What Telstra has said
In a joint paper reported by CommsDay, Telstra chief architect Mark Sanders and Ericsson’s Mats Karlsson argue that simply adding artificial intelligence to existing automation platforms will not be enough to reach highly autonomous networks. Operators first need systems that understand both network intent and the wider operational context, a “knowledge plane” that turns data into usable knowledge, and AI whose decisions can be trusted and explained. The destination they name is the higher rungs of the TM Forum’s autonomous network scale, which runs from Level 0 (fully manual) to Level 5 (full autonomy). At Level 4, humans set the goals and the network runs itself with minimal intervention.
Set that beside the strategy. Under Connected Future 30, chief executive Vicki Brady has told shareholders the 2030 workforce “will be smaller than it is today,” has called AI “a significant unlock,” and has guided to more than A$2 billion in annual savings from AI in customer engagement and more than A$1 billion in software and IT. Telstra has bought 21,000 Microsoft Copilot licences, runs 380 internal AI use cases, and wants 85 per cent of teams using AI weekly. For its part, Telstra frames AI as augmenting its workforce rather than replacing it, and says almost 9,000 staff completed a Data and AI Academy course in the first half of the 2026 financial year to build those skills.
On the public record, Telstra says the technology cannot run the network alone today, and it has told investors the workforce will be smaller by 2030 as that technology matures. This projection follows those two statements to where they point, a business that runs on as few people as the technology safely allows. The open question is one of timing.
The five technical gates to a self-running network
A telco that runs itself needs five things to fall into place, and Telstra is already working through most of them.
The first is clean data: one consolidated view of the network rather than a hundred scattered ones. Telstra has been collapsing 80 separate data systems toward a target of three, and is already down to about 20. The second is a way to turn a plain instruction, keep this customer’s service fast, into the thousands of actions that deliver it without a person in the middle, which is the part Telstra and Ericsson say is unfinished. The third is AI trusted enough to act without human sign-off, a question of accountability as much as code. The fourth is agents that act rather than advise; 2026 is the year those move out of pilots and into live networks, though for now a person still approves anything significant. The fifth is a safe place to rehearse, a digital twin where a change can be tested before it reaches a network carrying live traffic.
None of these is finished. Each is moving. On our reading, “not yet” describes work in progress.
The earliest window, layer by layer
Telstra employs about 29,000 people. Automation does not arrive for all of them at once. It comes in waves, and the order is visible.
The first wave is the work closest to a screen. Telstra’s own engineers are writing code roughly 20 per cent faster with AI, and the company has bought 21,000 Copilot licences and stood up 380 internal AI use cases. The second wave is the call centre and the back office, where Telstra has told investors it expects more than A$2 billion in savings from AI in customer engagement alone. The third is the network operations centre, the room that watches the network around the clock, which is the exact room China Mobile says it now runs at Level 4 with AI agents. The last wave is field crews and the safety-critical core, which moves slowest because accountability for those services is the hardest thing to hand to a machine.
The table below is a read of when each wave becomes technically possible.
Where the work sits | Roughly when AI could do it | Basis |
Writing software and running IT | 2026 to 2028 | Already underway; Telstra says AI has sped up its software work by about 20 per cent |
Answering customers and back office | 2027 to 2029 | The layer carrying Telstra’s A$2 billion customer-service savings target |
Watching the network (the NOC) | 2027 to 2030 | China Mobile says its AI agents already run this room at Level 4 |
Field work and network planning | 2029 to 2031 | Slower; it touches the physical world and needs safe rehearsal first |
The safety-critical core | Beyond 2030 | Last to go, because accountability for safety-critical services is hardest to automate |
The shape is consistent. The work closest to a keyboard becomes automatable first, and the work closest to physical, safety-critical infrastructure last.
The forecast: 2030
On the technology’s own trajectory, and on Telstra’s own guidance that its workforce will be “smaller than it is today” by 2030, our central forecast is the end of this decade. By 2030, most of the work that can be automated, the code, the contact centre, the back office and the network watch room, is technically replaceable in practice.
The scale follows from Telstra’s own figures. Telstra cut 7.4 per cent of its workforce in calendar 2025 alone. If that single year’s pace simply held, the 29,000 of today would fall toward the low 20,000s by 2030, somewhere in the order of 7,000 to 9,000 jobs. BT, a few steps further down the same road, has told investors it intends to go from 107,652 staff to no more than 80,000 by 2030, roughly a quarter of its people. These are our illustrative figures, not Telstra projections. They are included to show the scale the question operates at, measured in roles rather than percentages.
The precedents: China Mobile, BT and Verizon
The question has real precedents, which is what makes it worth asking rather than dismissing.
By China Mobile’s own account, it has reached Level 4 autonomy in its network operations centre using AI agents, and says the equivalent of 5,500 full-time manual roles were replaced by the end of 2024, with answer accuracy above 90 per cent. It reports it is now replicating the approach nationwide. These are vendor and operator figures rather than independently audited results, so they are best read as direction, though the direction is clear. On TM Forum benchmarking, only about 4 per cent of operators have reached Level 4 so far, and most pursuing it are aiming for 2030, a few for 2028.
The headcount precedents are already on the board overseas. BT is targeting no more than 80,000 staff by 2030, down from 107,652 in March. Verizon’s new chief executive cut 10,000 roles in a single quarter and has publicly forecast far higher unemployment as AI spreads. Across 20 tier-one telcos tracked by Light Reading, around 160,000 roles have gone since late 2022. Telstra, by this measure, is not leading the charge.
The real ceiling is trust, regulation and liability
The ceiling on full automation is not set by the technology.
Three things cap it. The first is trust and liability: handing safety-critical national infrastructure to a fallible system, with no human who can be held responsible, is a decision a board and a regulator have to make, not an engineer. The second is sovereignty: most foundational AI models originate offshore, and a carrier that built itself into a fully autonomous network on a foreign model would be exposed if access were ever restricted. The third is economics, which cuts against the hype. Labour is only about a fifth of operating costs at a typical large telco, so even aggressive automation moves profitability less than the headlines suggest, and analysts note much past telco shrinkage owed more to divestments and switching off old technology than to AI.
On the numbers and the science, Telstra could plausibly run most of its back-office, software and customer-facing work with very few people by the end of this decade. Running all of it, including the safety-critical core, is not a date that can be set, because the last barriers are human judgement, regulation and accountability rather than capability.