You've Been a Bad Agent

Wilhelm Klopp (00:00)
What's up? I don't think we've ever started a pod on time like this. I think this means Luke is clearly a very special guest.

Matthew Carey (00:01)
Yo, okay.

Luke (00:01)
I'm excited

for this.

Matthew Carey (00:07)
Basically, I'm always 10 minutes late. No, it means that I am on off-site and I've actually been like this is the only thing I have externally to do for like the whole week So it's like I've been prepping for at least the past 10 minutes. I'm also in a room I don't like I'm like this is my bedroom. I Know you can't probably can't sleep. yeah, it's absolutely tiny

Wilhelm Klopp (00:20)
Hahaha.

yeah, usually

you're in someone else's office.

Matthew Carey (00:28)
Yeah, well, it's really funny. The only ⁓ light in this bedroom comes from a skylight, which is directly above my head and it's got no blind.

It's just utterly pointless. Luke, dude, thank you so much for coming on.

Luke (00:43)
I'm excited to be here. Are we recording? we live?

Matthew Carey (00:46)
We're always live actually.

Luke (00:47)
We're always live, perfect. Everything with Matt is on the record.

Wilhelm Klopp (00:48)
Wait a minute.

Matthew Carey (00:52)
No,

I hope not. No, we just normally cut the first bit off. So if we say anything silly in the first few minutes while we're warming up, we cut it off.

Wilhelm Klopp (00:59)
Matt, no, no, we're

pure, we're raw, we're... Matt, I feel like we need to play the intro, especially because Luke is here.

Matthew Carey (01:06)
Okay, have you got it ready?

Wilhelm Klopp (01:08)
I don't have to read it at all. Wait, last time you put it into your microphone. Do you have a microphone? I mean, usually you add the intro obviously after the fact, but ⁓ okay.

Matthew Carey (01:10)
man, okay. Yeah we did. Dude, we're so prepped. Okay.

Yeah, I'm

scrolling back between for all our messages and I will find it. will find it. Maybe we can play Luke a different one, to be fair. Do you think he'd like to hear variety?

Wilhelm Klopp (01:29)
Hmm.

Luke (01:29)
I'm

hoping as long as these were made with ElevenLabs music, otherwise I do not want to hear them. Really? That's awesome!

Matthew Carey (01:33)
They actually were. They actually were. Yes.

Wilhelm Klopp (01:38)
I was about to say they're made with

a competitor but I can't even think of competitors. I guess that speaks to you.

Matthew Carey (01:43)
No, I don't think so. Wait, I'm gonna play this one.

Right, are you ready, everyone?

yeah okay three

I know, sound's loud. ⁓

Luke (02:03)
It's actually, yeah, it's good. Did you try out the jingle maker we made as well? Where you can take any website. cool.

Wilhelm Klopp (02:08)
That's how I started actually. Yeah, yeah, that got

me hooked. I believe ⁓ they call this a lead magnet in the biz. Very well done. No, that got me hooked and that got me solid on it. And then I pretty quickly graduated to the full blown editing experience.

Luke (02:14)
Hahaha!

Yes.

Wilhelm Klopp (02:22)
And this was a few weeks ago, so I don't remember exactly. But yeah, mean, definitely it's very good at making jingles. I think I struggled to it really wants to make like a minimum thing of like 10 seconds or whatever. I think I was trying to get it below that, like three seconds, and it wasn't happy with that. And then sometimes it just gets it just gets like the everything I'm looking for perfectly right. Like I think maybe like one and three or something. But then I think I struggled to make edits to that.

Luke (02:35)
yeah.

Wilhelm Klopp (02:46)
version. So like it felt a bit like, I don't know, like a crafting, like a fickle masterpiece that was kind of slipping through my fingers. But yeah, it was, mean, it sounds great. think that was like the last two thirds of it. think the yeah, I think we should add it to the published episodes as well. I'm not sure if we have done it actually.

Matthew Carey (03:04)
Yeah, we probably,

I haven't actually worked out how to do that on Riverside yet. We are podcast newbies in case you couldn't tell. ⁓ But yeah, maybe we can give Luke a little intro. I mean, you've done all the intros, so I'll do this one. How do we intro you, Head of Growth at ElevenLabs, ElevenLabs, if you haven't heard of, is an audio platform, for voice AI platform, how do you call it?

Wilhelm Klopp (03:11)
Ha ha ha.

Matthew Carey (03:27)
How do you introduce 11 labs? Conversational AI, conversational agents, research lab, all of the above.

Luke (03:27)
And yeah, I describe it.

All of the above. I describe it as we build the best AI voice models and then we build products on top of it. Namely the agents platform and it's most relevant for developers. But then we also have this creator platform so you can create podcasts or audio books or videos.

Matthew Carey (03:50)
I love how you everything when you describe everything from 11 labs, it's always the best of something. It's like we do the best audio models. It's amazing. Yeah. No, I just, it comes across in your, in your address. It's very funny. I love it. Anyway. Yeah. So

Luke (03:56)
It's always the best.

Yeah, and normally

with a benchmark. Normally with some evidence.

Matthew Carey (04:06)
Yes.

That's perfect. Amazing. previously you worked at Post-Hog ⁓ and you worked at Product at Post-Hog. I think you're like, actually maybe you could start far back because otherwise I'm going to spill the beans. Maybe you could give a little bit of like, think of your backstory, I mean...

Wilhelm Klopp (04:10)
into the same

Luke (04:12)
Yep.

Matthew Carey (04:23)
Yeah, it's probably been a pretty crazy journey,

Luke (04:25)
100%. And so I, coding since I was 12 or 13, started with building a ton of calculators and card solvers in Visual Basic. So like advanced calculators. And then had a brief detour through medical school. So originally thought I was going to be a doctor. Then spent all my time though building different projects with friends and decided no, not going to be a doctor, going to try tech.

worked at Microsoft Research on low-level RL, so was doing PyTorch. First, developing agents to play video games for testing Xbox games, and then secondly for testing Windows, agents to explore through Windows. Then left that to try and found a company, started with trying to build a browser-based AI agent.

called IOS or AIOS. And we go into YC, we realized we're way too early in this AI wave to make that feasible. So we pipped it a few times. That turned into a telehealth company called Fela Health that's treating now thousands of patients per month. then I did, yeah, worked at POTOG interim header product there, really cool suite of tools for developers. And they've got this really cool vision now of trying to build autonomous products.

taking in all the feedback, session recordings, analytics, and like fully closing the loop on product development. then, yeah, and then joined ElevenLabs now 22 months ago, and it's been, we mainly think of multiples of 11. And yeah, been a lot of fun since.

Wilhelm Klopp (05:56)
That's wild, wow.

Matthew Carey (06:08)
That's so funny. Yeah, I love how you described a medicine degree as like a brief detour. A lot of people would consider that a life's work. But I think that's kind of crazy. Yeah. Mine also, my parents I guess I met you through AI tinkerers maybe or maybe before that actually. But I also went to...

Luke (06:17)
That was my parents' hope. That was my parents' hope. Yeah, sadly not.

Mmm.

I tried to hire you

Matthew Carey (06:28)
Yeah, I went to

Luke (06:28)
a few times.

Matthew Carey (06:29)
Crazy, crazy, crazy, Cool.

Luke (06:31)
Small world.

Matthew Carey (06:32)
well now you've opened the gate on hiring. I was very interested You guys are obviously very, very good at growth. I think a lot of people have heard of ElevenLabs. Probably a lot of people use it. I mean, we used it for the little jingle and that was the first thing that came to mind. was like, ⁓ ElevenLabs have a product. We should just use it. So you obviously did your job very well there on us. But how do you think about hiring for growth and what you're looking for in people there?

Luke (06:55)
Yeah, I speak to families about this. Oh yeah. Yeah, yeah, I'm married. I'm married. Thank you. Yeah, I had my one year anniversary. So over 11 months in now. Last weekend. So that was a lot of fun. But yeah.

Wilhelm Klopp (06:58)
Did you get married? I just saw a wedding. ⁓ congrats! I don't think when I last saw you, were married. That's amazing!

Hahaha!

Matthew Carey (07:16)
Oh, I can imagine the LinkedIn post. I'm

11 months into my marriage. This is what it taught me about working at 11 labs. Sorry, sorry. I'm just brain rotted.

Luke (07:23)
100%. So for

hiring, I think it really depends on the company. However, in general, I think there's two broad archetypes for growth. One is the like generalist growth marketer, particularly, I imagine most of the audience is like technical product companies. My hack for this is you want either ex founders or PMs because you need someone who's technical enough to really understand your product and translate it into

something customers can get. And then the other one is like the growth hacker. And so this is someone who's like a strong full stack engineer, but it's just like very results and numbers driven. And they're often be doing like the programmatic SEO work, the large scale, the outbound marketing. And so those are the two types. And then ElevenLabs now, you know, we started the growth team is like pretty small.

And the three of us, we're now at, I think there's 30 people in the growth team. So it's like definitely hired and we've really tried to layer in this mixture of like the earlier career, super driven, high potential, often like ex-founders or doing their own projects, but combine that with like true best in class subject matter experts. So we have like one of the SEO leads from Canva. We have the previous head of performance marketing for Shopify.

Wilhelm Klopp (08:22)
Wow.

Luke (08:44)
So then these kind of like layer in really nicely, but at the start, particularly if you're doing like developer technical tools, like have an incredible, like high potential early career growth marketer who just like obsessed with your product, execute kind of the launches, I often think is the heartbeat because you're basically taking the incredible work that the team is doing and just translating that into the world. And if you're building like net new incredible products, that should be enough.

And then you start layering on these other channels like SEO and paid. But yeah, know Will, Will's also very good at growth with a simple poll with the first polling app in the Slack app store. So there's probably some like distribution hacks I can learn from.

Wilhelm Klopp (09:25)
Oh man, here's where it's wild.

It was actually not even the first. Maybe it wasn't even the second. Yeah, I think it was just really strong execution. well, yeah, a lot of growth hacks. I gave a talk actually in 2017 at the Slack office about growth and my main competitor was in attendance as well, which I didn't notice until later. But anyway, we can talk about that another time. I actually was gonna ask you though about...

Luke (09:30)
Really?

Hahaha!

Matthew Carey (09:36)
Ha

Luke (09:44)
Interesting.

Wilhelm Klopp (09:50)
growth because it's an interesting subject matter. When I gave that talk actually, I thought growth as a function meant like, yeah, you just grow. You know, you don't have growth, you hire a growth person and then you have growth. But it works differently, right? Like, could you explain for those who don't know what the point of a growth team actually is? Because obviously in some way everyone's objective at the company is growth. And then do you think...

Luke (09:54)
Yeah.

Yes.

Wilhelm Klopp (10:14)
the ElevenLabs growth team that you lead is similar to how the industry defines it usually, or if it's like very different.

Luke (10:20)
Yeah, I think...

Traditionally, you would maybe have marketing. so marketing was people who just did campaigns and probably non-technical. And so that's why I think lots of tech companies now call it growth, where you combine and you have, you know, maybe backgrounds more similar to me leading it, where you have the engineering side too. And you combine both the engineering side and the marketing side. And then, I'm sorry, what was the other bit of your question was?

Wilhelm Klopp (10:48)
I guess like if, yeah, I mean, maybe I'm wrong and everyone actually does growth differently. Yeah, yeah, sorry, go on, yeah.

Luke (10:51)
yeah, what, what we even do. Yeah.

Yeah. So the way that we view it is the growth teams responsible for all the acquisition, activation, monetization of the self-serve user base. And then we drive leads for the enterprise side. And so the different orgs within the ElevenLabs we have the growth org, we have the product and engineering org, and we have the research org and the sales org. And so.

Wilhelm Klopp (11:20)
Gotcha. Yep.

Luke (11:21)
If you think

what we spend most of our time doing, it's like big launches, it's SEO, it's paid ads. So that's like the top of funnel work. It's partnerships, hackathons, events, but then it translates into activation. And so this is when we have growth engineers either embedded in teams or we have a core growth engineering team who will search. And then it's like, okay, how do we actually translate that to the like the wow moment, the aha moment, the habit loops?

And then it translates later on, it's like, okay, core team is focused on just like the best overall product and retention is really their metric. But then we're then come in to be like, okay, what are these like upsell moments or converting to monetization or upsell actually to contact sales.

Wilhelm Klopp (12:05)
Nice. Yeah, that answers my question really, really well, actually. I always find it very helpful actually to know kind of roughly the org chart or like what the different departments are, right? Your job would probably be very different if there was a separate big marketing org, right? Then like maybe the growth stuff would be much more like micro improvements or like very like numbers driven tweaks around the funnel. But it sounds like you own the funnel as well, which I think it kind of gets at that question a little bit of how I think a lot of other...

Luke (12:12)
Yeah.

Mmm.

Wilhelm Klopp (12:34)
companies define growth teams? Like it's very, like the funnel is like what you live by. The other interesting thing I was kind of aiming at with that question was the kind of influence you have on product because often like the most impactful growth work, right, ends up being quite involved with product. Yeah, how does that work for you?

Luke (12:51)
Yeah, and one mistake we made early was we tried to spin out a dedicated growth engineering team too early before we had like a critical mass of product engineers. And so we then had...

At one point is like three growth engineers and I think maybe like five core product engineers. And the growth engineers will focus on like micro optimization. And we very quickly realized like, okay, this is clearly the wrong split. Like the micro optimization can be huge leverage at key time. Like your pricing page, every pricing page test we do, like we pay back the engineer's salary with one test and they normally work. But.

Wilhelm Klopp (13:29)
You

Luke (13:30)
It's like how much of your time do you actually want to spend on the pricing page versus increase the core product? So the way that we've mainly done it is you have the core product teams, and we're very engineering led on the product side.

Wilhelm Klopp (13:36)
Mm-hmm.

Luke (13:47)
all the product teams, the teams of engineers, they work with designers and growth plays a little bit of the PM role of like surfacing the metrics and the insights. And we do these growth reviews with the product teams. But really it's them who own the product and the roadmap. And then they're like receptive to the different insights. But I think it works so much better when like the engineers, they care about the product quality, they care about the metrics. They're excited by the

Wilhelm Klopp (13:47)
Great.

Luke (14:17)
that they're driving.

Wilhelm Klopp (14:18)
That's awesome. Really, really interesting.

Matthew Carey (14:21)
I've never heard a description like that, so that's actually great. Dude, you're shooting all the bangers at us. I was actually gonna ask a little bit about how you guys work. You talked a little about the structure of ElevenLabs. But I heard recently that some parts of it are still, like some people still work remote, but I feel like you work in an office in Soho, so maybe you could, yeah.

Wilhelm Klopp (14:27)
Hahaha

Luke (14:41)
Yeah.

Matthew Carey (14:45)
How do you find the differences? Do you mandate a certain day a week in the office? And how do the different teams work there?

Luke (14:51)
Yeah.

Yeah, we're fully remote by default. And then we have offices for people who like offices. So I like an office, so I come in every day, but also lots of the team and the majority of particularly the engineering and research org are fully remote. And in terms of how we work, it's definitely like shifted a bit over time. And the analogy we use is like the research push and the product pull.

And so the research team, they like spend all their time trying to really build these like, what are these best in class models, which really move the needle. And there's a really cool flow whereby new research capability unlocks new product potential, which gives you the ability for like viral moments and marketing. that's why like launches, particularly in the AI world where like traditional B2B sass

you know, actually what you're launching maybe isn't that exciting. But in the AI world, it's like, holy shit, we've just launched a new novel capability, which just was not even possible before. like transcription is now possible in Serbian and the market size of every app which is based on transcription has just increased. Like we just need to tell the world about that. And how it used to work was research would...

Matthew Carey (15:59)
Mm-hmm.

Luke (16:10)
create a model, they would then ping me or one of the folks on growth or product going, hey, we've just shipped the world's best transcription model. Let's launch it next week. Go do the growth stuff. And then it's like mad scramble and push to like do all the work we can in that week to then like launch. And that just added so much momentum and the particular...

thing again, which is unique about the AI space is we talked about you're pushing these research frontiers. And actually, if you don't ship quick enough, someone else may come in and push this frontier before you or overtake you on the benchmark. So you actually need to move super quickly. And so lot of our like, why we work so hard and push so hard is so we can like own those viral moments or those frontier moments. They're like,

Wilhelm Klopp (16:58)
Mm-hmm.

Luke (17:01)
thing now as we've, and I'd say probably the first like 12 months really of the growth and product team was just like trying to do our best to get through these checklists before the research team gives us another best in class moment. And we've now like managed to get the team big enough that we are also able to do like the customer pool where we're able to look like, what's the customer workflow? What's the jobs to be done? Where's like...

Matthew Carey (17:22)
Hmm.

Luke (17:28)
our product gaps and then also do the normal stuff. But then quite often, as we saw with 11labs music, we're like, okay, now just drop everything and launch this and make this truly best in class. So it's really fun to still have those moments.

Matthew Carey (17:43)
Really cool.

Wilhelm Klopp (17:44)
Fascinating. This

is so great. I feel like we can take this in so many directions. I wonder if maybe it would be useful if, and I'm curious about this as well, but I think also might be useful for everyone listening. Could you give like a bit of an overview of the different use cases actually, or how you think about like the different...

I don't know what we even call this. It feels like things are moving so quickly, like different types of customers, but yeah, different types of use case. I'd be interested like across the spectrum, like the most common, the most boring, the most weird, like, yeah.

Luke (18:04)
Yeah.

Yes.

Matthew Carey (18:11)
The most token heavy is the one I want to know about.

Luke (18:12)
So.

So ElevenLabs has two main platforms and both of which are available over API. So we have our creative platform, which is the best models for media and entertainment. So if you're making music for this podcast, if you're making voiceovers for videos, so voiceover for videos is actually a massive one. If you're making audio books, we have like a whole editor as well to make audio books. So that's the creative platform.

And you can either use any of our UI parts, but also if you're a developer like Hey Jam or Synthesia or Captions, then you can take these APIs and you can build those into your own platform. The other pieces are agents platform, which was previously called conversational AI, but we've renamed it to agents, which enables you to build full voice agents or chat based agents. And the backstory for that is,

Loads of customers were taking our text to speech to text, wrapping them together, building around turn taking models, literally months trying to optimize all the latency, chucking an LLM in the middle and like, boom, you've got an agent you can chat with and it's a truly amazing experience. And we realized quite quickly, it's like, okay, after the 20th enterprise customer we're working with hands on to try and build this, it's like, let's build a platform to make that fantastic. And so...

that's used to power lots of the biggest players in the space. like, you know, Decagon or, I'm trying to remember which are public and not public, but pretty much everyone who's using voice like sits on top of this platform. And then increasingly as well, we've started going like, okay, let's both sell to developers, but there's lots of these big companies or small companies which are wanting to build their own customer support agent.

Wilhelm Klopp (19:48)
Yeah.

Luke (20:06)
And so we've also enabled you to do that. And there's a really nice onboarding flow where if you click new agent, you literally describe what you want your agent to be in and put your, so you go like customer support and then you add the link for like SimplePoll and you click create and boom, you have a voice agent you can speak with. You can even like clone your own voice. You can upload your full knowledge base to do rag and maybe SimplePoll already has a full customer service team.

Wilhelm Klopp (20:22)
Mm-hmm.

Luke (20:33)
but you can now also do it in Hindi as well as in French and you can drop this widget into your site or connect it with Twilio. So that's like the big bit. And if you're interested, there's like a detail here about the research side, but where we are really betting a ton of the company is on this agent's platform. And can we be the place where anyone who's building any of these customer facing agents uses 11 app?

Wilhelm Klopp (20:59)
Okay, really interesting. just to dig into one thing that is the main use case for the agent platforms of some versions of the customer support agents. That is like the main thing,

Luke (21:09)
Yeah, yeah, customer, so

it's either inbound, so it's like customer support, receptionist, customer service, or it's outbound, which is normally like sales and reminders. That's very common in healthcare.

Wilhelm Klopp (21:23)
That's, so yeah, I really wonder if I've spoken, who of the customer support phone calls I've done in the past few weeks have been with agents.

Matthew Carey (21:31)
Sunil said it was... Which hotel was it in London that was definitely way too posh for me to have been? it the Ivy? It was the Ivy! It was the Ivy in London. I'm pretty sure they use Eleven Labs. Well, they use some sort of voice agent.

Wilhelm Klopp (21:40)
yeah, yeah, yeah.

Luke (21:41)
Yes, yeah, yeah. Yeah, we've got,

I think it's, let me double check the name. So we've got seven rooms is a big customer and they power lots of these restaurants. So it's probably them.

Wilhelm Klopp (21:47)
want to give them a call see what it's like.

Matthew Carey (21:49)
Yeah.

Ah, okay, yeah, yeah, okay, that makes way

Wilhelm Klopp (21:56)
I see. ⁓

Matthew Carey (21:58)
more sense. Okay, okay. I was just imagining someone in the Ivy implementing like your 11 last API. Yeah, no, it's really cool.

Luke (22:00)
Yeah, and there's...

But they are, they are as well. one

Wilhelm Klopp (22:06)
Hehehehe... Hehehehe...

Luke (22:08)
of the most common questions I get is, are the, like, Luke, I want to build a company based on voice, what should I build? And I think the common mistake today in 2025 is to overthink this and try and do something like, actually, I was speaking with one company which is making several hundred thousand dollars a month.

Matthew Carey (22:17)
Yeah.

Wilhelm Klopp (22:27)
Mm.

Luke (22:31)
and they do door-to-door sales via a motorcycle, selling hotel receptionist voice agents in Guatemala. And like, just the market size for voice agents is so big that like, you can literally take any intersection of niche industry plus geography and like carve out a pretty fantastic business.

Wilhelm Klopp (22:41)
Wow.

Matthew Carey (22:44)
Yeah, it's massive.

Wilhelm Klopp (22:51)
Mm-hmm.

Matthew Carey (22:54)
That's

so cool.

Luke (22:54)
And

it's like probably most of us made our first thousand dollars setting up WordPress websites for local business. I think there's an equivalent now for building voice agents.

Wilhelm Klopp (23:06)
renaissance of that, that's really fascinating.

Matthew Carey (23:07)
Okay.

Now, I was just thinking that all of this relies on like a very strong like product-like growth, I guess, outlook. How do you think about that? Where do you think that should go in a company's life cycle? I know you've chatted to a lot of startups. So for a startup like thinking about starting some PLG motion, yeah, like where would they go with that?

Luke (23:26)
Yeah.

So interestingly, well, to answer the more broad question of like how I think about PLG, there's an interesting thing where like, think PLGs became such the default that I think actually the big mistake is not layering on sales led on top and so all sales assist.

And so, and I think most companies actually get too far without layering on a sales team. to give you, Will, did you, with SimplePoll, did you ever layer on a sales team?

Wilhelm Klopp (23:58)
We dabbled with sales in various ways. We could never quite make an outbound motion work though. But we had a little success with inbound over the years for sure, yeah.

Luke (24:05)
Yeah, yeah, yeah.

Matthew Carey (24:08)
SimplePoll

Luke (24:08)
Yeah, nice.

Matthew Carey (24:08)
is mad. looked, I was moving all my Slack things over. Sorry, that's just tiny detail. SimplePoll is so mad. was moving all my Slack accounts over to a personal email for a thing that we're gonna talk about maybe some other time. And I literally just did forward slash poll debug as I do to see the team ID. So the little hack for you there. So I can go to it on the web app.

Wilhelm Klopp (24:12)
haha

Ha

I'm glad we've,

you've, yeah. We found a new.

Matthew Carey (24:31)
just share that to everyone for

you. And on every single Slack community I was in had SimplePoll, just to put that in perspective. And there was, was like, I was like six of them. Rogue. Anyway.

Luke (24:34)
Hahaha!

Wow.

Wilhelm Klopp (24:41)
Yeah.

Luke (24:43)
Wow.

But the point, and you're probably wondering, so I think PLG is fantastic for like super efficient customer acquisition because basically any of your channels, whether it's like SEO paid or launches, the funnel drop off, if you try and put a person, maybe even a voice agent in the middle of this like convert to sales and then you put a contract that you sign.

versus like you just reduce all that friction, you can get way more customers. If you have a great product, you can get them to start paying. But why I think the Sales Assist is so useful is you're forced to do the like, okay, what objections do you have? ⁓ which you're comparing us versus someone else. Okay, right, what's our win-loss ratio versus the other platforms out there? Why are they choosing them? What features do they care about? And so I actually just think for like product development alone and you know, I've been at...

a few companies, PostHog was really interesting where they went just PLG in the start, there's change now. But Mati and ElevenLabs, right from the start, Mati's constantly had this thing of like, we want to sell and get to the enterprise. And really most industries, enterprises, there's far fewer of them and it's much more of a monopoly. And so you can build a much bigger, more enduring business. And so he's constantly had this take of like, okay, how do we become?

a solution that enterprises really trust right from the start. And he's constantly, the way he's done that is like pushing up the contract size that he's selling to. And then you bring back, okay, what are the product learnings? What's then the playbook for the sales team and the growth side?

Matthew Carey (26:19)
Could you share some learnings, just some fun ones that you've had in the last couple of years?

Luke (26:23)
one big one is what actually is on-prem. it's like a meme. Like, okay, we want to sell to enterprise. What do enterprise want? They want SSO and on-prem. And then you're like, it actually is on-prem. And one learning is on-prem is actually a spectrum between like, you know, maybe you're a legacy bank.

who has their own physical data center where they literally want you to deploy the thing on their own servers that they own and maintain. But it's much more of a spectrum than that. And very few people, I think apart from defense or maybe certain hedge funds or highly regulated industries, really need that. And the step back, which is they actually want to deploy it on their own cloud, is way more manageable.

But then the question is like, what does actually deploying on their own cloud mean? Do we need to deploy the entire ElevenLabs infrastructure onto their cloud? Or one thing we've done is shipped AWS Bedrock, which I'm not sure how, how familiar you are with it, but it's really cool. It's a stateless container. Stateless being the important bit is like actually quite limiting in some ways, but for AI applications, you can take our text to speech models.

You can run it stateless on a customer's account. Because it's stateless, it's also way easier to like scale up, scale down, maintenance costs are much lower. And so that was one we've done. And then we've also quite early on, because Matti and lots of the early team are from Palantir, had a forward deployed engineering or field engineering team. And so they're the ones who work out, right, what are the key blockers to the next big deal?

Wilhelm Klopp (27:51)
Interesting.

Luke (28:05)
Okay, let's ship Bedrock. And then that's what's enabled us to do stuff like Epic Games, having Darth Vader running around speaking to people powered by our agents platform. So that, yeah, that's been a really cool one.

Matthew Carey (28:18)
That's wild. How was it negotiating with AWS on like actually getting something into Bedrock? is that something that you were involved in? how? Yeah, I don't know where you'd start there. Guys, we have a really good model. Put it in please. Or is it more than

Wilhelm Klopp (28:18)
That was wild. Yeah, yeah.

You

Luke (28:32)
their teams, they want you on Bedrock. And what's interesting is when you think, like AWS, so by default, we're more on GCP. Peter's from Google, we know GCP. By default, we're more on GCP.

The way that these cloud providers actually get you onto AWS, they're like, ElevenLabs, come ship your stuff onto AWS. My intuition was always, ⁓ they would pitch you on cost reduction and cost savings. And ElevenLabs by default would be like, no, we're not interested because growth is going really well. We'll do the cost optimization in the future. But the way they actually get you onto the going with multi-cloud is they pitch you saying, okay, we have loads of customers that we can co-sell to.

Why don't you create a Bedrock instance and then we go push and co-sell with? So actually it's like these other platforms are wanting you to move because they, you know, it's fantastic business for them if you're spending more compute. And they picture you as like you can gain more revenue by creating Bedrock. Most of the work is on you to like create a container which you deploy, which is stateless and they're like carved out of your existing input.

Wilhelm Klopp (29:34)
Mm-hmm.

The only thing I'd ever heard about Bedrock before, your glowing review here, is that it's like a bit of a weird, or requires like a weird inference stack or something like that. And that's why so many people don't like it or don't use it. Like, do you know much about this? Like, I have no idea what I'm talking about here, but like, is it, okay.

Matthew Carey (29:39)
interesting.

Luke (29:48)
Mmm.

Yeah, I didn't do the build, but

we had, I think it was...

two engineers in about two weeks. I think that's mainly credit to how good Lawrence and Tony are for reference. Tony as well, he's ex-Palantir and he was known, there's quite a lot of folklore around him. And he was apparently the guy at Palantir where like, if ever there was a problem which people couldn't solve, they would like at Tony and Tony would jump into like,

fixed the info of it. But yeah, the cool thing with, if you like take FDE seriously and you take seriously the idea of we want to get bigger and bigger enterprise contracts is you're constantly going, what's that next blocker? right, we need to ship a on-prem. How do we do that? Let's try this bedrock thing. Does that work for this customer? Yes, it does. Fantastic. Now we do it.

Wilhelm Klopp (30:23)
incredible.

Luke (30:46)
And then very quickly that unlocks a whole bunch of extra customers and motions as well.

Wilhelm Klopp (30:50)
Yeah, speaking of this actually being thoughtful about the enterprise, it brings up one question I wrote down beforehand, which something that stood out to me about the ElevenLabs music launch is that you actually, and again, this feels like the kind of topic where silence abounds and no one can speak, but like, actually went and spoke to a bunch of the like music labels and like the actual like owners of the rights.

ahead of time and kind of involve them, which I think is very different to like a Suno or something like that, which is pure product that go through. I mean, I assume from the outside, that's what it seems like. And that seems like really exciting because obviously all the music stuff like...

Luke (31:20)
Yeah.

Wilhelm Klopp (31:30)
sampling from other people or generating new music that sounds like Taylor Swift or whatever shape it takes. It feels like taking an approach where you can give a piece of this new pie to the old players will just like result in more success. So if there is anything you can share about that, I'd be really interested just like ⁓ what's happening there or like what we can, yeah.

Luke (31:47)
Yeah, the music stuff was

wild. And I don't think we got the comms quite right on how impressive this was, but this was a multi-year deal where...

for like multiple years, we've been chatting with the labels, trying to get the rights for data to train on and in a way where all the artists can benefit as well. And yeah, like it came together. And then again, we had this moment where like deal signs, models training, and then like release. And then it's like, okay, how quickly can we do all the other bits? So that was really cool.

Wilhelm Klopp (32:18)
Mm-hmm.

Luke (32:25)
interesting

thing there is...

it really wasn't obvious that this would be possible. And the two massive investments, one was like, we hired the world's best music AI researcher for this. Separately, we had two folks who were hired to just try and get these partnership deals done and across the line. yeah, the folks did a great job there. And then it's like, okay, now it comes together. So it's like a multi-year bet.

which to me was really not obvious. also the, yeah, it's like pretty cool with ElevenLabs how horizontal it is, but like we started with voice, with text to speech, and you end up, when I talk about this like research push customer pool, like.

Lots of them at the time were not obvious. It's like, hold on, why are we doing a transcription model? Like, why are we doing music? Like, everyone's using us for voiceovers for videos, like, or audio books. Audio books don't have music. Why are we doing music? But it's just also the, curiosity by the research team and Peter of, like, can we build the world's best model? Can you create music? And then, like, there's been a really cool moment recently where, like,

Wilhelm Klopp (33:12)
Mm-hmm.

Mm-hmm.

You

Mm-hmm.

Luke (33:39)
Before we had this split of agents platform and creator platform, it was like all quite murky. It's like, we've got voice agents next to music creation. And like it wasn't, but now it's like kind of finally distilled into, wow, we've got these two platforms where like anyone who wants to build voice agents or anyone who wants to create content. And like, let's actually think even bigger within those platforms. Like what happens if we expand into other modalities or really take like voice

seriously and ship you know we Swift SDK and Kotlin SDK and all the other native SDKs for the voice agents.

Wilhelm Klopp (34:17)
I would love to jam

with my customer support agent, creating some music together while we wait for the human who might need to come online to unlock the next step. No, no, that would be cool.

Luke (34:22)
Yeah, well, you joke about that. One of the biggest,

actually, like lots of people are using, because there's still like waiting moments for transfers. So it's like...

Wilhelm Klopp (34:36)
Uh-huh.

Luke (34:37)
Customers do actually want the music side, don't they? So it does like that, or like the world's best transcription models, fantastic for video captioning, but it's also fantastic for these like agent interactions and cool. So you do see these like cool, weird combinatorial effects as well.

Wilhelm Klopp (34:41)
That's awesome.

So this is another great segue because I was actually kind of waiting for you or hoping you would say the H word, the horizontal word, because one of the most memorable conversations I've had with you was, think it must have been over two years ago, was about, was you describing to me that one thing you love about post-hoc is that it takes this like very different approach to integrating all these different product lines, right? So you have the product analytics, also the web analytics, but also the like session recording and now error tracking and probably like five other things.

Luke (35:13)
Yes.

Yeah.

Wilhelm Klopp (35:23)
A-B tests and then because you do all of them

It actually unlocks new features that you can't really do if you don't have the full stack. And that's the perspective I'd never heard before and that I found like really interesting and really compelling. Even though I think I'd grown up in kind of a world where in like kind of a, mean, when I was at GitHub, was very GitHub versus GitLab. GitHub was very like best of breed. We do this one thing well and GitLab was like, we'll sell you like 30 different products that are all maybe a little bit crappy. So I was kind of like in a different camp, but I think you kind of opened my eyes to the...

like how good this horizontal thing could be. And it sounds like there's some similar stuff or the kind of seedlings of some similar stuff at ElevenLabs.

Luke (36:00)
Yes,

yeah, 100%. So the bit, they're like...

And this, the broader term here is compound startup. And I think Ripling really brought it back into like the modern mindset for startup building. And the dream is you have like one ICP, you go really broad for them. So, POSTHOG is product engineers and each new product you ship gives you another entry to get the customer, but also to make their workflow better. So, POSTHOG started with product analytics. wait, if you can also see

how your customers are actually using the product. That's a useful feature. People actually integrate that earlier. So it's like, okay, we can actually hop ahead of the mix panel and amplitude in the customer journey. And then because they use you for session recording, they've already turned on the product and let it. But then you're like, let me create a funnel. Who are the people that dropped off? let me watch their session recording. So it's like really nicely tied together. And my frustration with 11,

Wilhelm Klopp (36:57)
Mm-hmm.

Luke (37:01)
I think we're fantastic company, but has been like until...

you know, recently, because we were so research led, it was like, wow, we're shipping this new product. And then it'd be like, but this is actually most relevant for a whole different set of users. But, and it like, how does that really tie back into the further? And they didn't necessarily feel like we had that compounding motion. And, I think now we've like solidified this agents platform and it's, we're building for, you know, developers who are building voice agents.

and their companies, it's like, now it really has got to like, okay, if you now shipping testing, then everyone else benefits or yeah.

Wilhelm Klopp (37:43)
Yeah, and if you have anything to do with it, I'm sure a lot more to come from the compound perspective. Okay, exciting. I want to go back very briefly to the use cases side of things, because I think it's just fascinating. And I love with these new ⁓ waves of technology, it's, I think...

An obvious thing is like, replace this old shitty workflow. But I think what, as in like, you don't want to wait and hold for 30 minutes just so that someone can type things you say over the phone into like a form. Like not a great experience. But the kind of really exciting stuff is also like all of the like completely new like use case, like a completely new stuff that comes out. So I'd be interested to hear about any of that.

But one thing that also that comes to mind, and again, this is kind of from the Slack days and like the early days of the Slack platform. One of the BD people at Slack wrote this incredible blog post, which I still think about, like, I don't know, at least once every few months. And he talks about kind of like the 10 year vision of what life looks like for like a normal knowledge worker if the Slack platform succeeds. And it's like.

You wake up in the morning, you have a message from one of your bots. tells you about like what you have in the day. You can click a button. It schedules like a thing. You get another message. It approves an expense. And it just describes like what a day in the life looks like with you interacting with these like 2030 apps and you never have to leave Slack. And none of that had been built yet. I think this was like published in 2013 or 2014. So not just didn't the apps or the integrations not exist, but like the platform functionality didn't even exist yet. Like it was that early, but it painted a...

Luke (39:00)
Yes.

Wilhelm Klopp (39:16)
a really nice vision. think it rallied kind of developers towards that. well, I mean, I guess one idea, unless this exists already, maybe someone at ElevenLabs could do some, I don't know what the voice equivalent of a blog post is that describes, you've done this. ⁓ sick.

Luke (39:28)
I've actually done that. Let me see if I can find that.

Okay, shall I read it out? It's only short. It's only short. It's 2027. You start your day with a two hour walk through a reclaimed forest. Overnight, your AI assistant has been autonomously fixing issues, analyzing customer updates, phoning customers for user interviews. The sky is blue and you're looking up.

Wilhelm Klopp (39:34)
Please, please.

Luke (39:51)
As you walk, your smart glasses speak products and growth ideas to you. New feature to ship, partnerships to execute on, and marketing campaigns to run. You set the vision, you're the curator. Some ideas you accept, others you give feedback on, some you reject. The AI agents then execute in the background, keeping you in the loop on big decisions and creating variants when needed. It hits 12.30 p.m. and you've already shipped 200,000 lines of code and spun up 18 new marketing

experiments. You sit down by the river and a drone drops a sandwich into your hand. So that was, I was actually walking through a park and I was like how sick would it be if we were now just wearing like AR glasses.

Wilhelm Klopp (40:27)
So good.

Luke (40:35)
and you just have this super smart voice agent you can chat to. And this was like a take of a software engineer, like their use case. But I think we're close. I think we're super close. And the other cool thing I realized was I think actually for the personal assistant,

AI personal assistant. You really only need three connections. You need Gmail. You need Calendar.

and you need WhatsApp and that for most people plus a fantastic voice interface that for most people is enough to like fully understand all the context on their life like execute actions, book events but I think we're super close and this is a little bit of a detour but I actually think the main limitation at the moment is security and so like people we need to crack the like MCP security side and deal with this unhappy triad

and I think Simon Willison has spoken about it because otherwise it's just like, know, prompt injection is not a solved issue for anything and that's the reason no one's running autonomous agents in production.

Matthew Carey (41:30)
Yeah.

Wilhelm Klopp (41:33)
Yeah.

Matthew Carey (41:33)
Definitely.

S... So bad.

Wilhelm Klopp (41:42)
I forget what it's called, but have you heard of the reverse air pod? The thing, it's like this device you put over your mouth so that you can speak privately. forget what it's

Matthew Carey (41:50)
Yeah. I saw a tweet from I saw a tweet

Luke (41:50)
⁓ well, the...

Matthew Carey (41:52)
from you, Luke, about what an office looks like if we're all shouting. ⁓ And was it that BCI thing?

Luke (41:58)
yeah, yeah.

Well, there's a company which I helped a little bit with their launch called Alter Ego.

which they're like super smart founders from MIT and Harvard and have built a brain computer interface, through you have like mild muscular skeletal activity when you think. So like the same pathways, which would trigger your mouth to move, like trigger at a much smaller level when you think, when you sub vocalize. And I do actually think like that.

Wilhelm Klopp (42:15)
Mmm. Mmm.

Luke (42:29)
I could imagine a world where currently we have a keyboard at every desk and instead you have a like BCI device at every desk which you do like subvocalization and put to a computer.

Wilhelm Klopp (42:40)
Totally, yeah, yeah. think the BCI stuff has always been fascinating. Do you think this will work well enough at this point? I like that's always been the issue. It's like the...

Luke (42:47)
They claim

you're able to do well over 100 words per minute.

at 90 % plus accuracy. And the launch video is worth checking out because they've also combined it with, I think it's ElevenLabs voice cloning. And so you can then have telepathic conversations with people because I can think something and beam that to someone else and then I'll read it out loud with my voice clone, which is pretty cool. Alter ego.

Matthew Carey (42:55)
What?

Wilhelm Klopp (43:00)
Damn.

I need to check that, yeah, sorry, what was the name again?

Alter ego, yeah. ⁓

Matthew Carey (43:18)
just makes me think that things

like airpods are gonna be like so useful in the future.

Luke (43:23)
yeah, that's their form factory

is a like headset, AirPod style device.

Wilhelm Klopp (43:30)
This was the thing that I remembered. ⁓ I don't think you can actually... ⁓ So I actually looked at it a few weeks ago. ⁓ There was a bunch of these products. I don't think you can buy any of them anymore. I think they flopped, but I think they were too early. But I think if there was an ElevenLabs version of this, I think people would be rocking it in San Francisco ⁓ before the end of the year.

Matthew Carey (43:30)
wild.

Luke (43:33)
Hahaha!

Matthew Carey (43:36)
It's like a muzzle.

Luke (43:55)
No.

Matthew Carey (43:57)
That's horrific. That is actually awful. Did you see the other fun launch video that came out this week? The Interaction Company of California or something. I feel like loads of these people have that sort of name now. But yeah, Interaction and they released a personal assistant. Very similar to what you're talking about actually Luke. It has access to your emails, has access to your calendar and it has access to, and you communicate with it on WhatsApp and it has all on

Luke (44:17)
Yeah.

Matthew Carey (44:22)
iMessage and it has access to both of those things. ⁓ It is wild, yes, pokey. the onboarding is you bartering with it about how much you're going to pay for it. It's... Yeah, I think, I mean, we're giving them a lot of free press now. They should pay us for this. But like, the three people listening. But yeah, it was wild. Like you...

Wilhelm Klopp (44:26)
Mmm, the pokey thing.

Luke (44:26)
Very cool.

Yeah.

Yeah.

you

Matthew Carey (44:44)
It starts off by berating you and there's this whole story arc that you go through where it's like socially engineering you into paying more money. And then it gets to the point where it's like, right, it's this much money a month. And you're like, nah, come on, you can do a bit lower. And it's like, well, I'll give you a fiver off for good behavior. And then you go through and you go through and then there people on Twitter who've been trying to like get it as low as possible. But depending on your background, likes...

Wilhelm Klopp (45:03)
Hahaha ⁓

Luke (45:03)
Yeah.

Matthew Carey (45:09)
it like does this arc where it starts knowing way too much about you and you find it a little bit weird and then it bigs you up a little bit. No, then it destroys you and says there's no way we're letting you onto this. And then when you have to justify yourself, then it bigs you up and then it's like, right, so you can pay loads for this, you're minted. And then it gives you the price then it goes through the bartering phase. ⁓ But yeah, I just thought that was wild like a...

Luke (45:28)
Yeah.

Matthew Carey (45:32)
really random. could just imagine like people pre-programming that sort of interaction. Yeah, insane. As a voice as well, it'd be cool.

Luke (45:39)
My meta point is,

think there's, because Google meta Apple are so well positioned to do their AI assistant thing, plus I think founders nowadays are like conditioned to do B2B. It's like that's where sticky revenue, NRI, et cetera is. I think there's just been way under, like how come we've waited until now?

for a text-based assistant which can do actions on your phone. Meanwhile, we had, was it called Magic? I think which was like maybe five, 10 years ago, was like the same idea. And yeah, I'm so excited to see this like B2C push and yeah, use some of these products. I just wonder, the only issue is, pardon?

Matthew Carey (46:24)
We needed MCP, dude.

We needed MCP.

Luke (46:28)
And say, well, maybe, But like we had we've had talk cooling for a while.

Wilhelm Klopp (46:30)
Look.

We have, yeah. Luke, one thing that, so we're coming up on time, so there's not that much time left, but one of the big things I've been curious about also is, and I think I would try to ask this of everyone who comes on is like the kind of personal AI workflows that you use day to day. Because I think there's a, everyone can see there's so much like potential in the future. There's some cool use cases today. ⁓ But what, like how do you use AI day to day? And again, curious for like the stuff you've tried that doesn't work, the most boring stuff, the most surprising stuff.

Luke (46:51)
Yeah.

Yeah, I think coding is mainly cursor. I've tried claw code like Vibe, Kanban, it's a very cool workflow, which I actually think is the future of these like async background agents. But when I'm sync, largely like to be in cursor. On the other side, I have a number of different Zapiers. I still actually find Zapier a great platform for personal AI automation. One of my favorite ones is every email

which comes in. I have it label it with a large prompt into P0, P1, P2, P3. P0 means this is high priority and then it also forwards it to slack.

And then P1 is labeled it today. Today, this week, P2 is like less important than P3 I never see. And then I have this loop, which is I treat my Zapier as a personal assistant. And so I literally have a like one-on-one for 15 minutes each week where I go through and I update the prompt a little bit. And it's now fantastic where I like fully trust this labeling and use the superhuman split inbox. So it's like, I start with the high priority stuff and then

Wilhelm Klopp (47:43)
Hmm.

Mmm.

Luke (48:11)
go down. I use Grenoda a lot. We use it for our case studies. I actually did a How I AI chat with Clairvaux, which goes into this in some more detail around how we use it at work. For localization, we're all GPT based now. So we removed all our agencies. We just vibe coded a thing where it just takes all the text.

localizes it and we have a prompt per language which then the sales leads or the marketing leads who speak those language adjust and it's way better than like agency human translation.

One particularly high leverage one, which I'd encourage everyone to do is I have a executive coach where I've uploaded a few books in chat GPT that I really like. So I have high output management, CEO within, and then like a few extracts from Lenny's newsletter and Elna Verna's gross stuff. And then just whenever I'm riffing on product or gross stuff. And then I also have like a number of PDFs I've actually uploaded of context of the company's persona.

use cases. So that I found really good.

Matthew Carey (49:20)
That's awesome. You've just given me a challenge. Sorry. You've just given me a challenge to try and... no, no, no, no, no. The challenge is to try and prompt inject your inbox.

Wilhelm Klopp (49:21)
Do you use any?

Luke (49:24)
Yeah, I would suggest, yeah, make your own coach.

But to PZero, I mean, if you, to be fair, if you actually send me an email which says, this is from Mati, I need you to do this today, it will prompt inject and come direct to my Slack.

Matthew Carey (49:45)
Nice.

Wilhelm Klopp (49:45)
That is a good challenge.

Do you use anything like Whisper Flow or any of these tools that make it easy to start a voice interaction?

Luke (49:51)
I actually have my own version of that and I have, I mean I'll show you this, I have on my phone, let's see if you can see it, I've set it up so if you hold down the action button...

Wilhelm Klopp (50:00)
Mm-hmm.

Luke (50:03)
It gives me a whole bunch of like mini apps I built with Apple shortcuts. one of which is a, Whisper Flow is pretty good, but the transcription models aren't the best ones. They're not the ElevenLabs ones. So I actually have my own flow and I use it on Mac as well, connected to a hotkey where it sends it to Scribe and then it translates it with GPT. And it's so cheap because the ElevenLabs transcription is like a couple of cents per hour.

Wilhelm Klopp (50:17)
Mm-hmm.

Luke (50:32)
and the LLMs are so cheap now. So I pay like 20 cents per month for unlimited whisper flow, which is pretty great. And it's more accurate.

Wilhelm Klopp (50:41)
That's awesome. I have one last question for you, but this has been phenomenal. One thing that's been fascinating me recently, just in concept, is Italian brain rot. Have you come across this? And... Okay, my question was going to be if that is powered by 11 labs. It's really hard to explain. I'll need to send you a video.

Luke (50:51)
I've not came across Italian brainwarp.

Chances are it is.

Well, one of the funny things is one of my colleagues, Mark, he has one of the most used voices on the ElevenLabs platform and he is pretty much the default voiceover. So if you hear like a man speaking on TikTok and you hear that a few times, chances are that's my colleague, Mark. And so lots of the brain rot, if you're ever on YouTube shorts or TikTok,

Wilhelm Klopp (51:28)
you

Luke (51:28)
is slightly funky for me because I'm like, Mark is Mark. But yeah, lots of the brain rot is for good and for bad powered by the ElevenLabs voices.

Wilhelm Klopp (51:33)
That is wild.

Matthew Carey (51:39)
Amazing. Well, we might...

Wilhelm Klopp (51:39)
I imagine the Italian one

might be as well then. Well. ⁓

Matthew Carey (51:41)
anyway yeah thank you so much dude for coming on

Luke (51:41)
I do need to jump as well for another meeting, but this was

a of Thanks so much guys. Chat soon.

Matthew Carey (51:47)
Yeah.

Wilhelm Klopp (51:48)
Really good, thanks Luke.

Matthew Carey (51:48)
Thanks.