If you’ve followed telco industry coverage this year, you could be forgiven for thinking autonomous networks have arrived. Operators are announcing Level 4 milestones. TM Forum’s DTW Ignite in Copenhagen was wall-to-wall agentic AI. Vendors are declaring the era of the self-healing network open.
The truth is more interesting — and more useful — than the headlines. AI Ops is real, the momentum is real, but the Level 4 claims cover far less ground than the announcements suggest. If we’re all being realistic, the frontier that matters for the industry in 2026 is Level 3 — and understanding why is the best way to work out what your network should be doing right now.
AI Ops has changed its meaning
Two years ago, AI Ops in a telco context meant machine learning applied to monitoring: anomaly detection, alarm correlation, predictive maintenance. Useful, but fundamentally advisory — the system flagged a problem and a human fixed it.
In 2026 that definition looks dated. The conversation has shifted to autonomy: closed-loop systems that don’t just detect and recommend, but decide, act, verify the outcome, and learn from it. The engine behind this shift is agentic AI — reasoning systems that can plan and execute multi-step operational work, increasingly organised as networks of specialised agents that share context across domains. Even conventional AI Ops platforms are now criticised for “stopping at recommendation.” The bar has moved.
The industry’s shared yardstick for all of this is TM Forum’s Autonomous Networks framework, which grades operations from Level 0 (fully manual) to Level 5 (full autonomy). The levels that matter in practice:
- Level 3 (conditional): the system senses, decides and acts within specific domains; humans intervene by exception.
- Level 4 (high): the system self-configures, self-heals and self-optimises across a domain with minimal human involvement.
Level 4 is where the headlines are. Level 3 is where the industry’s real work is — and the distinction matters.
What Level 3 actually means
Level 3 is the point where the relationship between humans and the network inverts. Below it, systems assist and humans act. At Level 3, the system senses, decides and acts within defined domains, and humans intervene by exception. That inversion — the machine drives, the expert supervises — is the meaningful line to cross, and most of the industry hasn’t crossed it yet. Surveys this year put the large majority of operators at Level 0–2, with automation still fragmented across scripts, vendor tools and manual correlation.
What about the Level 4 announcements? They’re genuine achievements, but look closely at what they cover. The first validation under TM Forum’s assessment scheme — Ericsson with TDC NET in Denmark — covers predictive cell energy management. NBN Co’s Australian-first validation covers service assurance for a single service in the fixed wireless domain: managing atmospheric ducting, a weather condition that degrades radio performance. Each validation certifies one named combination of operational flow and domain. A validated Level 4 scenario is best understood as an island of high autonomy inside an operation that runs at Level 2–3 everywhere else. Everything either side of the certified scenario is still integration work, still cost, still humans.
That’s not a criticism — it’s how autonomy should be built, one bounded domain at a time. But it reframes the honest question for 2026. It isn’t “when do we reach Level 4?” It’s “where can we credibly operate at Level 3 today?” The industry’s own numbers support that framing: roughly one in five operators expect to reach Level 4 by 2027, while 81% have set a goal of Level 4 or beyond by 2030. Ambition is nearly universal; for the rest of this decade, conditional autonomy in well-chosen domains is the realistic benchmark of progress.
Standardisation will take until the 2030s. Starting shouldn’t.
The standards work is moving fast — TM Forum’s AI-native Open Digital Architecture extensions, its “Race to 2030” strategy, intent APIs, digital-twin frameworks — but that framing is itself the tell. The industry body coordinating this effort has put 2030 on the label. Cross-domain, multi-vendor, verifiable autonomy at network scale is a next-decade outcome, and anyone who tells you otherwise is selling something.
TM Forum isn’t working alone. Project Sylva — the Linux Foundation Europe initiative through which Deutsche Telekom, Orange, Telefónica, Vodafone, Ericsson and Nokia are building a common open-source telco cloud stack — has turned its attention to the same problem from the infrastructure side. Sylva’s own analysis this year of why telco AI projects stall at pilot scale identified three culprits: fragmented infrastructure that stops AI being replicated across the network, disconnected data systems that block scaling, and missing execution paths for autonomous operations. That heavyweight names are collaborating on these foundations is encouraging. That the foundations still need building is the point.
There’s a second reason for patience on the grand vision: trust. The obstacle to letting AI run networks is no longer intelligence — it’s whether the AI can be trusted to understand what it’s looking at, and whether its actions can be verified before they touch production. This is why digital twins have become central to the autonomy roadmap: a continuously updated virtual replica of the live network where an agent’s proposed fix can be rehearsed and evaluated before execution. And it’s why so much of the real work inside operator autonomy programmes is unglamorous: data quality, inventory accuracy, integration. A twin fed stale data verifies decisions against a network that doesn’t exist.
So the honest 2026 picture is this: the destination is clear, the standards are a work in progress, and the leaders are proving autonomy one bounded domain at a time.
Which raises the practical question: if bounded domains are where autonomy actually works today, which of your domains is best suited to start?
The messaging network is a better starting point than you think
The AI Ops conversation is dominated by RAN, transport and core. The messaging layer barely features — which is odd, because it fits the profile of a strong early autonomy domain almost perfectly:
Clear goals. Messaging operations have unambiguous, measurable objectives: delivery rates, latency, throughput, fraud loss, A2P revenue leakage. Intent translates cleanly into metrics.
Reliable, well-structured data. Every message event is logged. Traffic patterns, routing outcomes, sender behaviour and delivery receipts form exactly the kind of high-volume, well-labelled telemetry that AI systems need — without the multi-vendor data-integration problems that stall autonomy programmes in the RAN.
Reversible, contained actions. Blocking a sender, rerouting traffic, throttling a suspicious campaign, adjusting a firewall rule — these are fast, bounded and recoverable interventions. The blast radius of a wrong decision is small compared with reconfiguring a live radio network.
A live, adversarial problem worth automating. Messaging fraud — smishing, artificially inflated traffic, grey routes — already moves at machine speed. Fraudsters iterate campaigns in hours. A detection loop that stops at recommendation, waiting for a human to review a report, is structurally too slow. Closed-loop detection and response isn’t a futuristic ambition in messaging; it’s the only approach that matches the threat.
In TM Forum terms, an operator running AI-driven campaign detection with automated blocking and human oversight by exception is already operating at conditional autonomy — Level 3 — in the messaging domain. That is further up the maturity curve than most operators’ flagship network programmes, achievable now, and it generates something the wider autonomy journey desperately needs: organisational experience of trusting a closed loop, governing it, and measuring its business impact.
Start where the loop can close
The lesson from the operators furthest ahead isn’t “wait for the platform.” It’s the opposite: pick a bounded domain with clean data and reversible actions, close the loop there, prove it, and expand. The pioneers chose cell energy management and fixed wireless assurance. For most operators, the messaging network — with its structured data, measurable outcomes and urgent fraud problem — is sitting there as an equally strong candidate for reaching Level 3, largely overlooked.
Full autonomy across the network is a 2030s story. Autonomy in your messaging network can be a this-year story. The operators who treat it that way won’t just protect revenue and subscribers sooner — they’ll arrive at the standardised, cross-domain future already knowing how to run a network that acts for itself.
