The cost an AI saving is measured against

The pool an AI saving has to come out of was $150.9 billion in 2025, and it has hardly moved

An operating AI can only save money in two lines of a carrier’s income statement. One is the cost of running the network and serving customers, which Verizon and T-Mobile call cost of services. The other is the selling, general and administrative expense wrapped around it. Add the two together for AT&T, Verizon and T-Mobile and you get $150.9 billion for 2025 (Exhibit 1).

The captions differ a little from one carrier to the next, and it is worth knowing how. Verizon and T-Mobile report cost of services; AT&T calls the same line other cost of revenues. All three leave depreciation and amortisation out of it. Verizon’s own definition of cost of services runs from “salaries and wages” through “network access and transport costs” to “computer systems support”. Its SG&A takes in sales commissions, advertising, bad debt and “call center and information technology costs”. Between them, those two lines hold every cost an operating AI could plausibly touch.

Two costs we leave out on purpose. The cost of the handsets a carrier sells rises and falls with device sales, and says nothing about how the operation is run. Depreciation is the other, because it does not change in the year an operating AI is switched on.

Look at the three separately and they went in different directions. AT&T’s pool fell $1.5 billion to $54.4 billion while its revenue rose 2.6%; its other cost of revenues came down 5.8% and its SG&A crept up 0.2%. Verizon’s pool went the other way, up 1.3% to $61.6 billion on revenue that grew 3.1%. Strip out Verizon’s own special items and the picture softens a little, with cost of services down 1.5% and SG&A up 0.8%. T-Mobile’s pool rose most, by 6.1% to $35.0 billion, but its revenue rose 12.4% and its 2025 figures take in UScellular from August of that year (Exhibit 2).

Is that the right pool? One could argue for a narrower one. No operating AI runs advertising or decides who defaults. Take out those two lines, and Verizon’s special items with them, and the pool was $49.0 billion at AT&T, $53.0 billion at Verizon and $29.9 billion at T-Mobile in 2025. Two years earlier the same lines came to $51.3 billion, $53.3 billion and $29.6 billion. The narrower pool tells the same story as the wide one.

So whatever AI has saved so far is buried inside a cost base that held level in dollars and fell as a share of revenue. It may well be in there. But the only way a claimed saving can be checked is as a movement in one of these lines, and that is the test the rest of the series applies.

AT&T and Verizon have 32,370 fewer people, and both say why, and neither says AI

The part of the pool that has moved most is people. AT&T ended 2025 with 133,030 employees, 16,870 fewer than at 31 January 2024. Verizon ended it with 89,900 full-time equivalents, 15,500 fewer than two years earlier. T-Mobile went the other way, from about 67,000 to about 75,000, because it was buying businesses (Exhibit 3).

Each carrier says why, and in every case the reason is a programme rather than a technology.

• Verizon. By the end of March 2025 about 4,800 people had left Verizon under a voluntary separation programme. In the fourth quarter it announced a cut of more than 13,000 positions, and “Over 80% of the affected employees exited in December 2025”. The severance for the year came to $1.5 billion before tax. The 8-K of 20 November 2025 that announced the cut does not mention AI anywhere. What it names is “a significant reduction in outsourced, contracted and other outside labor expense”.
• AT&T. AT&T’s 10-K for 2025 explains lower Business Wireline costs as “primarily driven by lower personnel and customer support costs associated with ongoing transformation initiatives”. Its 10-Q for the second quarter of 2026 cites “lower personnel and other costs resulting from the decommissioning of our legacy network”. The words are transformation and decommissioning. AI does not appear.
• T-Mobile. T-Mobile cut about 5,000 positions in August 2023, “primarily in corporate and back-office functions”. In late 2025 it began what it calls a workforce transformation, taking out layers of management and duplicated roles, and booked a $390 million charge for it.

The pattern holds outside the United States (Exhibit 4). Deutsche Telekom’s German headcount fell 4.9% in 2025, “due in particular to efficiency enhancement measures”. BT’s labour resource, counting contractors, “fell 7% year-on-year to below 108,000”. Telstra cut 1,219 direct roles in FY26 and saw its labour expenses fall 4.6% to A$3,807 million. Vodafone is the exception in the table, with more employees at the end of FY26 than a year before.

Not one of the Forms 10-K, 10-Q or 20-F we read for the series attributes a fall in headcount to AI. Where a filing explains a fall, it names a separation programme, a restructuring or the shutdown of a legacy network, and the American ones put a severance charge beside it. AI may well have played a part. Nobody has filed a document that says so.

No carrier files what customer care costs

Customer care turns up in every carrier’s statement about AI and in none of their accounts as a line. Verizon is the only one that even says where the cost goes: “Aggregate customer service costs, which include billing and service provisioning, are allocated between Cost of services and Selling, general and administrative expense”. So the cost is in the pool, split two ways, and nobody outside the company can see it.

What operators disclose instead are proxies, and they are a mixed bag (Exhibit 5). Vodafone’s Form 20-F is the one filing that counts care staff: 56,922 of an average 92,988 employees in FY26 worked in customer care and administration, which by our arithmetic is 61.2% of everyone it employed. Telstra’s annual report is the one with an AI share of contacts, and it says its AI Assistant has handled 34% of enquiries since March 2026. T-Mobile reports “more than 50% reduction in calls to care since 2021” without saying how many calls that was. The nearest thing to a base is Light Reading’s 2022 figure of 3 million calls a week.

Two things follow. A care saving stated as a percentage has no filed base, because no operator here files either its contact volume or its care cost. And so the care figures in part 2 are shares rather than sums, and they will stay that way until somebody files a base.

Energy is the only cost with a physical unit, and only Vodafone files the bill

Energy is different from every other cost in the pool, because it is counted in watt-hours before anyone puts a price on it. That makes it the easiest saving in the business to check. Every operator in this set publishes its watt-hours. Only Vodafone says what they cost (Exhibit 6).

Vodafone’s figure is useful because it prices the unit. Its €0.9 billion for 5,967 GWh works out at about €150 a megawatt-hour. That is an average across every market it operates in, Africa included, on a spend rounded to the nearest tenth of a billion. The nearest thing to an American price comes from AT&T’s own energy projects. AT&T says they saved $81.1 million a year for 709 million kWh conserved in 2025, which is about 11.4 cents a kilowatt-hour. The same arithmetic gives between 8.7 and 11.4 cents in each year since 2022.

The totals themselves have moved little, which matters for anyone claiming a large energy saving. AT&T used about 6% less energy in 2025 than in 2022. T-Mobile, on a series recast for its acquisitions, used 14.9% less, and the energy it burns per petabyte carried fell from 263 MWh to 86 MWh. Verizon’s assured total for 2025 is 0.2% below its 2022 figure on the same SASB definition, which is about as flat as a number gets. Vodafone went up, from 5,701 GWh in FY24, restated to include Three UK, to 5,967 GWh in FY26. BT came down, from 2,989 GWh in FY25 to 2,678 GWh.

Price surfaces in one place only in an American 10-K. Verizon’s Consumer segment reported a $75 million rise in building and facility costs in 2025, “primarily due to higher utility rates”, and that is all. None of the three carriers publishes its energy bill, so a dollar saving on energy is a claim about a number the reader cannot see. A kilowatt-hour saving is different, because it can be checked against the published total. That is why part 3 ranks energy first.

The terms, briefly

• Cost of services. The cost a carrier attributes directly to providing a service: wages, network access and transport, contracted services and systems support. AT&T’s equivalent line is other cost of revenues.
• SG&A. Selling, general and administrative expense: sales commissions, advertising, bad debt, administration and, in Verizon’s definition, call centres and information technology.
• Operating cost pool. Our name for the two lines added together, which is the base any operating saving is measured against.
• Special items. Charges a company separates from its underlying results, such as severance and asset write-downs. Verizon tabulates them in its 10-K.
• Full-time equivalent. Headcount expressed as full-time positions, so two half-time employees count as one.