How revenue teams deploy voice AI agents for customer service

Tom Eggemeier at Zendesk says voice AI agents for customer service are ready for real calls. One Zendesk customer solved 92% of its interactions with an AI agent. Getting there, he said, takes good data, an internal team and service architects.

Lex Sivakumar, VP of Growth at Regal, described agents that go well beyond answering the phone. They place five to 10 parallel outbound calls to vendors, which he estimated cut resolution times in half.

Are voice AI agents for customer service ready?

Eggemeier said the ceiling for automation is high, and Zendesk's A/B testing showed AI agents making fewer mistakes than human agents on a range of tasks. Regal has transcribed and analyzed over 400 million calls to understand how customers use voice AI most effectively.

How leading revenue teams deploy voice AI agents

Regal deployed voice agents for roadside assistance to find available providers. The agent verifies the caller, grabs their location and pings a maps API to find nearby help. Sivakumar explained that “The voice agent then makes five to 10 outbound calls instantaneously, all at the same time, in parallel” to gather availability and pricing.

Regal also set up agents for labor marketplaces that handle negotiation. Sivakumar said that “in the previous human example, the labor was actually gaming the system. They figured out exactly how humans were coordinating the negotiation process” and were getting the highest prices. Regal randomized the negotiation logic to make it as fair as possible for both sides.

Real-world results: pipeline metrics and call outcomes

Revenue teams are also using AI to protect existing accounts. Alphonso H. Calanoc, Director of Customer Success at Momentum, recalled managing a highly engaged 400K account with 90% entitlement consumption: “two months into managing this relationship where I thought we had a solid renewal, pulled the rug out from under us, and I took a 300K, almost 400K churn.”

Customers would tell him, “Yeah, it's great working with you guys, but we're not picking it up yet. We're not quite at ROI.” That forced the team to change its delivery model. By operationalizing call signals and summaries, Calanoc said, “I think for the past year and some change now, we've been able to maintain roughly 120, 125% NDR.”

Quality, deliverability and brand risk

Over the last six months, Regal ran into problems with dialects and accents. Sivakumar recalled, “Someone said that their name was Jesus, and then the voice agent responded and called them Jesus. That's not acceptable.”

The Regal engineering team fixed it by modeling specific examples for Latinx names: “we specifically prompted with the agent and asked it to point attention to Latinx names in order to make sure those names were developed properly.” They are now expanding the practice to other languages.

Integration with CRM and the revenue stack

Pol Peiffer, Head of Agent Development at Sierra, said some data doesn't need to pass through the model at all. If an agent needs to give an account balance, sending that number through the LLM risks it coming out wrong. “What should go into the language model is a template string, and you swap it at the end.” Sierra also uses tools that let agents bypass the LLM entirely to read legal disclaimers.

Peiffer said voice agents have to be built for latency: “On a simple turn of a conversation, we might make a dozen plus inference calls to different models.”

Impact on RevOps roles and team structure

Peiffer said that when AI handles 80% of customer experience cases, the remaining 20% get more interesting for human agents, and “we've seen retention go up for those roles as well, just because it's a more interesting job.”

Difficult calls still take a toll. Sivakumar said, “Humans take twice the amount of time to recover to baseline productivity as opposed to when they're dealing with a neutral call.” Agents usually jump straight onto the next call anyway, and that customer gets a worse experience.

The teams getting results from voice AI agents for customer service are the ones handling the edge cases, like Regal's work on name pronunciation, and saving human agents for the calls that need them.

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