The candidates, ranked by what it costs to find out

A number costs least to trust where there is a meter and a control

The order here is the order in which a pilot can produce a number that survives a finance review. Rank the same ten by the size of the prize and you would get a different list, and we say so because the two get confused.

The logic is borrowed from the AI opportunity portfolio model, with value held equal across the ten. That model divides expected value by the cost of the cheapest test that would settle the question. It rates confidence High only when the effect has been measured “in your own business or in a published result you can check”, and Low when “Somebody believes it”. A published result with a site count, a period and a baseline moves a candidate up that scale. A vendor’s “up to” does not. So we marked each candidate on the five things that set what a test costs (Exhibit 1).

The marks produce the ranking in Exhibit 2, and the shape of it is plain enough. Energy, agent assist and RAN optimisation all use data the operator already holds, read within a quarter and allow a control. Fault correlation, field dispatch and procurement carry the largest claims and the thinnest baselines, and they rank low for that reason and no other.

Energy is the cheapest to test, and BT has the only national figure

Energy has the shortest path to a number anyone can check, because the meter and the control site both exist before the pilot starts. Nobody has to build anything to measure it.

BT holds the one national figure. In June 2024 it switched on cell sleep across “over 19,500 sites across the UK” of its EE network, using “machine learning to predict quiet periods of demand”. The function itself came from “the respective RAN equipment supplier” on each site. The release gives two numbers, and they are worth reading together. One is “up to 2 KWh per site per day”. The other is “4.5m KWh per year across EE’s estate”. Divide the second by the sites and the days and you get about a third of the first.

Every other energy figure we found is a forecast, a trial, an estimate or a vendor’s own statement (Exhibit 3). SoftBank expects about 5 million kWh a year once its sleeping cells grow from about 14,000 to 24,000, and that is a forecast. Samsung says its AI energy manager on Verizon’s network gave “an energy saving gain of 15% on average, with a maximum of 35% per sector during low traffic periods”. Ericsson says Vodafone UK cut the daily power of its 5G radio units “by up to 33 percent at select sites across London”. Neither vendor says how many sites or for how long. Telefónica’s engineers estimate, without a measured result, that power-saving features save 20% to 30% of RAN energy and that AI might add another 5% to 10% on top.

Why do the percentages scatter so widely? The standard itself explains it (Exhibit 4). 3GPP’s Release 18 study, TR 38.864, reports ranges from many companies’ simulations rather than single values, and the ranges are enormous. Nine sources found that lengthening the synchronisation-signal period up to 1,280 ms “could achieve BS energy savings by 0.9%~84.8% in range, meanwhile when traffic occurs and load increases, the UPT significantly decreases”. Aligning devices’ connected-mode sleep cycles with the cell’s saved anywhere from 0.2% to 71.4% across six sources. Three of those found user throughput fell by 0.91% to 15.5%, and one found losses of up to 62.4% in some configurations. Load and the starting configuration move the answer by two orders of magnitude, which is another way of saying that a percentage on its own is not a result.

Two cautions travel with any energy result. The first is about intensity. Light Reading showed that Telefónica’s energy per petabyte fell from 268 MWh in 2016 to 54 MWh five years later, while traffic grew more than fourfold. Total consumption fell “by as little as 11%”. An intensity figure can collapse while the bill barely moves. The second is about attribution. Vodafone Türkiye’s July 2026 trial, up to 10% from software and hardware power saving, credits “advanced algorithms” rather than AI, and we see no reason to relabel it.

Energy ranks first because a test is cheap and the unit is physical. It is also, and this is the catch, the candidate where headline and measurement lie furthest apart. BT’s “up to” figure is three times its own average, and the standard’s simulations span two orders of magnitude. The pilot that settles it counts kilowatt-hours at matched sites and measures user throughput alongside, so that a saving bought with dropped sessions shows up as what it is.

Care has results by journey, and no operator publishes an all-contact rate

Customer care is where most operators’ published AI results land, and every one of them is for one journey, one market or one tool (Exhibit 5).

BT says its assistant Aimee handles about 60,000 conversations a week, double what it handled two years earlier, with “automation success rates on several types of customer journey now approaching 50%”. Vodafone’s SuperTOBi lifted first-time resolution for appointment booking in Portugal “from 15% to 60%”, though its early answers were right only 75% to 78% of the time before rework took them to about 90%. Its chief technology officer reckoned a less cautious operator could launch in “six weeks to three months”. Telstra’s agent assistant, in the hands of more than 8,000 staff, is “cutting average customer call time by over a minute”. These are real results. They are also narrow ones.

Two documents mark the limits. Verizon’s own survey of 5,000 consumers in seven countries found “88% of consumers are satisfied with interactions handled mostly or fully by human agents”, against 60% for AI-driven ones. McKinsey described an agent copilot that “realized just a 5 percent improvement in productivity” until 90% of its budget was moved to training and change management, after which the gain passed 30%. That is why we rank agent assist above deflection. A person stays in the loop, and the test splits cleanly by agent group.

A deflection pilot that reports containment without repeat contacts and complaints beside it is reporting the one measure that rises whether or not the customer’s problem was solved. No operator in the published record states an all-contact containment rate with its definition, and until one does, every containment figure is a number without a denominator.

The network-operations claims are the biggest, and most of them are self-assessed

The fault-correlation and dispatch figures are the largest in the published record, and they are the hardest to test. The data needs cleaning before a pilot can start, results take quarters rather than weeks, and the baseline moves as the process changes under it.

• Fault correlation. Indosat reports resolving “up to 85% of common wireless access faults in a single step” and cutting raw alarms “by up to 95%” across more than 55,000 base stations. It also says, candidly, that “wrangling and cleansing the data is 70% to 80% of the time and effort”. It publishes no repair time and no ticket count. China Mobile’s pilots with Huawei report average repair time down 30%, on its own assessment.
• Automated change. Verizon made “more than 70 million network configuration changes autonomously” in 2025. That is a measure of activity, and activity is not a saving.
• RAN optimisation. T-Mobile and Ericsson ran AI link adaptation on live traffic at about 43 sites and reported spectral efficiency up to 10% and throughput up to 15% against “rule-based methods”. Orange changes 20 million radio parameters a month automatically. We rank RAN optimisation third because a cluster trial reads in weeks on counters the operator already holds.
• Field dispatch. Nobody in the documents we read publishes a fall in truck rolls with a sample behind it. Orange says AI diagnostics “prevented 180,000 maintenance callouts in France”, with no year and no base. Indosat’s field copilot raises “field productivity by 20% to 30%”, by its own account.

Reported TM Forum on Indosat (3 Sep 2026) and China Mobile with Huawei (3 Sep 2025); Light Reading on Verizon (24 Jun 2026), T-Mobile and Ericsson (22 May 2026) and Orange (15 Feb 2024); Orange AI page

The rest of the list

• Fraud and scam. BT stopped more than 2,430,000 scam calls in four months for its 2.5 million Digital Voice customers, which is about one call each. Bell Canada’s machine-learning risk score gave a “10% increase in fraud detection”. The catch is that a blocked call is a volume, and a volume is not a loss avoided.
• Churn and offers. China Unicom reported churn down 36% in 2020. Safaricom, with Huawei, doubled conversion on targeted packages for about 3 million subscribers. Neither published a control, so neither result can be separated from whatever else was happening to those customers at the time.
• Software engineering. BT’s first four months with CodeWhisperer produced over 100,000 lines of code with 37% of suggestions accepted. Verizon took Claude Code from fewer than 500 users to 33,000 in about six weeks. Both figures measure adoption, and neither operator has published what the code did for it.
• Procurement. BT put “around £2bn of spend” through an AI sourcing platform and cut simple purchases “from days and weeks to seven minutes”. Process time is published. Realised savings are not.

Reported BT newsroom on scam calls (3 Oct 2024) and CodeWhisperer (20 Feb 2024); TM Forum on Bell Canada (20 Oct 2021), China Unicom (23 Jun 2020), Safaricom (22 Oct 2025) and BT Sourced (6 Apr 2023); Light Reading on Verizon (24 Jun 2026)

One result belongs beside the ranking rather than in it. Vodafone’s net benefit of over €100 million in FY26, which part 2 describes, came with “hundreds of use cases” turned off along the way. A portfolio ranked by the cost of finding out is built to stop the same use cases earlier, before they have cost much. Part 5 sets out the rules for doing that.

The terms, briefly

• Cell sleep. Switching off a carrier, a cell or parts of a radio when traffic is low. 3GPP’s Release 18 and 19 features let more of a cell sleep; the decision about when is the vendor’s or the operator’s.
• Containment. The share of contacts an automated channel resolves without a person.
• First-time resolution. The share of problems solved in the first contact.
• SON. Self-organising networks: software that tunes radio parameters automatically.
• Holdout. Customers kept out of a treatment so that its effect can be measured against them.