TPG Telecom reports 40% churn drop in early agentic AI cohorts

By Daniel Okafor

TPG Telecom reports roughly 40% lower churn in the cohorts using agentic AI, ahead of a plan to let customers trigger those agents themselves.

At Cisco Live! in Melbourne, TPG Telecom general manager of technology operations Gaston De Arriba presented the customer-triggered vision. He did not announce a launch date, price or deployment scale.

What has TPG already deployed?

In March, TPG Telecom deployed a suite of Splunk tools to make service operations more AI-driven, an approach often called AIOps.

TPG stated three aims: real-time visibility across networks and IT, faster detection and resolution, and machine-learning prediction of operational issues before they affect customers.

Today, AI agents work in the background rather than directly with customers, collecting network performance data, key performance indicators, incidents, changes and complaints for correlation.

They identify each customer's problem in real time.

The system recommends customer actions and proposes network solutions for staff to implement.

De Arriba said: “One of the main pain points when customers have a problem is they call the call centre, and the call centre doesn't have enough information to tell them where is the impact and what we are doing to resolve it.”

The March disclosure named Splunk but identified neither individual products nor software versions.

Current deployments therefore support staff instead of handing customers direct control over network repair.

How strong is the churn claim?

De Arriba reported that agentic AI reduced churn by around 40% in the cohorts where staff used it.

He called the result an early metric and high value for TPG.

The figure means four in 10 on a percentage scale, not four in 10 customers, because De Arriba disclosed neither the starting churn rate nor the cohort size.

That distinction prevents readers from applying the result to TPG's entire subscriber base.

The remarks offer no company-wide trial or independent validation.

The metric also does not isolate AI from other service interventions.

De Arriba also described an unexpected benefit: proactive interventions have deflected many calls to contact centres.

Customers may never learn about a problem because TPG resolves it first.

He disclosed no call-volume baseline, reduction percentage or cost saving from that deflection.

The remarks therefore establish neither scale nor savings.

What changes for customers?

De Arriba said: “Our appetite is to give the power to the customer to trigger the agents from their own services, and then the agent to take care of the service performance for these customers.”

The proposed workflow would let customers start an agent from their own service, view performance on screen, receive optimisation guidance and, if needed, raise a trouble ticket.

The ticket option could replace a call, while routine self-diagnosis could eventually remove customer contact altogether.

De Arriba described the wider goal as more self-checking, self-healing and better understanding of the customer.

TPG says that model could prevent dissatisfaction faster than today's process.

No launch date, customer eligibility rules, service coverage, performance targets or commercial terms accompanied the proposal.

The remarks did not specify when direct customer access would begin or the degree of agent authority.

This remains a development plan, not a disclosed general-availability release.

What could derail the rollout?

TPG says it will move slowly, keeping agent actions within defined operational boundaries.

De Arriba identified customer experience as a primary concern, requiring agent actions to improve service outcomes.

Security becomes more important as customers interact with agents directly.

Cost is another material risk. De Arriba called the “tokenomics” of deeper agent deployment a clear point of risk.

He warned that releasing more agentic AI can easily escalate operating costs.

Neither token consumption, model pricing nor a spending ceiling appeared in the remarks.

The disclosure also omitted the number of agents, the Splunk deployment's cost and the scale of network automation.

The current division of labour remains limited: AI collects and recommends, while human staff implement network fixes.

Customer-triggered agents would extend that model without a published schedule, cost forecast or reliability threshold.

The 40% churn result therefore remains a management-reported cohort metric, not a promise for every customer or service.

TPG still has to disclose broader performance, deployment economics, operational limits and a dated release plan before customers can treat self-healing as a committed service.