When you call a company that uses Poly AI, the technology isn’t the first thing you notice. It’s the lack of friction. You won’t hear a robotic voice with a clip telling you to “press one for billing.” After saying something off-script, don’t pause for too long. The voice on the other end simply responds, sounding like someone sitting in a quiet office. It may seem insignificant, but anyone who has shouted “REPRESENTATIVE” into a phone for thirty minutes will attest that it is anything but.
A group of researchers from the University of Cambridge founded Poly AI in 2017. This is the kind of origin story that typically results in a paper that no one reads. Rather, they ended up creating a system that currently manages millions of consumer calls. In May 2024, the company raised $50 million in a Series C round, surpassing $120 million in total funding. Investors appear to think that this is the area of AI that will yield the greatest returns before the more ostentatious aspects catch up.
Their method is intriguing because they didn’t add voice to a chatbot. First, they developed the dialog agent and considered voice to be its natural habitat. Over a billion business conversations were used to train the Raven model. It’s a unique training set. The majority of language models incorporate the internet. This one absorbed contact centers and all the mess that goes along with them, such as irate clients, mumbled account numbers, and people who cut themselves off in the middle of a sentence.
In strangely specific ways, the results speak for themselves. According to the marketing director of Fogo de Chão, the system is expected to increase revenue by about $7 million. One voice agent reportedly earned over a million dollars for a hotel and casino brand. A retail clothing company saved a million dollars by reducing seasonal hiring. These are not the kinds of numbers that are typically associated with chatbots, which in the past have offered mild annoyance and transformation.
As this field develops, it seems that intelligence isn’t the true litmus test for voice AI. Tolerance is what it is. Will a client allow it to complete a sentence? When they hear it, will they immediately hang up? Poly AI won because the voices sounded real, according to Dan Weinberger of Audibel. The healthcare company Howard Brown Health’s Lauren Sullivan expressed surprise at how natural it felt. These are not marketing compliments, but rather useful ones. Gushing is not common among those working in operations.
Naturally, you can’t sign up for the platform for free in five minutes. With deployments that take about six weeks, Poly AI sells to businesses on a custom per-minute basis. It has built-in SOC 2, HIPAA, GDPR, and PCI DSS compliance, which is important when the agent is managing insurance disputes or fraud cases in multiple languages. Platforms like Voiceflow, which are more adaptable for teams without seven-figure budgets, are frequently used by smaller businesses searching for less expensive, self-serve options.
Whether voice AI will eventually completely replace human agents or only handle the tedious middle layer of calls is still up in the air. According to reports, a large percentage of calls are resolved by a global delivery company using Poly AI without ever escalating to a human. Depending on which side of the headset you’re on, that could be a warning or a glimpse of what lies ahead.
It’s difficult to ignore how carefree the entire experience seems. There is no longer any hold music. The reading of the script has stopped. Whatever AI develops in the future may begin with something this subdued, this commonplace, and this oddly similar to a real conversation.
