What the standards add

Every 3GPP release since 2019 has added something an AI programme can use

Three groups inside 3GPP write the parts of the standard an AI programme uses, and it helps to keep them apart. The core-network groups define where analytics data comes from and how a model may be shared. The management group defines how a model is trained, tested, deployed, monitored and rolled back. The radio groups define where a model may live in the air interface, and the signals that let a cell save energy (Exhibit 1).

The core got its analytics function in 2019 and a way to check it in 2023

The network data analytics function, NWDAF, was “specified from the 5G’s initial Release (Rel-15)”, in 3GPP’s own words, and its specification, TS 23.288, reached version 16.0.0 in June 2019. Release 16 said what it collects and what it puts out. Release 17 split model training from inference, added functions to coordinate data collection and store analytics, and allowed “Trained data model sharing between multiple NWDAF instances, limited to single vendor environments”. Release 18 added federated learning in version 18.0.0 of December 2022 and then, in version 18.1.0 of March 2023, the means to compute “the accuracy of the ML models and analytics”. Release 19 added vertical federated learning and policy recommendations driven by predicted quality of experience.

Two of those dates matter more than the others (Exhibit 2). The first is March 2023, because a standard way to check whether a model in the core is still accurate arrived almost four years after the specification’s first full version. Until then an operator could run a model but had no standard way of knowing whether it had drifted. The second is March 2021, when sharing a trained model between vendors was written out of Release 17. Put the two together and a multi-vendor analytics estate built on NWDAF is recent territory for the standard, and we have found no operator that has published a result from a commercial NWDAF deployment.

The radio study recommended the uses one vendor can build alone

The air-interface study, TR 38.843, was approved in December 2023, and its conclusions are as interesting for what they decline as for what they recommend. On compressing channel-state information with a two-sided model: “From RAN1 perspective, there is no consensus on the recommendation of CSI compression for normative work”. The reasons given are the “Trade-off between performance and complexity/overhead” and “Issues related to inter-vendor training collaboration”. On predicting channel state there was no consensus either, for “Lack of results on the performance gain over non-AI/ML based approach and associated complexity”. What the study did recommend for normative work was beam prediction and positioning, both of which one vendor can build at one end of the link.

Release 19 then gave “normative support for a general framework of one-sided (at the gNB or UE sides) AI/ML models as well as … beam management, positioning enhancements and CSI prediction”. Offline training is assumed for positioning. The two-sided model went to a Release 20 work item instead (Exhibit 3). 6G part 8 covers that alongside the rest of the AI-native argument.

For operations the relevant radio items are smaller and more useful. Release 18 specified AI/ML in the radio access network for energy saving, load balancing and mobility, with an “Energy Cost” index that base stations exchange with one another. Release 19 extended the work to slicing and to coverage and capacity optimisation. The energy work adds signalling rather than intelligence, and the distinction matters. Release 18 specified cell discontinuous transmission and reception, secondary cells without synchronisation blocks and the barring of older devices. Release 19 added synchronisation and system information on demand, woken by a device’s uplink signal. Each of those is a hook that lets a scheduler put more of a cell to sleep. The decision about when to sleep, which is where the AI lives, stays with the vendor or the operator. That is why part 3 found its energy results in company announcements rather than in specifications.

The radio standard has adopted the AI uses one vendor can implement alone, at one end of the link, and sent the uses that need two vendors to agree off to a later release. So the near-term AI value in the RAN arrives as software bought from the operator’s existing RAN vendor.

The management plane now names the controls a kill gate needs

The management specifications are where the standard meets the kill gate. Management data analytics, TS 28.104, was first published for Release 17 in June 2022 and covers coverage, fault and energy-saving analytics. The AI/ML management specification, TS 28.105, covered training in Release 17 and grew to the whole life cycle in Release 18: training, testing, emulation, deployment and inference. Release 19 adds reinforcement learning, and with it a set of named controls (Exhibit 4).

The autonomy levels belong to the same family of documents. TS 28.100 defines six levels, from L0, manual, to L5, full, across five kinds of task. L4 is the level at which “All the execution, awareness, analysis and decision tasks are accomplished automatically by telecom system itself. And intent handling tasks can be partly accomplished automatically”. Its first note settles who wins when a person and the system disagree, and the answer is the person.

So the standard now writes down what a kill gate needs for a model that learns on a live network. It names a degradation that counts as tolerable, an action to fall back to, and a human decision that overrides the system at every level. None of that is new thinking. Part 5 shows why each of those controls was needed long before AI came into it.

One third-party rApp runs in a live network, and nobody has published what it did

O-RAN’s specification set reached 147 unique titles by February 2026, 71 of them new or updated since November 2025. It includes an A1 service for managing ML models and conformance tests for rApps. That is a lot of paper. The adoption record is short by comparison (Exhibit 5). AT&T announced on 30 July 2025 that it had become “the first CSP globally to deploy a third-party rApp to optimize its live production network”, on Ericsson’s automation platform. It did not name the developer and it has not published a result. Every other entry in Exhibit 5 is a validation, a test environment or a vendor’s own figure.

The near-real-time controller and its xApps have done worse. Ericsson told Light Reading in December 2024 that the RAN itself provides what xApps would, and Samsung saw no near-term path for a standalone near-real-time RIC. Omdia found live deployments limited in February 2025. The clearest RIC-hosted use with an operator’s name on a number is Samsung’s energy manager at Verizon, which TelecomLead reports runs on Qualcomm’s RIC. Part 3 marks its figures as the vendor’s own. Our reading is that an operator planning AI in the RAN is buying from its RAN vendor, whatever the architecture diagram shows.

Autonomy levels are self-assessed

Autonomy levels are the industry’s scoreboard for operational AI, and every score on the board was awarded by the operator to itself (Exhibit 6).

China Mobile set the boldest target. Its Lingli Deng told TM Forum Inform in August 2023: “We are targeting Level 4 by 2025 for almost all the services that we are currently providing”, at a time when its production average was 2.85. Its 2024 results said the network “has already reached an advanced level (L4) of smart autonomy in some scenarios”. The interim chairman’s statement of 13 August 2026 says “Autonomous networks achieved L4-level capabilities in 21 high-value scenarios”, with 38 planned over three years. Notice what happened to the unit. The target was written in services, the result comes back in scenarios, and no document we read gives an average. Telefónica reports a group average of 3.42 at the end of 2025, up from about 1.1 five years earlier, with targets of 3.75 by 2028 and Level 4 by 2030. Verizon says it is working beyond Levels 2 and 3 and aims at L4 “in the near future”.

The surveys put the claims in proportion. TM Forum’s survey of March 2026, which asked 125 people at 80 companies, found 21% at L3 or above overall, against 19% a year before. Bain’s survey of 22 operators with TM Forum found 20% at L4 or L5 “in select domains”. The AI-RAN Alliance had 132 members by February 2026 and has published reference architectures rather than measured results. And the GSMA’s benchmarks of language models on telecom tasks found that performance on turning an intent into a configuration “remains low, with top models scoring under 30”. That task, as it happens, is the one closest to an AI changing a network by itself.

China Mobile’s claim moved from L4 for almost all services by 2025 to L4 in 21 high-value scenarios in 2026, and nothing in the standard stopped it. TS 28.100 defines the level and says nothing about who may award it. Our advice is to treat an autonomy level as a self-assessment of a named scope, and to ask for the scope.

The terms, briefly

• NWDAF. Network data analytics function: the 5G core’s function for collecting data and producing statistics and predictions for other network functions.
• MDA. Management data analytics: the same idea in the management system, for coverage, faults and energy.
• One-sided and two-sided model. A one-sided model lives wholly in the phone or wholly in the network. A two-sided model is split across the link, so two vendors have to train it together.
• RIC, rApp, xApp. The RAN intelligent controller, in non-real-time and near-real-time forms; rApps run on the first, xApps on the second.
• L4. In TS 28.100, the level at which the system performs all execution, awareness, analysis and decision tasks itself.