TELUS voice agent built on ElevenAgents more than halves 30-day cancellations

TELUS Digital, the technology arm of Canadian carrier TELUS (C$20 billion in annual revenue, 21 million connections), has built an outbound voice agent on the ElevenAgents platform for new home internet subscribers. The agent calls during the first 90 days after activation, walks through the first bill and shows customers how to manage their account on their own. According to TELUS, among customers the agent actually reached, the 30-day cancellation rate was less than half that of the overall cohort of new subscribers. The material was published on the ElevenLabs blog, which makes this a vendor case study, and the figures should be read with that caveat.

Why customers were leaving before the first conversation

Home internet is one of TELUS's highest-volume consumer products: in 2025 the company added 123 thousand net new customers there. An analysis of the first 90 days after activation identified four friction points that drove up inbound calls and early churn.

• Confusion over the first bill. Prorated charges for a partial month made the amount higher than expected.
• Low use of self-service. Many customers did not know that the MyTELUS app and telus.com work around the clock.
• Self-installation of equipment. Customers without a technician left more often.
• Missed discounts. Customers with both home and mobile plans did not know the accounts had to be linked.

A short call in the first weeks would have resolved most of these problems. But having people call every new customer was too expensive, and TELUS simply had no outbound customer support line.

How the call works and what the agent is not allowed to do

From November 2025 to February 2026, TELUS Digital compared enterprise conversational AI platforms against five criteria: voice expressiveness and tone, flexibility of the agent framework for complex outbound scenarios, real-time emotional modulation, security and compliance, and response latency. It chose ElevenAgents. The company's reasons: the speech sounds like a conversation rather than a recording, and branching scripts, tool calls in the middle of a call and human-handoff rules work without custom-built orchestration.

The call script is tightly structured:

1. The agent greets the customer, confirms their name, says it is an AI and asks for consent to continue; until then, no account details are discussed. The customer can decline at any moment. 2. It explains prorated charges and what to expect on upcoming bills. 3. It introduces MyTELUS and telus.com. 4. For customers with both home and mobile plans, it explains how to link the accounts for a discount. 5. It gives a setup tip and collects a rating for the conversation.

The most engineering-heavy part of the design is the handling of billing. If a customer asks about charges, the agent sends a one-time SMS code, verifies identity without interrupting the call and requests data through a verified API connection. Answers are tied to what the source returns: the agent quotes only amounts from the billing system. If verification fails or data is missing, it does not guess; it directs the customer to self-service channels. Account changes, complex troubleshooting and requests to speak to a person are handed off to TELUS staff.

About six weeks passed between first contact with the vendor and the first customer call. That window covered a security review, API integration, dialogue design, voicemail scenarios and a legal privacy review. Prior experience also factored into the choice: the Fuel iX Agent Trainer, a simulator where new agents rehearse conversations, is also built on ElevenAgents. By June 2026 it had run more than 90 thousand simulations, and employee onboarding time had fallen by 20%.

What the pilot showed and what the numbers leave out

In the trial phase, the agent made roughly 2300 outbound calls. The results TELUS reports:

• 30-day cancellations among customers who received a call were less than half the rate of the overall cohort of new subscribers;
• for every 1000 customers the agent spoke with, C$55 000 in lifetime revenue was retained;
• the average conversation rating was 8.5 out of 10;
• the agent reached the intended person in 19% of calls, matching TELUS's best human agents.

The questions about these data are obvious. The comparison is against the overall cohort, not a control group that would have been called by humans or not called at all. People who pick up the phone and agree to talk may differ from the rest even before the call. The material gives no absolute cancellation figures, no cost per call and no infrastructure costs, so the "agent versus human agent" economics cannot be calculated from it. The 19% figure also tells a different story: in most cases the intended person was not reached, and "on par with the best agents" means only parity, not superiority.

The practical takeaway for teams building similar agents: the value here lies in the combination of AI disclosure at the start, SMS identity verification, answers drawn only from live data and clear rules for handing off to a human. That scaffolding is what let the team ship in a regulated industry within a month and a half. According to Elena Bayich of TELUS Communications, the pilot became the foundation for proactive customer support: five additional scenarios are already in development.