Kane Simms: Hello there. I’m joined here by Nikola Mrksic, who’s the CEO of PolyAI here at CCW Europe in Amsterdam. two things I’ll be interested to talk to you about. For those of you that tune in regularly, you’ll know that Nikola was on the podcast fairly recently. We had a really good conversation about Raven, your fine-tuned language model at PolyAI, the kind of approach to that researchled kind of innovation. Since then, two interesting things have happened. one is that you now have Wren which is the kind of self-improving agents kind of idea. and the second thing is we’ve started to see the beginnings of kind of agent-to-agent communication uh which was always on the agenda but with Muse and Dots and Grokbot in the last sort of two three weeks it seems to have just really it’s there waiting to happen now so to speak. So maybe we’ll start with that and then we’ll get on to the idea of the self-improvement agent. What do you make of the the state of agent-to-agent communication? What do you make of Muse Grokbot dots and how are you starting to think now at PolyAI about this potential world where me as a consumer are going to have my own agents doing my own stuff and how do businesses prepare for that? How are you starting to think about that kind of stuff?
Nikola Mrksic: No, no, look, I’m really excited about that. I mean, I started Poly after leaving Siri because it wasn’t amounting to what people expected of it fast enough. So I’ve been hugely passionate about both personal assistants and voice assistants especially where they overlap for a very long time. I think that ever since the clawbot mania of about you know now 6 months ago is we could kind of see the moment coming and I think you know the valley has decided the hyperfunding’s coming in and you know we’re going to take a crack at it for the first time properly since Alexa was a thing right so I think it’s huge I think we see the number of calls growing from admittedly from a very small basis but like you know I we still don’t have one full month for me to tell you how much It’s jumped but it’s at least like 10 times. back in 2020 we had uh the early kind of like you know sign of this with Google Duplex calling one of our agents. I think to us it was like one of the first like you know the Dr. Livingstone I presume moments of AI talking to AI and it wasn’t a very good conversation back then. I think agents on both sides are far better equipped now to have those good conversations. I find it hilarious that they’re communicating with human language, right? that they’re burning these tokens to produce voice to understand voice. You know, it might be someone speaking with uh you know, Georgian accent calling and someone with a Geordie accent picking up and we’re doing all that for really no reason in that they should be communicating at least with like text if not uh voice. And it’s also funny in that in the CX strategy of companies, voice has always been meant as the escalation point away from it, right? So now that you kind of get parity on the caller side with the automation that was mostly facing inbound calls by enterprises, it’s a whole new world of having to like have an arms race between what you can do in terms of provisioning demand and the demand being created. But I think we’ll reach a steady state where a solid chunk of the calls are uh aagenticc. I think we did another kind of like survey of those calls that were able to detect as automated and u we thought it would be dominated by scammers and there are a lot of like scam outbound calls but it’s not. I think already with these launches now well over half if not quite a bit more are legitimate real calls. So I think companies are going to have to accept this as facts on the ground. They might have a slightly different level of engagement and handoff to AI agents versus the non-AI ones. But I think that Jevons paradox where you know instead of just getting to a level of service and stopping, we’re now going to overproduce this like labor automation on both, you know, the demand side and the supply side.
Kane Simms: Interesting. is there anything in particular that you do in order to detect that it’s a an agent calling? And secondly, this could be something that happens quite quickly. You know, ChatGPT went from zero to a million users in like days. So there’s always a potential that this thing can really take off pretty quickly. One, how do you go about detecting it? And then two, as a business, how do you how should you be thinking about dealing with it? Is there any difference between how you should treat an an an agent-to-agent interaction versus a human to agent interaction or not?
Nikola Mrksic: Yeah, there are many unanswered questions here so far, right? So I think one key question is like how are they authenticated and I think one of the better things about muse and instinct and you know is that well or at least like the grade of agents that will become used at scale like that especially muse is that in theory I should know that it is Kane Simms and that I can trust that it’s you and therefore whether it’s human or not like if I pre-authenticated you then we can skip a part of the conversation but that should be true in CX already and it’s not really implemented at scale. So I think it just adds another variable to something that sadly as a community we’ve not really solved at scale yet. So maybe it increases the need to do a lot of the digitalization and standardization in order to get to a point where this Kane. I know it’s Kane. I know why he called before. you know, a lot that we’re doing with our customers already lends itself well to, you know, keeping, you know, we refer to it as a dialogue that sits, you know, in the above a conversation. It’s like an ongoing set of interactions about you, what I know about you, your habits, your history of interactions, so that I can just kind of like get to the point a lot more quickly, ideally know what you’re calling about immediately. It’s really what a lot of true omnichannel has always been about, but never really reached. So I think as the technology reaches a point where we can do a lot of that more you know effortlessly that’s where I think these things just give us a reason to do it all the more cuz otherwise you’ll have contact centers collapsing and you won’t be able to pro you know provision enough humans or AI potentially to to deal with the demand. Of course, you can get enough AI, but like if it’s not able to do anything, then well, I mean, we’ll love it cuz we’ll get paid as will the outbound call originators, but it won’t really lead to improved customer satisfaction at the end of the day.
Kane Simms: And what does that mean for PolyAI where there is an increase in kind of like that agent-to-agent conversation type of thing? Do you need to do anything differently? Is that good news for you? Because as you just said, you know, voice was a was the last resort channel. Now maybe if I can get Muse to make a call that might be a preferred channel therefore calls might rise therefore the need to automate more might rise. So how what does that mean for PolyAI in terms of
Nikola Mrksic: Yeah I mean that that’s back to the Jevons paradox which I think is one of the like best things that has historically happened in every like industrial revolution. So rather than people just like driving to work they buy two cars move further out cuz you know the fuel prices went down. similarly here you know you you’re able to do more with AI. So, you know, you automate and then people provide their own agents to deal with your automation. So, it’s kind of like just a necessity rather than necessarily an improvement. Of course, it will be a better experience when we’re done doing all that. I think agents should talk to each other, you know, in text. They don’t really have to do it in voice.
Kane Simms: Do they even need text? Wouldn’t kind of like a series of API calls just being made behind the scenes solve this problem or is there
Nikola Mrksic: Depends on how language solves how agentified and AI-skilled you are. I think where deep learning has triumphed as we always knew it would is in the unstructured data being passed back and forth. LLMs allow you to not have to pre-specify like you know API fields and stuff. No, that’s what MCP is about. In reality, what we see in practice is that most MCP or A2A connectors are not yet solid and then you know that language around that interaction is something that happens because cuz you know if you just had clear API connectors you’d probably have an app that did it anyway and you don’t. So that’s where it’s just you know one question is the enterprise adoption overall. The other is just what do we need to do? What do we need to build to make it all work? Most MCP right now is quite slow and agents using APIs fare a lot better. It’s a lot faster, a lot more effective. Those are also usually far more battle tested. This true outside of CX. This true for using Gmail through Claude, right? so yeah, I think like there’s a lot like like it’s really exciting because we’re doing so many things that we could only dream of before. On the other hand, you know, the taxing work of standardization like actually we just see like the wild west. It’s full of opportunity but it’s also unregulated and that’s I think okay for the most part.
Kane Simms: Yeah. Interesting. finally I wanted to chat to you about the idea of self-improving agents as you know if we’re saying that there is a world in which personal agents create more demand businesses are automating more and more kind of conversation automation happening. it then becomes part of the kind of job of scaling that means that you’re not only building new stuff but you’re also needing to improve what you’ve just done. I kind of had a bit of a sort of ramble at the beginning of the year where I was saying that I think there’s a world actually where AI agents design themselves, build themselves, test themselves, analyze themselves, improve themselves and therefore you know it doesn’t really need kind of people to be involved cuz the AI has got models than we have. And this year we’ve seen quite a lot of that happening. You know there’s another some of your competitors have released something similar. You released Wren Self-improving agents. talk to us about what’s the concept of self-improving agents and how does it kind of work behind the scenes?
Nikola Mrksic: Yeah. Yeah. I mean, I think there are three levels of Wren, right? Like the first level is a coding agent where you finally bridge that like self-service for, you know, the business side where they’re finally able to do things. You know, we’ve got people sending 2,000, some have reached 3,000 messages a month with Wren. Now, I am terminally addicted to Claude Code and I don’t do that many in an average month. So like there’s a real kind of like AI-pilling of the contact center that coincides with the adoption of Wren by our customers where those guys are finally that like self-service you know human layer above the sea of AI agents dealing with your customers. That’s really good because it’s badly needed because the combined motion of you know hyperscaler brings in a gsi to deliver something has not really ever gotten to the point where it iterated fast enough to build Like a robust thing that could do most of the things that your contact center does. Like it’s your contact center that must encode those rules into your agent. So that’s level one. Level two is you’ve launched it and then like how do you get it to self-improve? Well, you know, another agent looks at all your calls and suggests, “Oh, I’m seeing this in 10% of the calls. I’m seeing this in 5% of the calls.” It basically goes away in very kind of like long horizon tasks and just delivers something that is 10 proposed fixes for the morning, 10 where it got there almost, but it’s unsure about the context. You go and you go, “Oh, I love that.” Click, click, click, click, click, accept. Right? For the other ones, you add a bit more. That goes into Ren’s memory. next day when it’s looking for new things, it has a richer context. So you iterating with it every day is to build up the context layer that as it gets better and better and better and you connect Wren to your Google Meets, Zoom Teams to the channels where you’re discussing it internally and that all just feeds that context. The moment you reach a sufficiently well built out context layer, you can let it go into level three, which is like full autonomy where some of those fixes can just go in AB test, AB test, AB test. If you’re running, say 10,000 calls a day, you can go up to at least five potentially 10 parallel AB tests and still for the most part get enough data to within a day or two say this worked, this worked, this didn’t. And then you can take it back to the human or for those that are like unambiguous like flipping the voice to that voice or changing the phrasing to this phrasing where we know and the LLM judge knows that it works. Then you just go merge, right? So that level three is in beta with a few customers. U it’ll be released later this year. but I’m super excited because it was always true that there was this law of diminishing returns of like an extra change and an extra change and an extra change and when it had to work combined with CX teams, this was a very expensive maneuver. When the CX teams can do it on their own, they’re going to go a lot further than any forward deployed motion ever could. And you know, we’ve like died on the hill of pioneering and defending the FD motions. I kind of feel like I’m backstabbing myself by saying, “Hey, that age is over because we will never and you know, I’ve got customers like FedEx who have been my customers for four years. I know everything there is to know about their CX. I’m not in a position to make daily calls about their business in a way that their people can. This why it’s really important for them to get AI-skilled so that they can go to a higher and higher level of autonomy in using Wren to help themselves because you know maybe it takes 10,000 changes in tickets to go from like 30 to 80% automation and with humans you might have stopped at the first 100 and hopefully you’ve gotten a 100 of those that add the most. So maybe you would have gone from like 30 to 50, right? But why not do the remaining 10,000, right? We finally can. And uh I think where I also sleep better at night is you know we see people rising through the ranks and being the most AI-pilled in these companies cuz they’re now harvesting this data that they have and other departments do not have this level of data to work off of. This oil that they’re extracting right and through that I think we are training the next generation of not just CX leaders I think the next COOs of large companies come from the AI-pilled CX managers of today.
Kane Simms: Nice. Brilliant. Nikola, absolute pleasure as always. Thank you so much.
Nikola Mrksic: Thank you for having me.
Kane Simms: Nice one.
