What TPG Telecom’s AI agent rollout gets right
TPG Telecom reports early churn gains from AI agents. The more useful lesson is the order it built them in: data first, staff second, customers last.
At Cisco Live! in Melbourne, TPG Telecom’s general manager of technology operations, Gaston De Arriba, gave an early figure for its AI agents: in the customer groups where they’re in use, churn has fallen by around 40 percent.
That’s TPG’s own number, and it’s early. We wouldn’t build a business case on it. What’s worth studying is the order TPG has built things in, because it’s the opposite of how most businesses approach AI agents.
Agents behind the team, not in front of the customer
TPG didn’t start with a customer-facing bot. The agents first went to work in the background, supporting contact centre staff.
The problem they were solving is a familiar one. A customer calls with a service issue, and the person answering can’t see where the problem is or what’s being done about it. The relevant information exists, but it’s spread across network performance data, KPIs, incidents, changes and complaints.
TPG’s agents collect and correlate that information in real time, identify what the customer is likely experiencing, and give the staff member a recommendation on what to tell them. They also suggest fixes for the network team to implement.
The foundation for this was a set of Splunk tools TPG deployed in March, aimed at real-time visibility across its network and IT environment.
The side effect: fewer calls at all
De Arriba described a benefit TPG hadn’t planned for. Because the agents surface problems early, the team can act before customers notice anything is wrong, and those customers never call.
That’s the pattern we’d point any established business towards. The biggest gains from AI agents usually come from closing the gap between what your systems know and what your people can act on, not from a chat window.
Only then, the customer
TPG’s next step is letting customers trigger these agents themselves. A customer would check their own service, see how it’s performing and what they can do on their end, and raise a ticket if needed, without calling.
TPG is getting there slowly and on purpose. De Arriba named three constraints:
- Boundaries. Agents operate within defined limits, and only where they improve the customer experience.
- Security. It becomes a bigger priority once customers interact with agents directly.
- Cost. He described “tokenomics” as a clear risk, because unconstrained agents can quickly push up the cost of operating.
That last point gets less attention than it should. Agents that call tools, check results and call again can run up model costs with nothing to show for it, unless the workflow around them is tightly scoped.
What this looks like below enterprise scale
You don’t need a telco’s network to apply the same sequence. A trade business running several service crews, or a solar installer managing hundreds of systems, has the same split: job details in the CRM, schedules in dispatch software, history in email and call logs.
The order that works:
- Connect the data first. An agent can only be as accurate as the systems it reads. If your CRM and job records disagree, an agent will just give wrong answers faster.
- Put agents behind your staff. Let them pull context together and suggest the next step, while a person still decides. This is where AI workflow automation earns its keep.
- Keep fixed steps fixed. Anything transactional, such as booking a job or changing an account, runs through deterministic logic in your own systems, not at a model’s discretion. A custom web app is often what sits between the customer, the model and your data, handling authentication and limits.
- Then open it up. Once staff trust the output more than their own manual triage, it’s ready for customers, whether through a portal or an AI voice agent that checks availability and books through your actual calendar rather than guessing.
The takeaway
TPG’s churn figure will get the headlines. The more useful lesson is the sequence: data first, staff second, customers last, with costs and boundaries watched the whole way. That’s how AI agents become something you own and can rely on, not a demo you have to babysit.

Gurdeep Saroa
Founder & full-stack engineer
Software developer turned performance marketer. Builds the automation, tracking and web systems GRIVITY delivers.