The Ugly Truth About Voice AI Agents for Customer Service

Voice AI agents for customer service are ready to handle real calls, and revenue teams deploy them by targeting specific call volumes and prompting models for cultural nuance, according to Florian Otto at Cedar and Lex Sivakumar at Regal. Otto expects these deployments to eliminate around 40% of the call volume humans handle today.

Sivakumar, VP, Growth at Regal, recalled that his team prompted the agent to pay attention to Latinx names so those names were handled properly.

Are voice AI agents for customer service ready for production?

The production data says yes. Cedar has pushed 90,000 calls through its system with early clients, and Sivakumar said Regal analyzed 400 million calls of transcribed data to learn what works.

Somya Kapoor, CEO at IFS Loops, reported seeing close to 60% to 65% autonomy in CX and support. Arvind Mathur at Amazon Web Services Singapore Pvt. Ltd. reported that Druva cut its problem-solving time by about 63%. Sivakumar said multiple roadside assistance vendors partnered with Regal to automate their workflows and cut a three-to-four-hour timeline in half.

Which use cases drive revenue outcomes?

Ori Entis at Gainsight described a 20/80 reality. About 80% of customers are probably very small, so the cost to serve is too high and you cannot throw people at the problem. Sivakumar said voice AI agents now own outcomes, with literal P&L responsibility, and no longer just execute tasks.

To hit your KPIs, Mati Staniszewski at ElevenLabs argued, you need bespoke solutions built by deployed engineers working directly with your team. He said that approach produces conversational agents that pass the Turing test.

Ashwin Sreenivas, Co-founder & President at Decagon, said that after tens of millions of interactions, users often thank the agents. Some conversations run four to five minutes because the agents are not allowed to hang up. Sivakumar added that a bad call is hard on human reps, who take twice as long to recover to baseline productivity as they do after a neutral call.

How do revenue teams deploy voice AI agents?

Keeping agents on the line is half the work. The other half is the messy data those conversations produce. Entis explained that the system collects emails, calls, support tickets, and any other unstructured interaction between a customer and a vendor, cleans it up, and moves it into databases so automation can use it. When you build a digital twin from that data, he warned, you have to fuse in subject matter expertise or it becomes a bloat of data that is not very useful.

Staniszewski argued that voice is our most natural way to communicate, which makes it the key to AI's promise. Sivakumar said Regal is rolling out a new Copilot product that cuts time to value from weeks to hours.

Where do voice AI agents fall short?

Sivakumar recalled a caller who said his name was Jesus. The voice agent responded with the English pronunciation of the name. That dialect problem led to the Latinx prompt adjustment he described.

Entis admitted that Gainsight started out believing it could build one agent that does everything. Very quickly, the team saw that today's technology cannot do that. If you throw complex data schemas at these models, he warned, they do not easily understand them, and they pull data from the wrong parts of your database and give the wrong answer.

How should revenue teams evaluate voice AI agents?

One panelist advised against inventing new AI KPIs, because people will gamify them. Work with your existing business metrics to see real movement. Another panelist recommended a hybrid approach: use agentic AI where it removes friction, and keep deterministic logic for flows that must be 100% accurate.

Entis reported that over 60% of customer success organizations are still running early experiments and proofs of concept. He argued that large language models are becoming the new front door to software, and said Gainsight has launched MCP connectors and servers so customer context can be pulled into workflows through Claude, ChatGPT, or Gemini.

Revenue leaders deploying voice AI agents for customer service should focus on operational detail over hype. Constrain the system with strict data schemas and measure it against the business outcomes you already track.

To hear more from the people putting AI into production, see the HumanX 2027 program preview.