It's plain to see: The next generation of AI is here. But what isn't yet clear, is where these game-changing developments will take us next. That's why we've created the Future-gazing AI series. In an ongoing series of video Q&A sessions, we’re casting our gaze ahead — to see what our AI-enabled future might look like.
First up, we sat down with Nikola Mrkšić, the CEO and Co-founder of PolyAI, the leading provider of customer-led voice assistants. Nikola was in conversation with our CEO, Reetu. Together, they discuss their projections for how contact centers will evolve in light of these AI advancements.
You can watch the full session here, or read the transcript below.
Reetu Kainulainen: Welcome everybody. We are future-gazing right now, and today it’s a special — first of all, thank you Nikola for being the first guest for our series. We wanted to start with a big big bang. So, great to have you here and maybe you can start by introducing yourself. Who are you, and what are you doing?
Nikola Mrkšić: Thank you for having me. My name is Nikola Mrkšić. I’m the CEO of PolyAI. PolyAI is a London-based conversational AI company. We build voice assistants for customer service, and we’re trying to change the way that people feel about automated customer service over the phone.
RK: Beautiful. And I guess to kind of kick things off, you’ve heard of this thing called ChatGPT and LLMs. I’d love to hear from your perspective — we’ve shared a lot of our learnings and views on it in the past — but would love to hear from you: How has the last 6 months been for your company, and how do you feel about the whole hype that is going on?
NM: I don't think it’s hype. I think it’s all real, right? In the sense that we are genuinely going through something that’s going to be like an industrial revolution with AI.
You know, I’ve worked on deep learning since 2013 and — at one point I had a paper submitted called “Language Is Not a Thing” and that got rejected because it was too sensationalist. But you know, we were training things in English and dropping a few Italian examples and, just like that, the thing would speak Italian. And it was really great. But I didn't see it going this far, this fast, right?
And I think — more than most people — I was very close to what was going on. Working to build task-oriented dialogue, I think we got a bit too pragmatic. So I’ll admit that with GPT-3, people were showing me examples and — anyone who’s done an NLP course in the past 10 years has trained a language model and generated some Shakespearean-like thing, where you write the first 20 words and it continues, and you know, it’s semi-intelligent but definitely Shakespearean sounding — so I was just thinking, you know, it’s higher perplexity, better model. Of course it can generate longer things. So what? Right.
At one point I remember just playing with it at 5 AM and realizing that it can do three-level reference resolution [recognizing linguistic context and grouping words that refer to the same entity]. I was like, shit. There are people who have devoted their whole lives to solving this problem. It was well-known as a really difficult problem in NLP that people couldn’t touch and go above maybe 70% or whatever. And just like that, this thing that had no special structure for it, was doing that reasoning. And I was like, oh, wait, wait, wait.
I think I’ve just flipped and I’ve become a believer, and now I’m a dogmatic. So I think it is really, really, really incredible. It’s really cool.
“But I think on the general AI front, what’s about to happen is an avalanche of innovation. You know, models are getting smaller and — the proof is there. It exists.”
Now, you know, inevitably in two years there will be much smaller models, faster models that could do similar things. I’m really excited. You? How do you feel?
RK: Literally the same. I always say OpenAI gave us the proof of technology. They showed that this can work. We actually — there’s this long story of how we were started. We were literally doing the Shakespearean experiments. Actually we were doing it in Finnish, like the epic of Finnish, and it was writing poetry and it didn’t make any sense, but it looked correct. And this is cool, but then you’re like, ‘Hey, I can’t actually use this because it’s nonsensical, but it looks cool.’ But we got quite close to ‘Wow, it’s actually quite accurate.’
“I still think we’re kind of 90% there. The last 10% is always going to be the most painful, to really put this into production grade — I'm talking about letting an LLM talk to the end consumers and run long dialogues.”
But it’s also exciting how quickly it’s moving. It’s inspiring. I feel like everybody was so depressed last year, it was all doom and gloom. But then suddenly it’s like, ‘Hey, innovation is in startups.’
And then what I’m really excited about is the open-source community showing companies like Google, who their leadership is like, ‘Oh, this is hard, expensive, gonna take us years,’ and then a couple of guys in a basement are like, ‘Hey, look, it works on an iPhone.’
I think that’s cool. I think it’s very cool. So, you mentioned a little bit about the future, but what do you see, from your point of view, what’s going to happen in support teams? It is now an acceleration of the vision that you were pushing forward of automation, or what will support teams look like in 12 months?
NM: I think in 12 months you’re not going to see all that much in terms of things being uprooted. What we’ve seen for the past — well, since Covid really — is that agent attrition is huge. People are moving out of these jobs. They don’t have to do them, they’re not well compensated. They take a lot of abuse, right?
And I’ve always wondered when I shadowed call center agents — and I think especially with voice rather than text; people can get abusive over text as well, but a lot more emotion goes through voice — I’ve always wondered how people can even stand their jobs. Because often you’ll get someone who gets call after call, with nervous, angry, customers yelling at them, when all the agent is trying to do is help.
So I think that attrition is going to continue. And I think increasingly people are using technology to get better SLAs, to get a better customer experience, and they increasingly have to use it. So, I don't know if it’s in 12 months, but I think really what we’re moving towards is a contact center that’s probably a fifth of the size in staff, but the staff are better compensated, happier.
It’s basically your current management and tier two agents. People who actually have the grit and the emotional intelligence and desire to do that kind of job. They’ll be able to program assistants — I think over both voice and text — that will continuously get better. And they’ll be there as escalation points when an edge case happens. I think they’ll be happier. I think they will be knowledge workers. I think they will be better compensated and effectively they will slowly transition into being engineers.
RK: We actually get asked, are ChatGPT and LLMs existential to us? Mostly VCs ask this, but what about you? Is it existential or is it an enabler? Well, I know the answer.
NM: I think that with voice in particular, what’s interesting is these things are slow, right? And they take a long time. So if you were using GPT-4 now to power a voice assistant, if you had good streaming et cetera — so if everything goes according to plan — you’re still looking at on average a four second delay and 20 cents per term, right? Now, if you think about an average conversation, you go back and forth like, five times in a minute, four times, right? So you’re looking at a dollar a minute.
That’s at the upper range of the most expensive agents in the US. So you know, Moore’s law will work, and over time things will get cheaper, but right now it’s an expensive choice. So whoever’s thinking about doing it should probably think twice. Especially if they think, ‘Oh, it’s going to get even better.’ Well, guess what? It’s going to get more expensive.
RK: Yeah. And I think one thing you mentioned earlier is: First of all, the upskilling of support teams.
“I think the whole support organization is going through a huge transformation in a positive way. And also then having that kind of permission from the leadership.”
Because honestly, a lot of support teams in big organizations don’t have a lot of power compared to, let’s say sales teams or marketing teams, although it’s almost backwards because they should have a lot of power because they’re the ones integrated with the customers. But I feel like now, they are the ones actually adopting the latest and greatest of the technology and becoming the experts in it. So I think there’s going to be this flip in many big organizations, where the support is the spearhead.
NM: Listen, we’ve had dozens of buyers who started with us, you know, almost as a side project, something interesting, because they were — as you said — early adopters. These people are now a few levels up from where they were and they are making transformational changes to entire companies.
They’re transformation directors, right? And they’re doing a lot of work now across areas outside of customer service. So I think it’s a really good thing that upskilling starts at the top of the customer service organization. I think the transformation of the contact center itself will take a bit longer.
And you know, the one thing that we can’t do that much about yet is overall digitalization, right? You know, an airline that’s stuck with green screen support terminals — it’s really hard for them to benefit from this. So I think now they have to either not survive as companies, or invest into getting better tech stacks so they can then reap the benefits of AI.
RK: The pressure is on, the FOMO is real. So I think that’s the good thing about this market. You’re very excited. I’m very excited. We’re both kind of in the eye of the tornado in a way — in a good way. But is there anything for the future that scares you when it comes to this technology or the shift we’re going through?
NM: Yeah, I mean, I no longer trust any long-term projections that I have, so I’ll start with that. I think the risk is to just… You know, whatever happens in a technological revolution. The economy will transform rapidly and that will send shock waves throughout societies. And we have to be careful about that, you know. Revolutions happen in these moments and it could happen again.
Now that’s doomsday speak, but really it is kind of what happens. People change where they live, like with Covid we saw a push out of the cities for the first time ever after the industrial revolution. I think with this stuff, if you were studying to be a lawyer right now, the best of lawyers will probably do even better. But it’s going to get more competitive. Right.
And a lot of those jobs that people thought of as safe, relative to automation, they’re starting to feel like the most unsafe jobs. Because actually it’s just very highly specialized knowledge that seems to be a lot more automatable than the work of someone who, I don't know, does plumbing or complex manual labor. Because we’re actually moving slower on robotics than we are in knowledge work, which is very easily accessible through textbooks and more advanced software.
I think it’s gonna get really interesting and I think that people have to funnel their creativity in productive ways. So I think that is the risk. I think, you know — overall killer robots and stuff — I think we should adopt a bit more of a trans-humanist outlook on it. Because really, honestly, even if you look at it religiously — this thing is of us.
“People talk about bias as if it’s induced by an evil terminator-like sentient thing that means us harm. But it is us in our raw, terrible form that creates the things that we don’t like about the AI.”
RK: I agree. I think with everything we have to be careful, like whether it’s the first hammer invented. You can do a lot of damage with that, but also good stuff with that. I think that’s kind of a cliched anecdote, but also it’s really a showcase of us humans doing what we as a species are good at doing.
We build new technologies to improve our lives, and that’s the goal. I don’t think there’s a lot of people out there that — of course there’s always bad actors — that actually have some evil plan to take over the world with AI. People are actually genuinely building these technologies because they think they produce value for the whole of society and as a species.
And that’s what we’ve been doing since the beginning and I think now there might be a big leap. But I also think that — I think Bill Gates said, we all really overestimate the next 12 months, but underestimate the 10 years. But I think it’s exciting. Overall it’s a big net positive.
NM: Oh, for sure. I think so too. I just think we have to be careful with being honest, you know, about things like UBI [universal basic income]. I feel like people stopped talking about it, relative to maybe a few years ago. And I think politically we should think about organizing society in a way where people are protected.
RK: Maybe I’ll tie back before we finish off on the world of support — can you give some advice for the companies, the buyers, or the support leaders, or whoever is looking at these technologies where they hear about ChatGPT they hear about what’s going on in the markets, it’s overwhelming. How can you advise them, how to think about this and their support organization, when it comes to implementing and bringing in AI?
NM: Small steps, but a quick succession of small steps. I think a lot of people try to boil the ocean. And they get into a discussion of ‘What do I need?’ And these are good things to think about, but in the early parts of buying a new technology, you’re mostly investing in your own education.
RK: And I think that’s key when it comes to these new technologies because no one knows — during the sales cycle we look at your data, we analyze, okay, here’s what’s projected, I think this is what we can do.
“But in the end, it’s your environment, your customers, your processes. There always might be something unexpected — and that’s when the partnership comes in, not the technology. And I think that’s key.”
I think to end with this momentum we are feeling, I want to thank you Nikola for joining us, and thank you for the partnership. It’s only a few months old, but we’re already on a roll. There’s a lot happening together.
NM: No, thank you. Thank you for having me. Really excited about the partnership. You know, we spend a lot of time telling people that generic omnichannel is going to lead to disappointment on both fronts. So, I think when you and I first spoke, what you said, that you could do a search and replace, voice with text and text with voice, and it's literally what I've been telling people. So I think we saw eye-to-eye very, very quickly. So I'm really excited about helping spread it across both customer bases and, you know, new customers as well.
RK: Exactly. Well, thank you so much. Enjoy your day in the States and then the weekend hopefully. So thank you, and onwards.
NM: Awesome. Thank you, Reetu.