Several years ago, when handwriting recognition was the rage, I asked my company’s chief scientist, John, whether it would ever be accurate enough. He immediately said no — not unless you reteach people how to write because human handwriting varied too much (with the possible exception of Chinese handwriting, which tends to be more uniform).
But he added that voice would eventually be accurate enough. He was right. While the PalmPilot struggled to reteach us how to write, modern tools such as Dragon have achieved up to 99% accuracy with clear diction and high-quality audio. However, what happens when sophisticated voice technology is applied in less-than-ideal conditions — on phones used by distracted callers, with inconsistent audio quality and in more complex business processes? This question matters because 70% of agent interactions at businesses will occur via voice — where issues are often urgent and directly shape the customer experience.
I got a window into work being done to address communication gaps at the Salesforce AI Research roundtable. There, Salesforce AI experts Itai Asseo, Silvio Sarvarese and Madhav Thattai shared information on eVerse, a simulation framework to train Salesforce voice and text agents through synthetic data generation and reinforcement learning.
EVerse aims to deal with complex voice agent situations — spotty connections, distracted behavior and frustrated tone. These are the kinds of conditions that could significantly degrade the performance of tools like Dragon.
Dr. Sara Murray, chief health AI officer, UCSF
Agent augmentation at work: Healthcare billing
The most interesting capability demonstrated was the notion of human/virtual agent augmentation. Dr. Sara Murray, chief health AI officer at University of California, San Francisco, demonstrated this concept with a healthcare billing example. She explained that while simple questions like “Has my co-pay been applied to my bill?” can be handled entirely by an AI agent, more complex inquiries, such as, “What will my out-of-pocket cost be with multiple insurance plans?” still require human expertise, especially when several systems need to be traversed.
With eVerse, the people involved in the process can provide agent assistance and use no-code tools to train agents, improving agent native performance over time. This aligns with the perspective of authors like David De Cremer, who in “The AI Savvy Leader” advocates for human-AI augmentation and for having domain experts, rather than IT specialists, train the agents.
Madhav Thattai, senior vice president and COO, Salesforce
Clearly, being able to drive improvement for the most critical experiences is the promise of agentic AI, and this will require voice and a process of continuous learning. Thattai argued for a flywheel encompassing design and build, deployment and improvement. Murray underscored the need for workflows to undergo continuous refinement and get real-time feedback.
To make this work, eVerse enables the creation of personae that have one of the above problems and an enterprise simulation environment for training agents through synthetic data generation, stress testing and reinforcement learning. The framework aims to address these challenges by building trust through reliability, optimizing agents for both capability and consistency through three interconnected steps: synthesizing data, measuring performance and optimizing agent behavior.
Parting words
This marks another step forward in advancing agentic enterprise capabilities.
“With eVerse, we were able to test every nuance of human conversation before Agentforce Voice reached production,” said Jayesh Govindarajan, executive vice president of AI engineering at Salesforce, referring to a component of Agentforce 360. “That rigor is what turns breakthrough research into dependable customer experiences — and it’s how Agentforce Voice delivers the responsiveness and consistency enterprises expect.”
While this progress is certainly exciting, Salesforce views it as part of a longer journey toward what it calls enterprise general intelligence — the point at which AI equals or surpasses human capability. As Bob Muglia, former CEO of Snowflake, put it: “Unlike robots, AGIs will not be servants. They will be as smart or smarter than we are, and we must show them respect. It’s probably best to think of them as our peers.” Are agents leading us toward that world? Only time will tell.
