Voice AI agents resolve complex requests 32% more successfully than chatbots

A customer's flight has been canceled. They need to find a new itinerary, check the connection, confirm the surcharge, and receive a ticket. In chat, this kind of conversation quickly turns into a long exchange: the person replies in fragments, scrolls back to earlier messages, and becomes more frustrated with every clarification. Voice lets them explain everything at once—but only if the AI agent responds without agonizing pauses and can actually complete the transaction.

ElevenLabs compared voice calls and chats involving several hundred operational AI agents. The sample included agents that conducted at least 50 conversations in each channel between May and July 2026. For each business, outcomes were assessed using its own criteria, while the voice and text versions of the same agent were compared with each other.

After two minutes, the balance shifts

In conversations shorter than one minute, chat performed about 9 percentage points better. Between one and two minutes, there was almost no meaningful difference. After two minutes, voice moved ahead: the share of successful requests was 10.9 percentage points higher, or 31.7% higher relative to chat. Nearly two-thirds of agents delivered a noticeably better result in the voice channel.

Successful long phone conversations took an average of 4.6 minutes, compared with 7 minutes for chat exchanges. Among the 90 agents with enough successful requests of both types, 78% completed voice conversations faster.

In the canceled-flight example, the difference is tangible. Instead of a chain of ten messages, the customer says: “I need a flight today, with no overnight connection, and the surcharge must not exceed ten thousand.” The agent clarifies one ambiguous parameter, checks the available options, and processes the replacement. Voice does not make the model smarter here. It shortens the path between the person's intent and the system's action.

Latency can erase the entire advantage

In a separate group of 315 conversations, agents were switched to a faster language model. Average end-to-end latency fell by 16%, while the share of successful calls rose by 26%. Calls that ended within the first 15 seconds dropped from 21.1% to 3.8%.

This is correlation, not proof of direct causation: other parameters may have changed along with the model. Still, the product signal is clear. The opening seconds matter especially—a silence after the customer speaks can make them think the connection has dropped and hang up.

In a small user study, 98 participants handled a simulated canceled-flight problem through voice and chat in randomized order. Voice was preferred by 85%, most often because of speed and a sense of empathy. However, a laboratory scenario and a sample of 98 people do not allow this result to be generalized across all industries and audiences.

What to implement in practice

There is no reason to replace the entire interface with voice. For short requests—checking a status, getting a link, or changing a single parameter—chat remains more convenient. Voice makes sense for tasks involving multiple constraints, a high cost of error, and emotional pressure.

Teams should measure completed actions rather than call duration alone: whether a refund was issued, a booking changed, or a claim closed. It is also useful to track time to the first response, latency for every reply, the share of calls abandoned within the first 15 seconds, and how often the conversation is handed over to a human.

The study was published by a voice-platform provider that is also promoting ElevenAgents, so its figures require independent verification. Duration is also only an approximate measure of complexity: in chat, the user may simply become distracted. Even so, the practical conclusion is robust: for complex requests, a voice agent can provide a meaningful advantage if it responds quickly, retains context, and has access to real business operations. An attractive voice without those three capabilities is not enough.