Industry deploymentMicrosoft Azure OpenAI ServiceMicrosoft Azure AI SearchMicrosoft Azure Machine LearningMicrosoft 365 CopilotSalesforce AgentforceMicrosoft Frontier Company (Migration Payment Agent)

Telstra's three-year path from a 200-agent GenAI pilot to production AI agents handling refunds and billing migrations

Telstra Group Limited· Australia / APAC· Contact centre agent assist, customer self-service deflection, regulated compliance workflow automation, and billing reconciliation during platform migration
Share
Telstra Group Limited logo
Telstra Group Limited
Microsoft Azure OpenAI Service logo
Microsoft Azure OpenAI Service
AI deployment

Telstra moved generative AI from a 200-agent contact centre pilot in 2023 to production agents that calculate regulated refunds and reconcile billing across 272,000+ service migrations. The published metrics are real — and unevenly evidenced. Here is what the record actually supports, and what it does not.

Results at a glance · every figure cited

90%Agents reporting increased effectiveness with One Sentence Summary (pilot)
20%Reduction in follow-up contact on calls handled with One Sentence Summary (pilot)
84%Ask Telstra users agreeing it positively impacts customer interactions (pilot)
8,000+Frontline staff using AskTelstra
over 1 minuteReduction in average customer call time attributed to AskTelstra
almost 3xIncrease in queries answered by Telstra Assistant without escalation
2.5 millionProactive SmartFix network actions in FY25
nearly 1 millionSupport calls prevented by SmartFix in FY25
18,000+Employees with Microsoft Copilot licences (80%+ weekly use)
272,000+Service migrations supported by the Migration Payment Agent
95%+Migration Payment Agent recommendation accuracy in production (vendor-reported)
more than 1,000Contact centre representatives whose administrative load the CSG agent reduces
86%Consumer service interactions completed via digital self-service (1H26)
$179 million / 2.4%Underlying operating expense reduction, 1H26 (not AI-attributed)

The challenge

Telstra's frontline staff had to answer questions spanning mobile plans, NBN fault restoration, prepaid recharges and legacy billing from a knowledge base no individual could hold, while the company carried roughly 80 separate data platforms and 18 data and AI vendors. Separately, regulated processes such as Customer Service Guarantee interim refunds — validating eligibility, calculating compensation, monitoring restoration — consumed administrative time across more than 1,000 contact centre representatives and created compliance exposure when handled manually.

What was deployed

Telstra consolidated its data estate (from ~80 platforms toward ~20) and moved to a cloud-based, API-first architecture, then layered four distinct AI systems on top: AskTelstra, a retrieval assistant on Azure OpenAI Service plus Azure AI Search over the internal knowledge base; One Sentence Summary, a CRM button that compresses customer history using Azure Machine Learning, Azure AI Search and Azure OpenAI; Telstra Assistant, a customer-facing generative AI chat assistant launched in November 2025; and agentic systems — Salesforce Agentforce agents for CSG interim refunds and Salesforce permissions monitoring, plus a Microsoft Frontier Company Migration Payment Agent that reconciles billing across legacy and new platforms with human-in-the-loop case review. Delivery was accelerated via a seven-year joint venture with Accenture announced in January 2025 and an existing JV with Quantium.

The results

Pilot-stage survey results reported by Microsoft in 2024: 90% of agents testing One Sentence Summary increased effectiveness, calls required 20% less follow-up, and 84% of Ask Telstra users agreed it positively affected customer interactions. By April 2026 Telstra reported AskTelstra was used by 8,000+ frontline staff and cut average customer call time by over a minute; Telstra Assistant almost tripled the number of queries answered without escalation; and SmartFix performed 2.5 million proactive network actions in FY25, preventing nearly 1 million support calls. In October 2026 Microsoft reported the Migration Payment Agent supporting more than 272,000 service migrations at 95%+ recommendation accuracy in production. Telstra's 1H26 results reported 86% of consumer service interactions completed through digital self-service and underlying operating expenses down $179 million (2.4%).
TL;DR

Telstra, Australia's largest telco, spent three years moving generative AI from a 200-agent pilot to production agents that touch contact centres, billing migrations and network operations. The published numbers are real but uneven: a 2024 Microsoft case study reports pilot-stage survey results, Telstra's own April 2026 blog claims AskTelstra cuts more than a minute off the average call for 8,000+ frontline staff, and Microsoft's October 2026 blog says a Migration Payment Agent is supporting 272,000+ service migrations at 95%+ recommendation accuracy. What is missing is an audited dollar figure attributable to AI, and a clear accounting of the job losses that have travelled alongside it.

The problem: a knowledge base no human can hold

Telstra's frontline problem is the classic telco problem, just at national scale. Agents field questions about mobile plans, NBN fault restoration, prepaid recharges, business connectivity and legacy billing systems — and the product catalogue keeps changing underneath them. Microsoft's customer story describes the starting point plainly: Telstra's staff "require a vast technical knowledge base in the face of this increasing complexity", and the company saw generative AI as a way to shorten the distance between a customer's question and the right answer.

Behind that sat a messier structural issue. In an April 2026 post on Telstra Exchange, the company's Data & AI Executive described starting from around 80 data platforms, since consolidated to roughly 20 and still falling, plus a migration to a cloud-based, API-first architecture. That ordering matters: the AI work that gets quoted in press releases was preceded by years of unglamorous data and platform consolidation.

What was actually built

Four distinct things, built at different times, by different partners, for different users. Treating them as one "Telstra AI programme" is the main way the public coverage misleads.

Mid-2023

Ask Telstra trialled by about 200 customer service agents. Built on Azure OpenAI Service with Azure AI Search for retrieval over Telstra's internal knowledge base. One Sentence Summary tested separately by 100 agents — a button inside the CRM that compresses a customer's recent history into a single line.

Feb 2024

Microsoft publishes the customer story with pilot-stage results. Full deployment is flagged for early 2025.

Jan 2025

Telstra and Accenture announce a seven-year data and AI joint venture, 60% Accenture / 40% Telstra, explicitly aimed at agentic AI. Telstra consolidates from 18 data and AI vendors to two JVs.

Nov 2025

Telstra Assistant, the first customer-facing generative AI assistant, launches on telstra.com for tasks like plan checks, SIM activation and password resets.

Sep 2026

Salesforce Agentforce agents go live, including an interim service agent handling Customer Service Guarantee refunds, and an internal security permissions agent.

Oct 2026

Microsoft discloses a Migration Payment Agent built with Microsoft Frontier Company, reconciling billing across legacy and new platforms during customer migrations.

The agent that is easiest to evaluate

Of the four, the Customer Service Guarantee (CSG) agent is the most instructive, because it automates a rule-bound compliance process rather than an open-ended conversation. The CSG is an Australian regulatory obligation governing how fast a fixed-line fault must be restored and what compensation is owed when it is not. That is a workflow with a defined input, a defined calculation and a defined audit trail — which is exactly the shape of work where agents are currently reliable.

How the interim service agent runs

1
Validate eligibility

The agent checks whether the customer's outage qualifies under NBN and CSG rules.

↓
2
Calculate the refund

A guided workflow computes the compensation amount rather than leaving it to a representative's interpretation.

↓
3
Monitor for restoration

The agent watches for the service to come back before acting — the step humans forget and the reason refunds used to be late.

↓
4
Apply and record

Refund applied, customer records updated, compliance evidence generated as a by-product.

By automating complex, manual processes like interim refunds, Agentforce is helping us improve efficiency and accuracy across our contact centres

— Amy Childs, Customer Service Representative, Telstra, iTnews

The measured outcomes

20% less follow-up contact on calls handled with One Sentence Summary (pilot) Source: Microsoft Customer Stories, 2024
>1 min cut from average customer call time via AskTelstra, now used by 8,000+ frontline staff Source: Telstra Exchange, 2026
~3x increase in queries resolved by Telstra Assistant without escalation Source: Telstra Exchange, 2026
2.5M proactive SmartFix network actions in FY25, preventing nearly 1 million support calls Source: Telstra Exchange, 2026
272,000+ service migrations supported by the Migration Payment Agent Source: Microsoft, 2026
86% of consumer service interactions completed through digital self-service (1H26) Source: Telstra 1H26 results transcript, 2026

Two of these deserve a closer read. The 2024 figures are survey results from a pilot cohort, not operational telemetry: Telstra's then-General Manager of Customer and Channel AI reported that agents said they were more effective, and that their calls needed less follow-up.

Ninety percent of customer service agents who tested One Sentence Summary increased their effectiveness

— Rohit Lakhotia, General Manager of Customer and Channel AI, Telstra, Microsoft Customer Stories

The unanticipated benefit was onboarding. Lakhotia noted that "New agents found that Ask Telstra gave them a big boost to becoming more effective, faster" — because they stopped queuing for a team leader's attention. Ramp-time compression is one of the most reliably reported second-order effects of retrieval assistants, and it rarely appears in the original business case.

The 2026 Microsoft figure is the opposite kind of number: operational, large, and thinly specified. As Microsoft's commercial chief described it, the Migration Payment Agent is a human-in-the-loop system — "The agent uses intelligence across these systems to proactively identify inconsistencies, recommend corrective actions and create enriched cases for human review". The agent recommends; a person decides. That design choice is what makes a 95% accuracy claim tolerable in a billing context, and it is the single most transferable detail in this entire case.

Governance was built before scale, not after

Telstra has published more about its AI control environment than most listed companies. According to its April 2026 governance post, "All AI systems, whether built in-house or by a partner, are included on an AI register and screened for their impact level on various stakeholders." Higher-impact systems go to an AI Risk Oversight Council drawing on data science, engineering, legal, privacy, cyber security, people and culture, and customer experience. Every AI system has a named owner, and whether a human-in-the-loop is required is a documented decision rather than an implicit one.

There is also a hard gate on tooling: "every Telstra employee must complete responsible AI training before getting access to Microsoft 365 Copilot". Telstra says a Responsible AI "control plane" — a company-wide platform for monitoring models against policy — is being built and tested in non-production.

!

Reading pilot survey data as production ROI

The widely-cited Telstra numbers — 90% effectiveness, 84% positive impact, 20% less follow-up — come from agents who volunteered for a trial of a new tool, surveyed by the team that built it. Novelty, selection bias and the desire to be helpful all push those numbers up. They are a signal worth acting on; they are not a measured outcome across 8,000 agents.

Fix: instrument the system-of-record, not the survey. Average handle time, repeat-contact rate within 7 days, and first-contact resolution are already in the CRM. Run the pilot as a holdout experiment against a matched control group, and publish both numbers.

Pilot evidence vs production evidence

PhaseWhat was claimedEvidence typeWhat it supports
2023–24 pilots90% more effective; 20% less follow-up; 84% positive impactAgent self-report, ~300 usersA decision to scale
2026 assistants>1 minute off average call; ~3x unescalated resolutionCompany-reported operational metric, method undisclosedDirectional confidence
2026 agents272,000+ migrations; 95%+ recommendation accuracyVendor blog, measurement period undisclosedScale, not yet value
1H26 financialsUnderlying opex down $179m (2.4%)Audited half-year resultsCost discipline overall, not AI attribution

The parts the public record does not settle

Telstra's CEO has been candid that AI is central to the company's economics. In the January 2025 joint venture announcement she said: "We've made strong progress on our AI goals and already have hundreds of value-driving AI use cases across the business." "Hundreds of use cases" is also, notably, the phrasing Telstra still uses in 2026 — and hundreds of use cases is not the same as hundreds of millions of dollars.

Three honest gaps:

No audited AI P&L. Telstra reported underlying operating expenses down $179 million, or 2.4%, in the first half of FY26, and attributes efficiency to a bundle of causes — software partner consolidation from 400 vendors to two, API-first architecture, relentless simplification and AI adoption. No published figure isolates AI's contribution. TechInformed made the same observation about the Microsoft disclosure, noting the 272,000 figure is a count of service moves supported rather than payments checked or errors corrected, with no disclosed measurement period behind the accuracy claim.

No deflection or containment rate for Telstra Assistant. "Almost tripling" the number of queries resolved without escalation is a ratio with an undisclosed baseline. Tripling from 5% to 15% and tripling from 20% to 60% are entirely different businesses.

The headcount question. Telstra has reduced roles alongside this programme, and in February 2026 told investors that proposed changes to the Accenture JV — tapping Accenture's existing resources to deliver the data and AI roadmap faster — would mean that some roles would no longer be required. The company has not published a reconciliation between AI-driven productivity and net role reductions, and no quoted executive has offered one.

What an integration team should take from this

First, sequence matters more than model choice. Telstra spent years collapsing 80 data platforms toward 20 before the agent work landed. Nothing in the public record suggests a frontier model rescued a fragmented data estate; the consolidation came first.

Second, the agents that reached production are the ones with adjudicable outputs. A refund calculation, a billing reconciliation, a permissions audit — each has a right answer that a human or a rule can check. Telstra did not put an autonomous agent in front of a vulnerable customer; it put a summary button in front of the human serving that customer.

Third, human-in-the-loop is a product decision, not a compliance afterthought. The Migration Payment Agent creates "enriched cases for human review" by design. That is what lets you ship at 95% accuracy instead of waiting for 99.9%.

Fourth, the adoption flywheel is organisational. Telstra reports 18,000+ Copilot licences with more than 80% weekly use, 20,000+ employees through a Data & AI Academy, and mandatory responsible-AI training as a gate on tool access. Telstra's Data & AI Executive framed the whole programme as a reinvention rather than a tooling exercise: "We first set an ambition to become an AI-fuelled organisation and use AI to address major pain points in our business".

If you are planning similar work, the practical lesson is to pick the compliance-shaped process first, instrument it against the system of record, and keep a person in the approval path. Our case study library and automation practice cover how that pattern generalises beyond telco; start here if you want to map it onto your own workflows.

Frequently Asked Questions

What AI systems has Telstra actually deployed?

Four publicly documented ones: AskTelstra and One Sentence Summary, agent-facing tools built on Azure OpenAI Service and Azure AI Search; Telstra Assistant, a customer-facing generative AI chat assistant launched in November 2025; Salesforce Agentforce agents handling Customer Service Guarantee refunds and Salesforce security permissions monitoring; and a Migration Payment Agent built with Microsoft Frontier Company to reconcile billing during platform migrations.

How much money has Telstra saved with AI?

Telstra has not published an AI-specific savings figure. Its first-half FY26 results show underlying operating expenses down $179 million, or 2.4%, but management attributes that to a bundle of factors including software partner consolidation, an API-first architecture, simplification and AI adoption. Treat any single-number "Telstra saved X with AI" claim as unsourced.

Who built Telstra's AI systems?

Telstra built the early tools in-house on Microsoft Azure. In January 2025 it announced a seven-year joint venture with Accenture — 60% Accenture, 40% Telstra — to accelerate its data and AI roadmap, and consolidated from 18 data and AI vendors down to two joint ventures, the other being Quantium Telstra. Salesforce supplies Agentforce; Microsoft Frontier Company built the Migration Payment Agent.

Is the 95% accuracy claim for the Migration Payment Agent verifiable?

Not independently. The figure comes from a Microsoft corporate blog post in October 2026. Microsoft did not disclose the measurement period, the number of recommendations assessed, or the criteria used to judge a recommendation correct. The agent operates with human review of its cases, so the practical risk of an incorrect recommendation is mitigated — but the accuracy number itself is a vendor claim, not an audited result.

How does Telstra govern AI risk?

Every AI system, in-house or partner-built, goes on an AI register and is screened for stakeholder impact. Higher-impact systems are reviewed by an AI Risk Oversight Council spanning data science, engineering, legal, privacy, cyber security, HR and customer experience, with board-level Audit & Risk Committee visibility. Each system has a named owner, human-oversight requirements are documented, and responsible-AI training gates access to Copilot.

Twarx analysis

Original interpretation

The Telstra agents that reached production are the ones with adjudicable outputs — a refund calculation, a billing reconciliation, a permissions audit — shipped with a human in the approval path. Open-ended customer conversation was handled by giving the human a better summary, not by replacing the human.

Three transferable lessons. (1) Sequencing beat model selection: years of collapsing ~80 data platforms toward ~20 and moving to API-first architecture preceded the agent work, and nothing in the record suggests a frontier model rescued a fragmented data estate. (2) Compliance-shaped processes are the best first agentic target, because eligibility rules and refund formulas have checkable right answers and generate their own audit evidence. (3) Human-in-the-loop is a product decision that lowers the accuracy bar you must clear before shipping — the Migration Payment Agent creates enriched cases for human review by design, which is why 95% accuracy is tolerable in a billing context. Integration teams should also note the evidence-quality gradient: Telstra's pilot numbers were good enough to justify scaling but would not survive as ROI claims, and the company's strongest governance asset (a named owner and documented human-oversight decision per system, with an AI register and risk council) was built before scale rather than retrofitted after an incident.

Read the numbers honestly

Telstra has published no audited figure isolating AI's contribution to earnings; the $179m opex reduction in 1H26 is attributed to a bundle of causes including vendor consolidation and simplification, not AI alone. The widely-cited 2024 percentages are agent self-reports from trials of roughly 200 and 100 users respectively, not instrumented operational telemetry across the full agent population. 'Almost tripling' unescalated resolution for Telstra Assistant is a ratio with no disclosed baseline. Microsoft did not disclose the measurement period, sample size or correctness criteria behind the 95%+ accuracy claim for the Migration Payment Agent, and neither Telstra nor Microsoft has published that agent's internal impact classification. Telstra has reduced roles alongside the programme and told investors in February 2026 that proposed changes to the Accenture JV would mean some roles would no longer be required, but has published no reconciliation between AI-driven productivity and net headcount change. Finally, no independent audit, academic evaluation or regulator review of any of these claims is in the public record.
Telstratelecommunicationsagentic AISalesforce AgentforceAzure OpenAI Servicecontact centre automationcustomer service AIcompliance automationAccenture joint ventureresponsible AI governancecase studyAPAC
RS

Analysis by

Rushil Shah · AI Systems Builder & Founder, Twarx

Agentic AIMulti-Agent SystemsAI Workflow Automationn8nLangGraphAI Integration

From evidence to deployment

Want telecommunications results you can measure — and defend?

We deploy and audit AI agents the way we research them: real baselines, cited outcomes, and no inflated ROI. Twarx maps where an agent actually pays off for your team — and, just as honestly, where it won't.

  • A utility audit before you build — so budget goes where it moves the metric
  • Source-grounded benchmarks, not vendor marketing numbers
  • Deployed with measurement baked in, so ROI is provable later
Book a utility audit

The teams that win with AI agents aren't the ones with the boldest claims — they're the ones who measured honestly and deployed where the evidence actually pointed.

— Twarx Research