Salesforce's eVerse Simulates Realistic Customer Service Interactions

Simulating Lousy Conversations: Q&A with Silvio Savarese, Chief Scientist & Head of AI Research at Salesforce

Salesforce is developing eVerse, a simulation environment designed to “battle-test” AI agents against the complexities of real-world customer service calls. Chief Scientist Silvio Savarese explains the need to move beyond scripted interactions, noting that current phone menus struggle with nuanced, multi-step problems, and that eVerse allows for the creation of synthetic data to address these edge cases.

Why This Matters

Current LLM training often relies on clean, curated datasets, a stark contrast to the messy reality of human communication. This disparity leads to AI agents failing in unpredictable environments, resulting in poor customer experiences and increased costs for human intervention—estimated at 60-70% of inbound healthcare contact center calls requiring human assistance.

Key Insights

• Synthetic Data Generation: Salesforce leverages small amounts of real data to create a wide range of synthetic scenarios, ensuring agents aren’t simply memorizing responses.
• Agentforce Determinism: Salesforce’s Agentforce platform allows for control over agent behavior, enabling developers to dial in levels of determinism or creativity.
• Human-in-the-Loop Learning: UCSF Health pilot shows agents can learn from human experts, improving coverage from 60-70% to 84-88% of routine inquiries.

Practical Applications

• Healthcare Billing: UCSF Health uses eVerse to automate routine billing inquiries, freeing up human agents for complex cases.
• Pitfall: Over-reliance on pre-training data can lead to agents failing to generalize to real-world scenarios and exhibiting inappropriate responses.

References:

• https://stackoverflow.blog/2025/12/11/simulating-lousy-conversations-q-and-a-with-silvio-savarese-chief-scientist-and-head-of-ai-research-at-salesforce/