Build AI Voice Agents | Rounded

The Rise of AI Voice Agents and Rounded's Orchestration Platform

French company Rounded anticipates a future where interacting with businesses via AI voice agents will become commonplace. Rather than developing pre-built agents, they are concentrating on an orchestration platform empowering companies to construct customized voice solutions.

From Web3 to Voice AI

Rounded initially explored the web3 space before redirecting its efforts towards AI voice agents in June 2023. Co-founder Aymeric Vaudelin explained their early concept involved connecting a transcription service with ChatGPT and a speech synthesizer, believing this combination would deliver significant value.

However, the team quickly encountered the challenge of achieving product-market fit. After several months, it became clear the market wasn’t yet receptive to voice agent technology. This led to the creation of a complete product, effectively a packaged, initial agent.

Donna: An AI Agent for Anesthetists

This initial product materialized as Donna, an AI voice agent specifically designed for anesthetists. The startup strategically chose this niche due to the high volume of patient interactions handled by anesthesia secretaries, which are typically routine and transactional.

In France, a pre-operative consultation with the anesthetist is mandatory to identify potential allergies or complications related to anesthesia.

These secretaries manage a substantial number of calls focused on straightforward requests. Patients generally inquire about the anesthetist’s availability, schedule appointments, or request date changes.

Crucially, these interactions aren’t sales-oriented, meaning the AI agent doesn’t require persuasive skills or peak efficiency. Vaudelin noted early latency issues, sometimes reaching 4 to 6 seconds, were a concern.

Early Success and Platform Evolution

Despite these initial challenges, Rounded successfully convinced 15 private hospitals to implement Donna for call answering. The agent has since processed hundreds of thousands of conversations. Continuous improvements focused on accuracy, integration with existing systems, and, importantly, speed.

Current web call latency is now under 700 milliseconds, often closer to 600 milliseconds. Adding approximately 200 milliseconds for phone connectivity completes the process.

Introducing the Orchestration Platform

Rounded has expanded its offerings with an orchestration platform, enabling other businesses to develop their own AI voice agents.

The platform allows users to select from a range of pre-built AI models, including speech-to-text, Large Language Models (LLMs), and text-to-speech engines. For example, a voice agent could utilize Azure for transcription, GPT-4o mini as the LLM, and ElevenLabs for speech synthesis.

The platform also assists in defining the instruction trees and prompts necessary to tailor the LLM to specific use cases. Vaudelin emphasized that successful agent implementation hinges on identifying the correct prompting, parameters, and variables.

“Our core value proposition is empowering everyone to create effective prompts and agents, and our product will support them throughout the iterative process of building highly reliable agents,” Vaudelin stated.

Funding and Future Outlook

Rounded has secured €600,000 (approximately $620,000) in funding from UC Berkeley’s SkyDeck accelerator and various angel investors. Given the current enthusiasm surrounding artificial intelligence, further funding rounds are anticipated in the near future.

Stay informed with TechCrunch's AI newsletter! Subscribe here to receive it weekly in your inbox.