become indispensable to the major humanoid robot platforms (Figure, 1X, Apptronik, etc.).