What TPG Telecom announced
On 12 November, TPG Telecom's general manager of technology operations, Gaston De Arriba, told the audience at Cisco Live in Melbourne that the telco is building AI agents its customers can trigger themselves. iTnews reported that the agents would diagnose and fix service problems, potentially without the customer ever calling TPG. Under the planned design, a customer would see how their service is performing on screen, find out what to adjust, and raise a trouble ticket with the call centre when self-service can't resolve the fault.
TPG hasn't published a launch date, pricing or a list of the fixes the agents will be allowed to make. De Arriba said the company is moving slowly so the agents stay inside defined boundaries.
The groundwork came first
TPG didn't jump straight to customer-facing agents. In March 2026, Cisco announced that TPG had chosen a suite of Splunk products to unify network operations and security. Splunk Enterprise handles real-time monitoring and links incidents to specific faults, and Splunk IT Service Intelligence uses machine learning to project service health. The telco runs the Vodafone, TPG, iiNet, Lebara and felix brands on that platform.
A month later, iTWire described what the platform was already doing. Root cause analysis that used to take anywhere from 15 minutes to several hours now takes a few minutes in most cases. When a customer calls, the system checks network performance, incident records, coverage maps and earlier complaints so the contact centre agent has context during the call. De Arriba told iTWire that cases escalated to the network operations centre used to come back days later, often with no fault found. Automated outage notifications now stop about half of potential calls from reaching the contact centre, and customers who do call give those interactions a 20% higher net promoter score.
According to iTnews, TPG's agents currently work behind the scenes. They feed call centre staff network, KPI, incident, change and complaint information and recommend fixes in real time. De Arriba cited a reduction in churn of roughly 40% in the customer groups served with agentic AI. The 40% figure is TPG's own early measure from a conference talk, and the company hasn't published how it defined the cohorts or over what period it measured them.
Why the order matters
The customer-triggered agents are the last step in a sequence TPG has taken over at least eight months. The telco built a single view of service health first. Next came internal agents to help its own staff, and only after that did TPG talk about putting agents in front of customers. Each stage gave the company data on whether the agent's diagnosis matched reality before any customer saw it.
Many NZ organisations want to skip to the visible part. A customer-facing support agent is easy to demo, and vendors sell them as a quick win. An agent can only diagnose a fault if it can read accurate, current data about the customer's service. If incident records sit in one tool, change records in another and customer history in a CRM nobody trusts, a customer-facing agent will give confident wrong answers. TPG's iTWire figures show the payoff came from correlating data sources a human used to chase manually.
De Arriba also named the risks plainly: customer experience, security once customers deal with agents directly, and cost. On cost, he said agentic AI "can easily escalate your cost of operate". Every customer who opens the self-service screen triggers model calls, data lookups and possibly automated actions. A support channel that used to cost staff time now costs usage fees that grow with every customer interaction, and a bad week for outages means a spike in the bill.
What it means for NZ businesses
The TPG model applies well beyond telcos. Any NZ business with service faults customers can see can follow the same sequence. Internet providers, managed IT providers, utilities, equipment hire firms and SaaS vendors all fit. The customer-facing agent needs three things underneath it: a reliable record of what's broken, a record of what changed recently, and rules on what the agent may fix without a person.
Most mid-sized NZ firms already hold the raw material in their service desk. Incidents, problems and changes recorded consistently under an ITIL 4 style service management practice are the data an agent reads to explain a fault. When categorisation is loose and change records are incomplete, the agent has nothing solid to reason from, and staff will spend their time correcting the agent's answers.
The security point carries extra weight here. A customer-triggered agent is an interface the public can type into, and some of those people will try to make the agent do things it shouldn't. Restart a router, yes. Change account details or waive fees, no. NZ firms need to decide those limits before launch and enforce them in the agent's permissions. A line in the prompt isn't enough. The Privacy Act 2020 still applies to every customer record the agent reads or reveals, so the agent should only see the data for the customer it's helping.
Practical next steps
- Map where fault data lives today. List the systems holding incidents, changes, monitoring alerts and customer complaints, and check whether a person can join them for one customer in under five minutes. If staff can't, an agent won't either.
- Start with an internal agent that supports your own service desk or call centre staff. Measure how often its diagnosis matches what the engineer finds, and keep it internal until accuracy is consistently high.
- Write down the actions a customer-facing agent may take on its own, the actions that need staff approval, and the actions it must never attempt. Enforce the list through system permissions.
- Model usage costs before launch. Estimate interactions per customer per month, then test the numbers against your worst outage week from the past year.
- Pick a measure you can defend. TPG uses churn and NPS by cohort. Choose a baseline, a comparison group and a period before the agent goes live, so any result is credible.
- Plan the handover. Customers who reach the end of what the agent can do should land with a person who can see what the agent already tried.
If you're weighing up where agents fit in your own service operations, our AI agents page sets out how we approach design, permissions and testing.
Sources
- TPG Telecom wants customer-triggered AI agents (iTnews)
- Cisco to Power TPG Telecom's Next Generation Security and Service Operations Across Australia with Splunk (Cisco Newsroom)
- TPG Telecom's quiet AI play is already cutting fault times from hours to minutes (iTWire)
- DataForge Insights feed
- DataForge AI agents
- DataForge ITIL 4 framework
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help