ElevenLabs Launches Experiments Feature for A/B Testing AI Voice Agents.

ElevenLabs announced on LinkedIn a new feature called Experiments for its ElevenAgents platform, bringing controlled A/B testing to AI voice agent optimization. The feature allows developers and product teams to test agent variants in production with real traffic, measuring performance across business metrics before committing changes.

Experiments addresses a common challenge in deploying conversational AI: uncertainty about which configuration changes actually improve outcomes. Teams can now create agent variants, route a controlled slice of traffic to each version, and measure impact on key performance indicators like customer satisfaction scores, containment rates, and conversion metrics. When a variant proves superior, teams can promote it to production with full version control.

The feature supports experimentation across multiple dimensions of agent behavior. Teams can test different prompt structures, workflow logic paths, voice characteristics, and personality traits to understand what drives better results in real-world conversations.

This capability matters particularly for companies like Revolut and Klarna, which have deployed ElevenLabs agents at scale for customer support. Revolut selected ElevenLabs to bolster its customer service operations, while Klarna reduced time to resolution by 10X using ElevenAgents. For deployments handling millions of customer interactions, incremental improvements from testing can translate to significant operational gains.

The system includes version control, allowing teams to track changes over time and roll back if needed. This mirrors standard software development practices but applies them specifically to conversational AI configuration.

Experiments focuses on business outcomes rather than just technical metrics. While traditional A/B testing platforms measure click-through rates or conversion funnels, ElevenLabs built Experiments around metrics specific to voice agent performance. Teams measure containment rate (how often the agent resolves issues without human escalation), customer satisfaction scores from post-call surveys, and conversion rates for sales or support scenarios.

The controlled traffic routing ensures statistical validity. Rather than switching all users to a new variant at once, teams can test with a small percentage of calls, gathering data while limiting risk if the variant performs worse than expected.

ElevenLabs continues expanding its agent capabilities following partnerships with Deloitte for enterprise deployments and Deutsche Telekom for network-integrated voice assistants. The company raised a $500 million Series D in February at an $11 billion valuation, reflecting strong enterprise demand for voice AI infrastructure.

Experiments is available now to ElevenAgents users. The feature ships as part of the platform's existing interface, requiring no additional integration work for teams already running agents in production.