ElevenLabs study: voice agents beat chat 32% on long calls

ElevenLabs published production data comparing voice agents and chat agents on the same deployments. The headline: in conversations longer than two minutes, voice agents succeeded about 32% more often than text chat. In short conversations, chat won. If you run one AI support channel for every kind of question, you are probably losing the hard ones in the wrong place.

What actually happened

ElevenLabs looked at agents on its platform that handled at least 50 voice and 50 chat conversations between May and July 2026. Each agent was scored against success criteria the business defined, and each agent's voice results were compared against its own chat results. That design matters. It stops one high-performing, voice-heavy customer from skewing the whole sample.

The numbers, per the November 2 post:

• Under one minute: chat did better by about 9 percentage points.
• One to two minutes: roughly even.
• Over two minutes: voice success rates were 10.9 points higher, a 31.7% relative lift. Nearly two-thirds of agents did meaningfully better on voice.
• Speed: successful long conversations averaged 4.6 minutes on voice versus 7 minutes on chat.
• Latency: across 315 agents that moved to a faster LLM, a 16% cut in end-to-end latency correlated with a 26% lift in success rate. Calls ending inside 15 seconds dropped from 21.1% to 3.8%.

A separate user study of 98 people rebooking a simulated canceled flight found 85% preferred voice.

Two caveats. This is a vendor measuring its own product, and it uses conversation length as a stand-in for complexity. ElevenLabs says so itself. Nobody has replicated it independently yet.

Why the voice-vs-chat data matters for your business

Match the channel to the job. "Where's my order" is a 40-second chat. A disputed charge, a botched install or a rebooking is a phone call. The data says to stop forcing both through the same widget. Put chat on the quick lookups and give the messy problems a voice line that can actually finish the work.

Latency is the first impression. One in five callers hanging up inside 15 seconds is a dead-line problem, not an AI problem. Before you tune prompts, measure time-to-first-word. A faster model that is slightly dumber can win on resolved calls.

Define success before you launch. Every number above rests on success criteria the business wrote down. If you can't say what a resolved call is, you can't tell whether voice is helping.

Sources: ElevenLabs.