French startup round Believing that AI voice agents are going to become the default way for customers to interact with companies, so instead of building ready-to-use AI voice agents, the company is building an orchestration platform that allows companies to build their own voice agents.
Rounded began working on a Web3 product in June 2023 before shifting its focus to exploring AI voice agents. “The idea was that we would just put ChatGPT after a transcriber and before a synthesizer and it would be worth it,” co-founder Aymeric Vaudelin (pictured above, first from left) told TechCrunch.
But the team soon encountered the usual product-market fit conundrum. “After a few months, we realized that the market was not yet ready to listen to voice agents. So we built a product, and packaged everything to create the first agent,” added Vaudelin.
The result of that effort DonnaAn AI voice agent for anesthesiologists. Although it seems a bit random, the startup chose that market because anesthesia secretaries have to deal with a large number of patients and it's usually a very transactional experience.
In France, when you schedule a surgery, you need to talk to the anesthetist beforehand so they can make sure you don't have any allergies or potential complications to the anesthetic products.
Anesthesia secretaries have to handle a large volume of calls which is quite straightforward. Generally, people simply want to know when an anesthetist is available, schedule an appointment, or change a date.
Moreover, these are not sales calls, so an AI agent does not need to be persuasive or highly skilled. “In the early days, we were struggling with delays of sometimes 4, 5, 6 seconds,” Vadelin said.
Still, with Donna, Rounded has managed to convince 15 private hospitals to have an AI voice agent answer calls, and the company says the agent has handled tens of thousands of conversations so far. Over time, Rounded has improved its product so that it is more accurate, better integrated with other products, and importantly, faster.
“For example, with a web call, we now get latencies of less than 700 milliseconds – like 600 milliseconds. You add 200 milliseconds for the phone connection, more or less,” Vadelin said.
More recently, Rounded expanded its remit with an orchestration product that other companies can use to build their own voice agents.
Rounded lets you pick and choose your off-the-shelf AI models, which can be a speech-to-text model, an LLM, and a text-to-speech model. For example, for your first voice agent, you can use Azure to transcribe calls, GPT-4o Mini as LLM, and ElevenLabs as your speech-to-text engine.
The platform then helps define the instruction tree and prompts that will make the LLM work for your specific use case. “Implementing an agent means finding the right prompt, the right parameters, and the right variables in the prompt,” Vadelin said.
“Our pitch is that we'll enable everyone to build great prompts or great agents, and our product will support them in this iterative process to build super-reliable agents,” Vadelin said.
Rounded has raised €600,000 (about $620,000) so far from UC Berkeley's DeepTech accelerator. SkydeckAnd several business angels. But as artificial intelligence remains a very busy industry, I'm sure startups will raise more money in the coming months.