Self-Optimising Mobile Networks: Long in Use, but AI’s Own Share Is Unmeasured

The use case

Elisa, a Finnish operator, reports that an energy-saving tool covering its whole radio network cut energy cost by 14% on top of the static power-saving methods it already used (operator- and vendor-reported; the page shows no date)⁴. The claim is typical of the field: a specific number, a large footprint, and no published method.

A self-organising network (SON) is software that configures, tunes and repairs a radio network from measurements, instead of engineers adjusting each site. 3GPP, the body that writes mobile standards, introduced SON in Release 8 in 2009 and groups it into self-configuration, self-optimisation and self-healing functions¹. The standard leaves the content of these functions "left to implementation", so each vendor builds its own¹. SON platforms were already reported in use in 2G, 3G and 4G networks, with performance and cost benefits, in a 2022 peer-reviewed review (the review's assertion, no figures given)². Manual radio optimisation, the alternative, is described there as taking about 8 weeks per round².

Scope. AI here means machine learning, including reinforcement learning (software that learns by trial against feedback), used to set radio parameters, switch capacity on or off, or detect faults. Not credited as AI: threshold-based SON in use since the 2G to 4G era, manual optimisation, and human operations staff. 3GPP's own page says SON processes will "likely" adopt AI models in the future¹. Customer-service AI and core-network automation are out of scope.

Where it is running

TM Forum, the industry body, scores network autonomy from Level 0 to Level 5. These levels cover network operations broadly, so they mix machine learning with rule-based and manual work⁷⁹. All named deployments are in Europe, Asia and Latin America; evidence from the United States was not examined. SON in 2G to 4G networks predates all of these, from 2009¹².

What the evidence shows

No independent evaluation of machine-learning network optimisation in a live network was found. Every result below is operator-reported, vendor-reported or survey-based.

• Energy. Elisa reports 14% lower energy cost (operator/vendor-reported, undated, no baseline or period)⁴. China Mobile Hubei, with Huawei, reports a 10.4% energy saving with service indicators maintained (operator/vendor-reported, May 2025, no period or method)⁶. Rakuten expects "up to 20%" lower energy use (operator-reported, expected, May 2025, no baseline)⁵.
• Fault handling. In a trial in east Indonesia, Telkomsel reports traffic loss down 12.6%, repair time down 6% and a customer experience index up 14.7% (operator-reported trial, March 2026)⁷. China Mobile reports a 30% average cut in fault and complaint repair time and over 30% fewer back-office operations staff (company-cited via TM Forum, March 2026, covering AI and non-AI changes together)⁷.
• Maturity. In a TM Forum survey of 125 respondents at 80 companies, 21% said they operate at Level 3 or above, up from 19% a year earlier (self-reported survey, March 2026)⁷. In a Bain and TM Forum survey of 22 operators, 20% had reached Level 4 or 5 in selected domains (self-reported, 27 August 2025; small sample, co-produced with TM Forum)⁸.

The mechanism behind these gains is plausible: radio equipment can be put into low-power sleep when traffic is low, and faults can be found and repaired faster by software than by people. But the sources do not show how much of each figure comes from machine learning, how much from older automation, and how much from the process changes introduced alongside.

What it takes: cost, integration, governance

Cost. No public cost per site or per operator was found. Bain's surveyed operators target about 30% operating cost savings by 2028, which is a target, not a result⁸.

Data and systems. 3GPP states that an AI model "must be trained with lots of different data" before it reacts appropriately to changing conditions¹. The 2022 review notes that centralised SON is the most widespread design but can be a single point of failure, and that the software controlling the network needs full read and write rights to the operations systems². Bain's respondents name technical debt, vendor interoperability, legacy migration, skills gaps and organisational resistance as barriers (self-reported, August 2025)⁸.

Pace. Telefonica targets an average autonomy level of 3.75 by 2028 and Level 4 by 2030 (operator-reported targets, February 2026)⁹.

Rules. 3GPP specifies where SON inputs come from but not what the functions do¹. No regulator guidance specific to automated network changes was read in this research.

Where it falls short

The challenge question is what would make this use case look less successful than it is presented. Four answers emerge.

No baseline, no method. None of the energy figures states a measurement period, a control, or a baseline, and Elisa's claim of no customer impact comes with no service metrics⁴⁵⁶.

AI is not separated from older automation. Operator pages say "AI" but do not say machine learning or describe the model; Rakuten's announcement does not explain how its savings are achieved⁵. Because SON has existed since 2009, a share of any reported gain may come from rule-based functions¹.

Known conflicts. A peer-reviewed 2015 analysis (abstract only; modelled, not measured, and older than three years) found that concurrent SON functions can conflict and trigger chain reactions, so coordination is needed³. No newer field measurement of this was found.

Change risk in live networks. ENISA's 2024 data from national authorities record 188 telecom incidents, the highest on record, with 113 caused by system failures (60%) and 548 million user-hours lost (regulator-aggregated data, July 2025). Faulty software changes and updates caused 29 incidents; the report's summary and body give different totals for this category¹⁰. The data do not say which incidents involved automation, so they show the size of change risk, not the risk of AI. The 2018 O2 outage in the United Kingdom, lasting almost 23 hours and affecting about 25 million customers, came from an expired certificate in supplier software, not from automation (regulator decision, November 2019; older than three years)¹¹.

People and process also carry weight: TM Forum itself found its autonomy scoring tool "excessively subjective" and is working on enhancements⁷.

Phase: In early use — why

The hypothesis was "In wide use". Automated network optimisation as a whole is in wide use, with SON in standards since 2009¹². But only confirmed machine-learning deployments count towards the phase, and none of the named deployments confirms the method or the share of the network affected⁴⁵⁶. 3GPP still describes AI in SON as a likely future step¹. The lower phase is therefore chosen, and it should be raised if an operator publishes the method and a measured result.

What to watch next

• Telefonica's targets of an average autonomy level of 3.75 by 2028 and Level 4 by 2030⁹.
• Whether operators in Bain's survey reach their roughly 30% operating cost savings target by 2028⁸.
• The next edition of TM Forum's annual Level 4 progress report, after the March 2026 edition, and its promised fixes to the scoring tool⁷.
• Publication of any operator result with a stated baseline, period and method for its energy-saving figure, such as Rakuten's "up to 20%"⁵.

Questions to ask before adopting

• Does the vendor confirm in writing that the optimisation uses machine learning, and which functions remain rule-based?
• What baseline, measurement period and control group underlie the quoted energy saving?
• Which service quality indicators were tracked alongside energy, and what were the results?
• How are conflicts between concurrent automated functions detected, and who can reverse a bad change?
• What happens to network operation if the central controller or its data feed fails?