Melbourne-founded voice AI startup Phonely removes annoying pauses, sees revenue dial up

Melbourne-founded Y Combinator (YC) alumnus Phonely has seen revenue move "up and to the right" as its voice AI technology is adopted by more clients in the USA and Australia, including Perth-headquartered contact centre consulting group TSA Group which is reportedly already saving 12,000 hours a year.

Now based in California but with one of its founders Nisal Ranasinghe still in Melbourne, Phonely has finetuned its offering so that AI agents can be used for the majority of call volumes for many businesses, following pre-defined scripts around tasks such as appointment scheduling and lead qualification.

And with advancements in its AI training models, bolstered by a collaboration with AI fast inference provider Groq and LLM builder Maitai, Phonely co-founder and CEO Will Bodewes says the technology has overcome the traditional "big pause" associated with non-human phone operators.

"From our use case, around 70 per cent of people don’t realise they're not talking to a person when they call in, unless they're prompted by an AI," Bodewes tells Business News Australia.

"The goal for us is to get to 95-99 per cent of people not being able to distinguish that this is an AI or a person, in the best way possible, not because we're trying to trick people - only because we want people to have a good experience with the support that they're given by the technology."

He says one of the dead giveaways that you're talking to a voice AI is the lengthy pause between when you finish talking and you hear a response.

"A lot of times businesses when somebody hears about voice AI they think of this old IVR (interactive voice response) system that we all hate, including me," he says.

A thought for pause

This occurs because a model is trying to predict when someone is done speaking and there are numerous methods to make this calculation, all with their different flaws.

"One of them is you can just wait half a second to see if they’re done talking, but then that adds half a second latency to the response so that doesn't really work well," he says.

"The next thing you can do is look at when a period (full stop) is predicted from a transcription model, but then it will jump in randomly in the middle of a sentence."

Phonely's approach was to train a custom model that analyses transcriptions while "figuring out when someone's done talking without adding additional latency".

"State-of-the-art transcriptions are about 300-700 milliseconds of latency that we add into there. We did some intelligent routing on this and deployed it locally in Australia where we’ve got our own speech-to-text system, and that saves on latency," he says.

"The biggest piece in the middle is the logic piece, which is the large language model (LLM) in our case. That was the partnership with Maitai and Groq to get us set up and actually able to accomplish the real-time, human-like responses where we can get 300-400 millisecond response times out of on AI model."

Bodewes says this was needed because using larger models like those offered by OpenAI's ChatGPT and Anthropic's Claude have very high average latency and latency variance.

"OpenAI has this huge queueing problem where they get hit by a tonne of requests at any given time, and so they're constantly putting your request into a different queue, and sometimes it’ll take four seconds for that to come back. And when you’re on the phone, four seconds is a lifetime to get a response," Bodewes says.

"We simulated some data around a flow that we’d built in Phonely, and we were able to finetune a model specifically customers that was able to outperform OpenAI in terms of latency, because it was a lighter, smaller model.

"The reason Groq is valuable on this is because their custom chip set is designed for running language models really quickly and cheaply."

Model refinement post-YC

This all represents significant progress on the demos that got Phonely into Y Combinator in the first place, attracting an initial US$500,000 ($749,720) investment from the prestigious Silicon Valley accelerator, but more importantly the education and networks that the program engenders. Since then the startup has gone on to raise a further US$2.2 million ($3.4 million) and employs 12 staff.

The co-founder says that in those early days the company had demos up and running with a couple of small businesses that were deploying Phonely after hours.

"What we’ve done for the past year is we've just made it work consistently for a lot of different organisations, and we’ve built this robust platform around it to take what we had at the beginning – the demo – and fortify it so that it would work in every use case," says Bodewes, who started the business with Ranasinghe in 2023 when they were PhD researchers in Professor Saman Halgamuge's AI research group at the University of Melbourne.

"We’ve spent a lot of time making it easy for businesses to be able to not have to learn how to prompt engineer...we've also built a lot of external connections and technology on top of it. For example, you can integrate directly with Google Calendar.

"One of our customers is doing 60,000 calls a day with our platform, whereas in the beginning it was like 100 calls a day."

Phonely's users include more than 3,500 people who have signed up for its free plan, as well as a growing clientele of paid enterprise customers.

He says the technology excels in use cases where the pre-set script and call flow wouldn't require too much back and forth between the AI model and the internet.

"Every time we have to build a connection with the internet we have to make an API (application programing interface) request, and we just have to build that out," the co-founder syas.

"That plumbing, that takes a lot of time, but a lot of the industries we’re seeing big success in and where we can get onboarded in literally a matter of weeks are businesses that are qualifying for a certain type of customer and then scheduling them into an appointment.

"What we’re seeing in our business is it’s knocking out maybe 80 to 90 per cent of call volume, and then we have that human to do more complicated tasks - the things that might require more knowledge."

Pioneering Australian partnership in voice AI

Bodewes is bound by confidentiality agreements with many customers, but one partner he is able to reveal is TSA Group, and Australian customer experience (CX) specialist with 4,500 employees which works with large organisations such as airlines, as well as government, on their contact centre solutions.

"We’ve been working with them for about a year, and we’ve just set up a locally deployed Australian environment," he says.

"It's the first one that's been set up, so it means that all the data stays in Australia – we’re officially partnering with them to take AI solutions to Australia for these larger organisations.

"They do the data management, build out the flows on the platform, making sure it’s all enterprise ready, and drive a lot of analytics.

He notes that TSA has observed interest from brands in voice AI as they recognise the significant cost and time savings it can bring. 

"They’ve got a lot of value out of it. They were saying it was saving them 12,000 hours a year just by implementing the Phonely system right now, and this is just the beginning. It’s going to save them significantly more," he says.

"There are also use cases as well where the human is the best answer, where you need to escalate a call and that needs to go to a person," he clarifies.

"A really interesting problem out there as well is speed to lead. That's what we call it, and it's a very critical thing for these businesses."

Reflecting on the Y Combinator experience, Bodewes says a lot of the value came from the network experienced tech founders and individuals you can lean on for guidance or introductions, as well as the competitive environment and networking benefits of being in a cohort of other founders.

"I think a lot of the value that I got out of Y Combinator was the ability to be surrounded by 500 other people that were really excited about building something and really passionate about working crazy long hours. For everyone there, working 16-hour days was pretty much normal," he says. 

"A lot of people that are interested in working at a startup would love to work in a YC startup, and so we get access to really good talent from that.

"It’s been growing really fast. When we got out of YC I think we were in the top 5 per cent of companies in the batch for revenue, and we’ve continued to have this ‘up and to the right’ growth path as customers have scaled with us, we’ve gotten new clients and made really great decisions around partnerships and people that we decided to work with."