The PolyAI office might well remind you of a thousand tech startups. There’s spacious open-plan seating, desk beers are a regular event, and naturally, a friendly dog joins us for our meeting. Yet, unlike so many tech companies that believe they have a killer idea, only to discover it’s been done and discarded a thousand times before, what the PolyAI team is aiming for could genuinely disrupt the customer service industry that so many of us interact with.
PolyAI’s ace card is its “conversational AI” approach to customer service – an advanced AI system that can genuinely understand conversational patterns from the human side, abandoning the stuttering decision-tree process of typical AI chat or call services. PolyAI's laser-focus on the customer service applications of conversational artificial intelligence technology is what's giving the company such a huge leg-up over competitors.
It’s not just PolyAI that believes in the potential of its product. The company recently won $12m of funding from investors to help it realize the first step of its call center AI platform plan.
We sat down with Stefan van der Fluit, Business Development Manager at PolyAI, Kylie Whitehead, Content Manager, and Fig, dog, to talk about what exactly PolyAI is, and what it’s striving to do. We’re later joined by founder and CEO Nikola Mrkšić, a former Cambridge graduate, who had previously spent two years at Apple, working as a machine-learning researcher on Siri. Nikola flits between our conversation, his phone and his laptop, all at once, periodically contributing his thoughts on the industry with careful consideration.
The problem with customer service AI
AI-driven customer service has spread quickly in recent years. Perhaps, too quickly – think strained attempts at dealing with an automated voice on the phone, or typing in anger at the dreaded chatbot.
Stefan blames poor examples of AI for coloring the public’s perception of just how fluid conversational AI can be, and what it’s capable of achieving. It’s a dangerous PR problem, and one that PolyAI is keen to redress.
The issue with existing conversational AI, Stefan tells us, is that it’s too reliant on trying to predict conversations, rather than reacting to them. It’s not a natural way to converse, and it makes for a disappointing experience, especially when queries become complicated.
With systems like Siri and Alexa, human teams pour over transcripts and try to identify where the systems are going wrong. PolyAI, however, relies more heavily on sophisticated machine learning.
What is PolyAI doing differently?
“Imagine a magnet”, says Stefan. “It hovers above a high dimensional space, plucking out the answer that is the most relevant to the question. It’s about getting from the question to the answer in the fewest steps. We’ve based PolyAI on a machine learning model that’s based on reranking, rather than a predictive model.”
Instead of trying to architect conversations, its goal is to orchestrate conversations. “It’s just a question of looking at the equation differently,” says Stefan.
The “basic” model for PolyAI’s platform has been trained on a billion conversations at this point. Now, the team argue that it’s ready-made for customer service industries. “Even if companies give us their own data points, the improvement is only slight, as we’re already at such a good point, meaning the conversational model is industry agnostic” argues Stefan.
The reason for PolyAI’s different approach is that its founders come from academia, not business. All carried out research at Cambridge Dialogue Groups, under Steve Young, a man who Stefan calls “the Stephen Hawking of conversational AI”.
The tech driving the platform is IP protected, so the company isn’t worried about one of the tech giants overtaking them suddenly. If Google were to train a billion conversations, it wouldn’t mean much without PolyAI’s back-end tech, Stefan tells us.
Robots dealing with illogical humans
When Spock called us illogical, he was onto something. We’re strange creatures, and although we think we’re au fait with following processes, to an AI, we’re all over the place.
That’s bad news, as it means that we’re well set up to confuse our robot companions. This is no truer than on a telephone call, when conversation can easily flit from one topic to the next, sometimes without warning or context.
PolyAI has been working hard in this area to make its system as natural as possible to handle queries and respond in the same way a human agent would.
In a demo example, PolyAI shows a diner calling to book a restaurant. At the end, the caller asks how she can get there. In this context, “there” is the restaurant, and the system is clever enough to realize this. It’s also able to tackle questions about vegan food options without a hitch, understanding that you’re after menu items from the restaurant you’ve previously been speaking about.
That may sound simple, but think back to the last time you tried barking a question at Siri, followed by another one. Most systems would trip up if you didn’t ask a question with a full beginning-and-end context.
In customer service cases, this could send the caller down an unsatisfying path of further questions and clarification.
Is conversational AI suitable for all interactions?
This was the main thought on my mind on the way to PolyAI’s offices. There’s no doubt that its system is slick and impressive, but surely there are situations where it’s not appropriate? In an interesting twist, PolyAI is quick to agree with me.
“We feel that the system isn’t ideal for browsing or shopping, for one”, Stefan reveals. “These experiences are much better served with a GUI, and all that happens is that the user is read out a large scrollable page”.
I raise the question of the emergency services. Could an AI match the compassion needed in this industry? I’m told no. Actually, what Nikola says is “No, not yet”, meaning that it needn’t be totally off the table in the future.
However, there is an argument, Stefan says, for making the first part of the call AI-led. It’s purely transactional after all – which service does the person require? This could cut down on call time and lead to quicker responses when a human operator takes over for the full emergency call.
The future of jobs in the call center
Then we reach the elephant in the room. Jobs.
Everything PolyAI is planning sounds great for companies and consumers, but its implications for those currently working in call centers feel troubling.
With the rise of AI and a potential technological revolution on the doorstep, are those working in customer service centers soon to be a dying breed? According to PolyAI, the answer is yes, and no.
“We believe that jobs will get more interesting”, Stefan tells me, when I bring up the potential upset the technology could cause. “If you go to a call center – and we’ve been to many – the best ones are indistinguishable from an eBay or Paypal campus. People take pride in their work, and pick up every call as if it’s the first of the day. What we can do, is remove the transactional calls, and let them focus on the more interesting ones”.
PolyAI points towards the time and money spent on paying people to answer rudimentary, repetitive queries – the ones that would be best placed for an AI answer. Calls to chase missing orders are necessary, but not challenging for a human call handler. The reallocation of these calls to an AI, which doesn’t mind boring jobs, can only be healthy for the company.
“Yes, there will be fewer people in the call center,” Stefan follows up, “but they’ll be better paid.” In theory, the requirement for staff able to handle more sensitive or complex queries could push wages up. “Let's say you have a million dollars set aside for customer service staff. With AI, you need fewer of them, and can pay your human employees more”.
PolyAI also argue that while call centers will need fewer call center staff, AI will see a net increase in jobs overall. All the data and training that AI’s require doesn’t appear from nowhere.
The example used is Amazon’s Mechanical Turk, a platform that allows pretty much anyone to sign up and make money, by crowdsourcing important data collection, analysis and clean up.
Privacy and data protection
The conversation naturally turns to security and privacy. It’s big news in the tech space, with large corporations regularly being hauled over the coals for breaches, and where losing consumer confidence could lose you your customers.
PolyAi believes that talking to an AI could actually improve users assurance that their data is secure. They use the example of buying something from an adult store online. If there’s a problem with your order not turning up, you’ll have to contact customer services. Would you feel happier talking to a human about the issue, or an AI – a collection of ones and zeroes whose only job is to process your call and nothing else?
There’s also huge scope in the medical field. Callers are likely to be more relaxed talking about embarrassing health conditions to a non-human.
It’s not just potentially embarrassing information either. Stefan makes a case for human employers transferring callers to an AI for a payment transaction portion of a call, to set their mind at ease. With no human element, there’s no fear of anyone writing down your details or misusing them.
Handled the right way, and with clear permissions, the company feels that AI could enhance the user experience, if given access to the right information. With an associated profile, AI already knows what services and products the caller has, as well as their interests, personal details and anything they’ve happily shared.
This could make for a more personalized service when on a call, with the effect being two-fold. You get a happier customer who feels understood, plus get the chance to sell them the right product.
It’s important, though, that callers know they’re talking to an AI in the first place. We talk about the industry’s dabbling with live chat, and how it rarely comes clean about the ‘person’ behind the keyboard.
Does John, the company representative you’re talking to on a live chat actually exist? Is he a person in an office with a lunch break, who gets in his car at the end of the day, navigates the commuter traffic, before unwinding with Netflix before bed? Or is he, in fact, a programmed AI just masquerading as a person?
Last year, for example, Google unveiled its Duplex AI service. It drew an instant backlash for the perceived deceptiveness of an AI voice assistant “umming” and “ahhing” to imitate a human caller.
Stefan tells me that his company doesn’t want to deceive callers, and that there are lessons to be learned from chatbots. He suggests that callers are likely to be more receptive when told upfront that they’re communicating with an AI.
That’s especially true if, in the short term, you tell a customer that they can talk to a human employee if they prefer. He sees it as the perfect opportunity to impress the public with what AI can do – a great advertisement for PolyAI’s tech.
The big wins for PolyAI
It’s clear from talking to the people at PolyAI, they have a huge faith in what they’re doing, and there’s a sense of excitement and progress in the air that can’t help but bubble up in conversation. It’s contagious too, and it’s hard not to get drawn into the dream that PolyAI is selling.
As we’re wrapping up, I ask Nikola what he sees as the biggest win the company has had. Initially he struggles to answer – not because there aren’t enough to mention, but because the pace is so fast that everyday is a new discovery and milestone. Eventually he settles on an answer. “It was when we stopped talking about conversational AI as a concept, and settled on customer services. This gave us a focus that helped us define exactly what PolyAI was and what we wanted it to be.”
In this way, the company’s laser focus on one industry (even if it’s not, as they tell me, the most immediately glamorous), has allowed PolyAI to concentrate its energy into one specific solution. The company is working hard at generating an innovative product and keeping ahead of its competitors, whose more scattergun approach to conversational AI has seen a much wider scope, but arguably more limited success.
PolyAI is a company that is only a couple of chapters into its story. But, if even a small number of its lofty goals are realized, the business could rewrite the customer services rulebook for all of us.