Nikola Mrkšić, Co-Founder & CEO at PolyAI on Building One of the World’s Leading Voice AI Companies

0:00Hello and welcome to another episode of Riding Unicorns. Today we're delighted to be joined by Nikola, the co-founder and CEO of PolyAI. Nikola has once been on the podcast before from the Capsule event, but today we're going to dive in a bit more into the PolyAI journey and the future of voice AI, which is obviously very exciting. For those that don't know, PolyAI is a leading supplier of conversational AI for automated customer service. But I think the product suite is expanding, so it'll be interesting to get into that a little bit. And you recently closed an 86 million Series D from top investors.

0:34So you are, you know, one of the biggest voice AI companies in the world. And Nicola, you were previously a machine learning researcher and part of the team that built Siri, which I know that you have some pride and some sort of regrets about how it all went. So it'd be interesting to learn a bit more about that as well. So welcome to the podcast. Let's just go back a bit and go back to when you started. What did you see in conversational AI that the world is only now currently waking up to? Yeah, I think like, and to go back to kind of like your summary of my life story, our first company, VocalIQ, was acquired by Apple to make Siri more conversational.

1:19And Siri had a grand ambition of being in that future Jarvis-like state, right? And that's a lot to ask for. It's really a lot to ask for because the technology was nowhere near good enough. And I think the bet we made with Polly was like, how do we go for the step in between where we do something that many companies do in great volumes? How can we do it with technology in the way that humans do in contact centers today in order to move the technology and its evolution to the next step where it's not just a clunky IVR that saves cost. It's really a very poor vision and version of what we have, but really something that you look at it and you're like, wow, thank God they have AI.

2:04Thank God they've implemented poly AI because it means they care about me as a customer. That was the vision. That was what we set out to do, to build a few systems that people would like, or at the very least not hate the way they hate most voice automation right now. And, you know, we've been at it for a while. We've been very successful. I think we achieved that a lot more with our deployments. And then we've scaled it through a platform, through partners, through everything that goes into working with and scaling software to a point where now, you know, we process. I think we crossed the point of like half a billion conversations now.

2:42And, you know, we'll do a lot more. It grows exponentially. So it'll be more and more every year. But, yeah, the goal was always just like how do we take the next step in that human-computer interaction? And Nicola, I mean, I think you're doing many, many calls. I think I read somewhere you've done like 150 years of calls through the through the platform now. And I wonder, I think I think everyone knows all our whole audience will have heard of voice AI and will know that perhaps they should be using it in their in their own businesses. But tell us about an example or two of how your customers have transformed their businesses using voice AI.

3:20Yeah. I mean, look, a simple example is in hospitality, where we work with a few thousand restaurants around the world or with one in two different casino groups in Vegas. So like one out of two properties in Vegas run a poly AI system so that if you call reception, someone picks up immediately. Right. and that someone is AI and, you know, has answers to the questions. It can help you order, you know, fresh set of towels or anything else along those lines. And then, you know, beyond that, you know, for restaurants, it means that they never miss a revenue-generating phone call, right? They never fail to make an appointment.

3:59We've had cases where that increases revenue by 5 % to 10%. And, you know, the average lifespan of a restaurant is actually five years. So that difference of five to 10 % can be the difference between whether they make it that year or not. So it's a great example where you do it because it makes your business better. It gives you more top line, not just like, you know, I pinched a few pennies. And because of that, you know, like my business is better. That's small thinking. Beyond that, you know, we have many large customers. You know, think, like say, PG &E in California is the largest utility company in California.

4:37One of the largest, if not the largest energy company in America. And with them, we've had a day where we took nearly a million phone calls, right? It was a biblical flood where people were cut off from gas, from electricity, and they called in absolutely crazy numbers. And we picked up every single phone call, right? So it's a pretty bad day, but it would have been even worse had they not been able to service that demand and tell them, we know that the power is out. Like you don't have to report it. We're on it. ETA, X hours, right? It's not great that it's happening, but it's better that you communicate it clearly immediately at the right volume than that your phone's crashed as well on top of everything else going on.

5:26so you know and beyond that a number of retailers where we save their kind of like whole holiday season because the calls spike by a factor of like three to five times and the issue is they don't spike just for a single customer they spike for all of them so all the pools of extra kind of like labor that they can tap in like bpos or their local contact center kind of like suppliers like they can't they're all fighting for the same resource so you just can't get enough people and And, you know, you should see how a retailer prepares for a change freeze. It's like the whole year revolves around that period, much like in hospitality, it revolves around that period because that's where they make up the losses and turn into a profitable business.

6:07And, you know, we work with large banks like Unicredit or NatWest, Metro Bank, with insurers like Allstate, Liverpool, Victoria, Simply Health, really a bunch of different ones. And there it's all about just like picking up every phone call, doing what's right, reporting on the data, you know, like for one of the largest logistics companies in the world, we've been able to identify patterns of behavior of, you know, like operational issues in a depot that may arise even from just one rider doing something incorrectly. The writer doesn't know. They don't know why the overall metric is falling. But if you can kind of like dig into the data and triangulate with the call patterns coming, you're able to surface operational issues that lead to those calls.

6:55And that's always been the issue that, you know, people say contact centers are boring. They're a cost center. Well, they're a cost center because you're not doing something right with the rest of your business. And the more you mess up kind of like further upstream, the more you have to spend on contact centers just to do failure management. Right. It doesn't make your business better, but you have to do it nonetheless. Right. It's a tax on not doing something else right. and if that whole channel can come alive and feed that information back to the business well then you've got like a real thing then you've got like a brain you know we we call this like the agetic enterprise where you basically start having like nervous system that goes all the way out to those functions and goes oh this hurts please don't put your hand on the stove or you You know, hey, calls are spiking about billing issues in Newcastle.

7:46You've done something off in, you know, like the Northeast. Go check it out, right? And, oh, okay, let's send out communication. Like, we overcharge you for the month of May. We'll be sending out a revised bill in the next 48 hours. No scandal, no media reporting, no social media outrage. Business as usual, right? So that's really appealing about these systems. It's that the upstream effects of adding intelligence to that layer, like systematic, you know, hive mind thinking intelligence or the kind that AI today can deliver. You make your business a lot better. Yeah, I mean, you've done so much more there than just kill the cool queue because obviously you can handle millions of calls asynchronously.

8:33So that was probably a big early feature. But actually being able to have all that feedback feeding into every team and every system is absolutely massive. You mentioned a number of use cases there, different verticals. And we've recently seen a rise of voice agents that are very vertical specific. The one that comes to mind is Linda AI, which is, I think, for dentist receptions. It's like that's their thing. how do you think about the need for verticalization versus poly ai can do it all so these things are never really going to be as good as the the super agent which poly ai is you know depends what kind of answer you want give us the most controversial answer yeah look look by and large the latter like you don't need to be vertically oriented towards its dentists.

9:26We've got a number of very large enterprise clients that are dental chains. I don't think any of our deployment teams even thought of that as being particularly special, right? It has a set of backend integrations that we set out to learn, do, productize, and then when the next one shows up, it's easier. So on that front, yeah, if you're going to do the first use case, then the second one is easier if it's the same use case, right? And more broadly, It definitely doesn't have to be that specific because an appointment for a dentist is no different to an appointment for a vet or an appointment for a dermatology clinic or an appointment for at the end of the day, it's almost not that different from a restaurant or even a hotel.

10:06Hotel gets more complicated because the order itself is more complicated, but it's just really subclass of appointment. Right. And that is by far the largest use case for conversational AI full stop. right appointments scheduling order management like that's easily half of all the volume across all the vendors right and then you know you've got other pockets like outbound is big so outbound lead generation outbound sales uh we do quite a bit of that it's a good use case it's a bit harder to figure out where you are in the clear when it comes to the word of law who can you call when it gets too much whether it's the right strategy or not because you know sure i could call you a thousand times and you know if i do that i will almost really guarantee that you never buy anything from me ever again and then you probably put in considerable personal effort to tell others that they shouldn't buy anything from me either right so with that you have to be careful but it can be really good when part of like a really good orchestrator journey then you know you've got things like account management itsm use cases like you know password resets technical troubleshooting, a number of things.

11:16And it's all really, really interesting. And the potential of the whole thing is large. But the more of something you do, the easier it is to go live or to have a first party integration. I think the one thing that is particularly relevant now, especially in the aftermath of the SaaSpocalypse and those software vendors that have the right to be a system of record, deciding like, what are we going to do? Are we going to like open up freely to these things and let them like, um, Sachs said on the old podcast, it's like, if all the value goes to the authentic layer, right? Like all these guys become is a utility company.

11:52They're providing something that you have to have kind of like telephony or a system of record like Salesforce or Zendesk or ServiceNow, any one of them, right? Freshdesk. Do you just resign to being that system of record that they have and they flip between one out of four or five, and then all the value of the labor placement on top going to someone else. Well, you can fight that by saying, I'm going to close down. You can't have access. The only way for you to do something like that is to buy from me. Now, their solutions for that are not very good. And that's just because big companies move more slowly and don't innovate as fast.

12:27Will they get there? We will see. The issue is, can they afford to close down and piss off all of their customers that want to work with Apolly and risk that they go somewhere else because you know i think as a system of record the moment you've said you're going to do something like that you've crossed the rubicon you are no longer a sales force you're becoming maybe a bit more of an oracle or someone like that where it's like all right castle walls up like you know drawbridge up moat filled with crocodiles and it's like let's see if you can do it all yourself or you know you've heard carp talking about palantir I'm like, go on, rebuild it, rebuild it.

13:06I dare you, right? So yeah, I mean, they've spent like 15, 20 years building something. It's complicated. It serves a particular purpose. It's not easy to rebuild it because the insights, the integrations, the know-how needed to replicate everything that they do is hard. There's a mode in there, 5, 10, 15 years of mode. Maybe it's endless. Maybe they're the only ones that now fill that need and it's hard for anyone else to get in. You can put it to the test if you close down. I think for the systems of record, it's a dangerous game. And we don't see them doing that just yet. But yeah, we'll see.

13:42And Nicola, where do you see your moat coming from? How do you see the market for voice AI playing out? Will there be a number of winners? Or how can you build a very sticky product that's hard to rip out? Well, I mean, I'll tell you why it's hard to rip out, right? Like by the time you've started working with someone, you know, week one, two, three, you're very easy to rip out. It's a matter of a telephony integration. and then you're probably answering some questions. Maybe you've understood what's known as skill groups, kind of like where you route the different goals when you can't do something.

14:12That's not that hard to replicate. But then you start integrating with them. It's kind of like joining the EU. You go in and you change your laws and your paradigms integrations to work with their ecosystem. And they've got a special instance of this system of record. They've got a special system homegrown for their loyalty system. They've got another ticketing system for this, another action trigger for something else. Very complicated, right? So you go through all that. And there's a data point which says that basically your churn risk in enterprise software is directly correlated to only one thing, which is the number of different integrations with other systems.

14:53It kind of grows quadratically, your stickiness. So when you scale deployments to the point where your AI is doing the work of a thousand plus people for a single customer, and right now we've got at least five customers where that is true, it's not just that they'd have to hire a thousand people. Because in theory, if it's so easy to replace, they could buy it from someone else or build it themselves. It's really the fact that they'd have to go through that procedure again. Now, in an ideal world, a phenomenal software systems architect looked at it all, abstracted everything away, created an MCP service that worked for all of this.

15:27But I can tell you that in practice, that is complete science fiction. And it'll probably take a very long time for that to happen. And the reason we talk about forward deployed models is because these are hard environments where documentation is lost. You're often doing archaeology to figure out how to do something with a client. They don't know who to even reach out to internally to find out how to do something, right? How do you avoid the voice becoming commoditized? What do you mean by that? How do you avoid voice AI becoming the next utility company where margins are driven to near zero? Will there be pricing power in this market, do you think?

16:07Depends on what level you're looking at. So the application layer, where it's like a solution doing something, where if it's integrated in this way, it built up effectively an application which is built to play nicely and quickly to bring time to value to like 30 minutes instead of a development project for six months then you have like an inherent advantage then you have pricing power you have pool you have an ability to say okay you're a bank well like what's your bank from polyai is someone who's already been approved by the european central bank by the fca we have systems running where you know it's lived the test of time and your competitors have renewed a number of years you're buying a near risk-free kind of like addition to what you do our job there is to choke out all the other players so that they don't get those case studies the moment there's enough of us then there's a pricing award if our capabilities are the same we don't see that happening because most of the competition in conversational ai are not even voice companies right they're chat bots and they talk about being omni-channel and they plug into like deep gram or 11 labs or cretesia and they're like well okay now i have voice and they can demo it and then in production none of it works and that's kind of like where the market is right now so there are very few companies that are running serious enterprise agents you see a lot of demos and i think that we'll live in a time where a lot of these companies write case studies before they're even live and uh you know look power to them i think it's uh i don't know if it's good marketing, but it's definitely an impressive discipline.

17:41I think for us, you know, we've always been quite measured around like these claims also because it's not particularly popular to go around telling the market that, you know, you don't do the work of God knows how many tens of thousands of people, because they do definitely disrupt some jobs. And many of those jobs would not have been filled. And many of them are roles that face constant attrition, rehiring, untrained people, part-time workers jumping in to help, but like the impact is real. So we've always been measured around like those companies that want to improve their customer experience and are willing to pay a premium price.

18:14They work with us, but the whole like commoditization, we're not talking about electricity here. We're talking about, think of it as a composite electrical good for now, because it's not really voice AI that we're selling. We're selling voice solutions that solve particular problems. So the mode is in that like application layer versus the model layer. I think in the model layer, it already is commoditized. And Nicola, I heard someone recently say that no good innovation comes with good margins initially. But I wanted to know if that was true with voice AI, because it seems in the AI space, we've got obviously LLMs, and then we've got voice models and video models.

18:55Video models seem to be completely uneconomical viable from what I can tell. LLMs are great for lots of use cases, but the coding side of it, there seems to be some margin pressure for the wrapper platforms like Lovable and Cursor, etc, where they've had to adjust their pricing a lot to try and fit their margins. But in voice, you're literally replacing a human who would be quite expensive you are generally paying per minute and charging per minute so it feels like you can have quite reliable margins built into the product from the start is that correct or how have you seen things change in in the time that you've been building the company.

19:45If you're a pet owner who wants real talk on training, behavior, and everyday life with animals, check out the Fresh Patch Podcast. Hosted by Drew and Gabe, we cover the wins, the challenges, and everything in between. So wherever you listen, search the Fresh Patch Podcast, where good pets get it.

20:09I mean, when we started doing this, we were the only ones. And it was near impossible to even sell it to anyone because they thought I was nuts. The flip side was that when we got in, like the pricing power was incredible, right? And if you got 3x ROI over humans, you'd still be charging well over what you see today in the Boise AI market, right? Because the ROI was there. But the truth is, and I'd really, you know, having gone through a few cycles of these renewals and thinking about the whole thing. With technology, we, to our customers, owe that the use of our tools gives them serious leverage.

20:45Because it's not just paying us versus paying humans. They are committing to build a new kind of solution on top of Poly AI to, at the end of the day, make their business lever and get themselves financial leverage, revenue upside, seaside improvements, right? So I think to me, being transparent with pricing and saying, hey, I will charge X cents per minute. And that's what I want to do because it gives me healthy gross margins. And if it works for you, that's great. And then obviously there's a bit of haggling and volume discounts and all that. But we find that to be the cleanest model. We've done a load of outcome-based pricing.

21:25And in theory... Nicola, can I just ask a follow-up? Is that the sort of the bit that some people are missing with voice is that it does have this business model that has margins in it. And it's that's why it's a great space to be in. Like, you just sort of validated what I was sort of suggesting. But I don't run poly AI. There's two parts of it, right? So we've always had very healthy gross margins, right? And part of that is having our own models, right? I think you'll see a lot of the wrapper companies, right? Think like Sierra, Decagon, Parloa. They're built on other people's tech, right? And that just means that you're like a seller on top.

22:03And that, if you want healthy SaaS margins, makes it a lot more challenging to be competitive on price, right? So they have to be really expensive. Will they just be betting that the models get cheaper? They are betting on that, right? Is that their only bet? they're just hoping that it's like they're just gonna assume that that will pay off in the long term i don't think they have a choice so it doesn't really matter right for them like they just have to plow along and hope for the best the the truth is like they you know like maybe it gets cheaper maybe it doesn't they are they don't have strategic autonomy and that's problematic because they're basically just a value-added reseller to open ai and then you know we'll see we'll see what happens with that if you have your own models you have like one extra set of like you know positions to retreat to because your cost of service probably like that extra layer of adding five extra price to get open ai gross margins does not exist right so i think technology does matter and that autonomy does matter can you build a great business otherwise probably right so the way that some of them do it is they try to do outcome-based pricing where you know they charge people a lot for successful outcomes and that works for the likes of zendesk and intercom because they have a large set of clients with not large volumes who would not actually go and build the thing on their own they have a pre-baked first party integration with their system of record they are their system of record so they can offer them something that you know paying with it with higher OPEX, but not committing any CAPEX just makes sense because it's the risk compared to doing anything else.

23:48So I think they've built like really impressive businesses. I think both are a hundred million plus in ARR from the kind of like AI line of business on top of the standalone intercom or send this business, right? When it comes to other players, like say Sierra, where they try to price an outcome, I've done it as well. And here's what happens in practice. a year in like you say pay me two dollars if that works because that outcome is worth ten dollars right the moment you've done that the clock starts ticking and internal it is like well like those guys are charging two bucks for that when it could be like internally built 10 cents a minute so close like three minutes 30 cents they're charging two dollars like what's the Like by the time they built it and they prompted the living hell out of open AI systems in the backend with a few kind of like little modules they built on top.

24:45All that data belongs to the customer and can be used by the customer's IT team to do the same thing. And it used to be hard for them to do the same thing, but with coding tools and stuff, it's just getting easier and easier, right? So that whole like borrowing from the future by having that level of pricing, that will go away. So that pricing power of like charging way more, because like in theory, it makes sense. Like you can sell a few deals like that come renewal. They're going to turn around and either like just go and build it on their own or they're going to force you into a conversation where it's either, you know, per minute on a fair price or much lower rates for outcome based.

25:21But it comes back to that philosophical thing, which is like if you're a tech company, you should be building your technology. And that technology should be giving people leverage in their work, in their daily work. Right. So everything else is trying to be this kind of like geezer who's coming in and like, well, we do that. I get half and you get half. The other part that people forget is if you don't commit to doing something, building something and running it, which is parent pricing, you're kind of giving yourself an option to work on it. And if it's great, you'll get paid a lot of money. And if it doesn't work well, it's OK.

25:52You don't pay anyway. But if I'm a vendor, sorry, if I'm a prospect and I'm buying from one company that will do something and charge me transparently for it versus another company where it's like, hey, I won't pay anything if it doesn't work, but I'll pay through the nose if it does. Well, why? Why would you do that? So I think like the outcome based thing, at least in CX, for the most part, doesn't make sense. You know what does make sense? Like licenses. Many companies are complicated businesses. that kind of like spread costs around constituent units, right? You mentioned dental offices, right?

26:27They will probably be passing the cost down to the individual dental office, especially if it's a franchise model. At that point, they kind of want to know what they're going to pay and they will find both outcomes and the license, both the outcomes and the consumption-based model to be too difficult. Because if I'm like a restaurant and I get like 50 calls a day, I don't want to like every month look at like how many calls did I take? How much should I be paying? How do I budget for it? I'd rather just know that I pay 300 bucks a month because then I can budget for it. So there's really just like real considerations of real businesses buying this technology.

27:02And much of the mode is like knowing what they need, building a system of like value additive things around your solution so that it's super valuable to them. And at the end of the day, we're back to like, you know, like square one. We're just trying to build good, useful software. James you had a question about growth and I had lots of questions but I think we we need to um get towards uh the latter part of the show so well in that in that case yeah I would like to hear the story of Gordon Ramsay Nicola I think you were working with Gordon Ramsay and I'm sure our audience would like to hear the use case there yeah I mean Gordon Ramsay has been a client of poly AI for about two years now, his restaurants.

27:47And, uh, it's exactly the kind of client we love someone with like crazy standards, really passionate about what they do, the best at it, like we are. And with them, it's simply like, okay, well, like if they're not afraid of putting the CA in front of their customers, why should you be right? That's why we work with Gordon. That's why we work with Caesars with Marriott. Right. And, um, you know, we had this idea of kind of like showing him getting like really frustrated with like a restaurant he called that has an automated system that's not very good. And then kind of like calling his own restaurant and just being delighted by how good it is.

28:22We reached out and recorded the whole thing. It was a really fun day. He is a really awesome guy. You know, he doesn't act. He is who he is. He's that on screen and off the screen. And, you know, I think that's, that's my favorite kind of actor, even though, you know, he insisted that he's not an actor, but Adarki is probably one of the better ones. brilliant is that is that and is that video out now it's available yeah i mean you can look it up on youtube it's like i think we have something like you know four or five million views on it it's more than like the open ai and tropic super bowl ads put together times some right it's definitely the most watched super bowl ad of the last super bowl so i think we also really like the idea you know as as a british company that is now like 80 american revenue we're kind of like gordon right like he's a brit who went to america and made it really really big there i mean not that he's not big in the uk but really you know i think america made his career from hell's kitchen kitchen nightmares and all that and i i really like gordon ramsay you know i think watching him has a has a therapeutic effect on me which probably says more about me than about anything else but um yeah like uh it was a real drink i'm sure because like you think like who would be a good brand ambassador and i think for us how we think about our product about quality about good experiences like he's the guy yeah very cool he's got very recognizable voice as well but also i think it opens up a whole nother conversation which unfortunately we don't have time for about like b2b companies having like celebrity brand ambassadors and influencer partnerships and things like that but i can see the alignment and it's very very cool and it's great that his actually he's actually using product in his restaurants there's obviously clear alignment there very high authenticity which is critical for any influencer stuff so nicolo we've got our final two questions which we've just got time for so the first is a future unicorn prediction so is there another company that you've seen that you think has a lot of potential yeah i i think i saw you guys interviewing uh manny medina from paid ai yeah he's great he's really really great he's you know i think uh one of one of this like new wave of you know serial entrepreneurs who've like found their way to london and uh i think that's really good thing for the uk and the european tech ecosystem he's trying to basically like find ways for companies to monetize agents in the kind of post-sass world and you know i had a few really really strong conversations with him and you know i think that the beauty of like second time entrepreneurs who've like taken the first company to a great heights like he did with uh with outreach is just like you know the next time they're playing the game they kind of know how to get to a similar level in like one tenth of the time so i had to bet on one especially locally it would probably be him yeah very cool yeah he's a bit of a machine and they seem to be doing really really well and it's very he's really thinking about how how ai is actually going to change business which is really really cool and then our final question is our dinner party guest game so if you could have dinner with any three people not manny um who would they be it's a hard question right i mean topically just in this space and stuff i think like one founder i really really like and always have is jensen huang i mean i've met him i've not had a chance to like sit down have dinner with him where, you know, if drinks start flowing, you can ask even more authentic questions.