Telstra says an AI agent built with Microsoft is already helping reconcile billing and payments across more than 272,000 service migrations, with recommendations at over 95 per cent accuracy in production. The agent flags inconsistencies and prepares cases for human review, putting agentic AI to work on the decidedly unglamorous business of getting customers from one service to another without losing the plot, or the payment.
Watch Desk analysis
What happened
The agent is part of work between Telstra and Microsoft’s Frontier Company, an AI implementation and consulting unit Microsoft launched in July. Microsoft says the system uses information across legacy and new platforms to spot discrepancies, recommend corrective action and create enriched cases for people to review. The reported figures are more than 272,000 migrations supported and recommendations at 95 per cent-plus accuracy in production.
Telstra has also appointed Dayle Stevens to an enterprise-wide AI leadership role and deployed an AI control plane, developed with Accenture, to monitor AI costs and governance. Those are further signs of the telco building oversight and organisational capacity around its AI deployments, not just adding agents one workflow at a time.
Why it matters
This is a live operational deployment at substantial volume, not a polished demo. Billing reconciliation is the sort of repetitive, cross-system work where agents could save staff time, but mistakes can have a direct effect on customer accounts. Microsoft says the agent recommends fixes and sends cases for human review; that distinction matters more than the word “agentic” on the slide deck.
The accuracy figure is Microsoft’s reported production result, not an independent evaluation in the article. Still, the scale gives readers something more concrete than a promise that AI will transform work: a named workflow, a reported deployment count and a human-review step.
Our read
The interesting part is not that AI has discovered billing. It is that a system is already operating across legacy and new services at scale, while people remain in the loop for recommendations. That is a practical model for enterprise agents: give them a bounded job, make their output reviewable, and measure whether they get it right. The numbers are promising, but the reporting does not explain how accuracy is defined or what kinds of errors remain.
What to watch
- Whether Telstra or Microsoft explains how the 95 per cent-plus accuracy figure is measured.
- How often human reviewers accept, amend or reject the agent’s recommendations.
- Whether the deployment expands beyond migration billing into other customer-service workflows.
Discussion spark: For AI agents handling customer billing, is a human review step enough, or should companies also publish error rates and the costs of mistakes?
Sources and evidence
- Source update (28 October 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.