London-based PolyAI, a Cambridge University spin-out led by dialogue-system veterans, has built technology so natural that some callers never realise they’re speaking to a machine.
Press 1 for sales, press 2 for support, press 3 to abandon all hope.
For decades that joke has summed up the frustration of calling a contact centre equipped with sub-standard technology.
But a new generation of AI-powered voice agents is turning the joke on its head. In the process, it is helping some of the world’s major brands – Metrobank, Whitbread, Pacific Gas & Electric, Unicredit, for example – to pick up every phone call, and customers to book tables or cancel subscriptions in seconds.
This is freeing up human agents to handle the problems that truly need the personal touch.
John Murphy, Director of Customer Service at French digital services company Atos, said: “PolyAI has allowed us to accommodate periods of peak customer activity and to minimize complaints by significantly increasing our ability to answer Frequently Asked Questions through the voice assistant.”
And Dan Eddie, Customer Service Director at Simplyhealth, another PolyAI customer, said: “This is next-level conversational AI, and it’s far-and-away better than anything else I’ve heard in the market.”
From research to prime-time
PolyAI’s origins lie in the Machine Intelligence Lab at the University of Cambridge, where co-founders Nikola Mrkšić, Tsung-Hsien Wen and Pei-Hao Su were completing PhDs when Google first published its revolutionary transformer architecture in 2017 and started the Large Language Model [LLM] race.
“They’ve had a front-row seat at the development of that technology since the very first transformer-based models,” explains Michael Chen, PolyAI’s Vice-President of Strategic Alliances and Corporate Development.
Chen joined shortly afterwards, when the company was a 30-strong London start-up.
Today it employs almost 300 people across the UK, US, Serbia, Canada and the Philippines, and has raised more than $120 million in investment to scale-up its vision.
Voice first
While most conversational-AI firms began with text-based chatbots and tacked on voice later, PolyAI chose the opposite path.
“We’re a much more integrated player – we have our own speech-recognition engine and we also have our own LLMs, bridging the gap between hearing and reasoning at production scale,” Chen explains.
Rather than chasing ever-larger models that can solve quantum chemistry and make up silly limericks, PolyAI tunes its LLMs for the controlled, predictable performance businesses demand.
“For customer-service conversations, the reasoning ability of LLMs is already intelligent enough to perform customer service. It’s now about control, reliability, getting the LLM to follow an exact brand experience,” Chen says.
Fine-tuning allows the model to follow an exact dialogue style, while PolyAI’s proprietary speech-recognition system can swap between domain-specific vocabularies – UK postcodes in one moment, US Social Security numbers the next – mid-conversation.
Zero-wait service
In production, the effect is dramatic. Because voice agents answer every call the instant it arrives, there is, as Chen puts it, “no wait time anymore.” The mantra inside the company, he adds, is equally bold: “‘Hold times are for the old times.’”
That elasticity matters because voice is uniquely hard to staff: one human can juggle 10 chat windows yet only a single phone call.
“Even after years of ‘digital transformation’ there are still massive phone-based support operations around the world, and every human being can handle only one call at a time,” Chen notes.
By absorbing routine tasks like password resets, PolyAI’s agents free human staff to focus on higher-value moments that require empathy and emotional intelligence – cancelling an account for a recently deceased family member, say, or escalating a complaint to senior management.
Better not less
While some observers worry that automation inevitably leads to redundancies, Chen sees a different outcome. Contact centre staff turnover can exceed 30% a year and many businesses are currently understaffed as a result, he says.
“If AI removes the repetitive, low-value tasks, you leave room for humans to build real careers in customer service,” he argues.
The goal, then, is not to replace human agents but to elevate them, turning a reactive cost line into a proactive relationship engine where people handle complex, loyalty-building interactions and the AI takes care of the mundane – such as handling in a personable manner those common questions buried in a company’s online FAQ. This can even improve Net Promoter Score (NPS) for clients along the way.
But successful automation, Chen insists, is as much about identity as efficiency.
“We think of ourselves as custodians of a brand experience, not a cost-saving measure,” he says, describing how PolyAI works with each client to craft an agent persona that reflects company values and customer demographics.
Dial a rural pub in Yorkshire and an agent can greet you in a gentle local lilt, for example; phone a global bank and the tone shifts to one of professional reassurance.
Building on Microsoft Azure
PolyAI recently joined the Microsoft Partner Network and is now available on Microsoft’s cloud platform, Azure. The move is pragmatic as well as technical.
“Being on Azure allows us to work seamlessly with enterprises in regulated sectors like government and healthcare,” Chen explains.
The company has already built an integration with Microsoft Dynamics 365 Contact Centre, enabling voice agents to push real-time context into a customer-service dashboard. If an issue escalates, the human agent sees not only the call reason but also every data point the AI has gathered so far.
PolyAI is also exploring Microsoft Teams, which is evolving from collaboration hub to full telephony platform.
“If you’re using Teams to manage reception desks or storefronts, you probably also want an AI agent fronting those calls,” Chen says, highlighting a future where missed calls – and missed revenue – become relics, too.
From frustration to fluency
PolyAI’s agents already handle many of the calls they receive without human intervention, and internal tests show customer-satisfaction scores climbing as hold music disappears. That, Chen says, is precisely the point.
When voice bots stop sounding robotic and start being genuinely helpful, the phone call reclaims its place as the most direct, reassuring channel a brand can offer, he believes.
Soon, the simple act of picking up the phone could once again become a pleasure, and the phrase “please hold” consigned to history.