What operators say AI saves

There are three kinds of AI claim, and where it was said decides how hard you can hold it

When an operator talks about AI savings it is making one of three kinds of claim, and it helps to know which. The first is a programme total in which AI is one named lever. The second is an AI-specific operating figure, such as the share of contacts a bot handled. The third is a figure for AI revenue. Any of the three can be made in a filing, on an earnings call or to the press. In our experience the venue says more about how hard the number can be held than the number itself does.

A filing is the company’s document of record, and anything in it can be traced to the accounts. A call is on the record too, but it reaches most readers through third-party transcripts, and what gets said on it is a forecast as often as a result. A release or an interview is attributed and nobody checks it. Sort the statements of eleven operators that way and what you see is lopsided. The large numbers are programme totals, and the AI-specific ones are small, spoken or both (Exhibit 1).

The 10-Ks tell the same story from the other side. AT&T, Verizon and T-Mobile mentioned AI 11 times between them in their 10-Ks for 2023 and 56 times in the ones for 2025, five times as often. Not one of the nine documents attaches a dollar amount to any of those mentions (Exhibit 2). AT&T’s 10-K for 2025 repeats its 10-K for 2024 almost word for word. It expects “cost savings through AI-driven efficiencies in network design and operations, software development, sales, marketing, customer support services and general and administrative costs”. Verizon’s, for its part, adds a caution. There “can be no assurance that the usage of AI will meaningfully enhance our products or services or be beneficial to our business, including our efficiency or profitability”.

The big numbers are programme totals, and AI is one lever among several

Every large number an operator has published is a programme total. Each lists AI beside other levers, and not one says how much of the total AI is supposed to deliver.

• AT&T. AT&T first tied AI to cost in its 10-K for 2023, where it expected “cost savings through AI-driven efficiencies in our network design, software development and customer support services”. By the middle of that year it had already met a $6 billion run-rate savings target. Its investor-day release of 3 December 2024 then set “$3 billion+ in run-rate cost savings by the end of 2027” and said nothing about AI at all. On 28 January 2026 the chief financial officer said AT&T had “achieved over $1 billion of cost savings in 2025” and expected “an additional $4 billion annual cost savings by the end of 2028”. In April, in the one transcript we found, he listed the levers: workforce optimisation and vendor rationalisation, AI enablement, digitalisation and lower legacy support costs. Four levers, no split.
• Verizon. Verizon’s target is the largest of the three, and it is the plainest about what is in it. On 30 January 2026 its chief executive described “an in-year war chest of $5 billion in OpEx savings, with a substantial portion realized by headcount reductions”. The rest of his list was “marketing efficiencies, real estate rationalization, contract renegotiations and more”. The chief financial officer added “reducing call volumes, or on the IT side, in terms of rationalizing platforms and including AI enablement”, and gave no AI share. The 8-K on the 13,000 job cuts does not mention AI.
• T-Mobile. T-Mobile is the one carrier with AI in the headline of a furnished filing. Its results release of 11 February 2026 promises “a nearly $3 billion incremental contribution to Core Adjusted EBITDA expected from digitalization and AI, by the end of 2027, relative to 2025”. Read on to the guidance section of the same release and the credit goes to “significant 2025 investments in greenfield network expansion, AI, and digitalization”, so the attribution shifts inside a single document. On the call, in the one transcript we found, the chief financial officer put the savings at $1.3 billion in 2026 and $2.7 billion in 2027, both against 2025.
• BT. BT’s filed number is its cost programme: £1.5 billion realised so far and an extension to FY30 “to enable the delivery of £3.7bn of cost savings”. Beside it is the line that “AI will continue to change how we work and the types of roles we need”. BT’s AI figure is older, and it was never filed. On 18 May 2023 AP, The Guardian and RCR Wireless reported that about 10,000 of the up to 55,000 roles to go by 2030 would go “through digitization, automation and the use of AI”. The Annual Report 2026 repeats neither number.
• Telefónica. Telefónica’s capital markets deck of November 2025 lists an “Autonomous Network Journey” and a “B2C call centers AI virtual assistant” among the levers behind €2.3 billion of gross efficiencies by 2028. There is no split by lever. Its exit plan for about 5,500 employees does not name AI.

Filed AT&T call of 28 January 2026 and Verizon call of 30 January 2026, each in two transcripts (MarketBeat and The Motley Fool) · AT&T and Verizon 10-Ks and 8-Ks · T-Mobile 8-K exhibit of 11 February 2026 · BT Annual Report 2026 · Telefónica 6-Ks of 4 November and 22 December 2025

Reported AT&T’s list of levers, call of 22 April 2026, and T-Mobile’s $1.3 billion and $2.7 billion, call of 11 February 2026, each in one third-party transcript (MarketBeat) · BT’s 10,000 roles, AP, The Guardian and RCR Wireless, 18 May 2023

Each American target can be set against the cost pool in part 1, and doing so puts them in proportion. Verizon’s $5 billion is 8.1% of its 2025 pool of $61.6 billion. AT&T’s $4 billion is 7.4% of $54.4 billion. T-Mobile’s $1.3 billion for 2026 is 3.7% of $35.0 billion, and its nearly $3 billion for 2027 would be 8.6% (Exhibit 3).

Every programme total the eleven operators have published bundles AI with something else, whether headcount, property, contracts, legacy shutdowns or digital channels. Not one says how much of it is AI. So a reader can hold AT&T to $4 billion and Verizon to $5 billion, and nobody can hold either to an AI figure, because neither has given one.

The figures that credit AI were said on calls, and only one made it into an annual report

Only a handful of figures credit a result to AI by name, and all but one of them were said out loud on an earnings call rather than written into a report.

Filed Verizon call of 27 April 2026, MarketBeat transcript, with RCR Wireless (28 April) and Light Reading (30 April) carrying both figures · Telstra annual report, year to 30 June 2026

Reported T-Mobile call of 28 April 2026, one third-party transcript (MarketBeat)

The most specific statements by an American carrier are Verizon’s of 27 April 2026, and every one of them is missing either a period or a base (Exhibit 4). Its chief executive said Verizon had “over $200 million of energy savings as a result of deploying AI into the network”, and that “85% of all of our issues right now are autonomously resolved”. He did not say which issues, out of how many, or over what period. Is $200 million a lot? Set it against Verizon’s own meter and it turns out to be a surprisingly large share of what it uses, which is what makes the missing period matter.

The other two figures want the same treatment. T-Mobile’s 60% is one third-party transcript’s rendering of what its chief financial officer said, with no definition of what counts as contained. Its filed care figure, “more than 50% reduction in calls to care since 2021”, runs from three years before its OpenAI platform was announced in September 2024, so the two cannot be matched up. Telstra’s 34% is the one AI share in a filing, and it is a share handled rather than a saving. The same report shows labour expenses down 4.6% and underlying operating expenses down A$454 million, or 3.0%, and attributes neither to AI.

Press figures are weaker still. AT&T’s chief financial officer told Fortune in June 2023 that AI had cut employee effort by nearly 17 million minutes a year, worth “hundreds of millions of dollars”. In the same breath he added that the saving was not necessarily related to generative AI. The “up to 90% cost savings” its data chief described to VentureBeat in February 2026 turns out to be the cost of running its internal assistant. That is a different thing from a saving in the operation. The most useful number of the lot comes from Vodafone. Its group chief technology officer told TM Forum Inform that Vodafone “recognized over €100 million net benefits from AI in the 2025/26 financial year” and had “turned off hundreds of use cases” along the way. We like that figure for two reasons. It is net, and it comes with a record of stopping things. It does not appear in the 20-F for the year.

The accounts cannot test the 2026 claims until February 2027

A gross saving only shows as a fall in the pool when nothing else is growing, and in 2026 acquisitions and special items are growing faster than any saving could show through. The first half of 2026 shows why the test has to wait (Exhibit 5).

Verizon’s SG&A rose 5.9% in the first half of 2026. Inside that line is a $746 million loss on a business held for sale and $397 million of new severance, while its Business segment cut personnel costs by $294 million “related to workforce reduction initiatives”. The saving is there. It is swamped by the rest. T-Mobile’s SG&A rose 8.4% against revenue up 9.2%, with UScellular in the base for the whole half. AT&T’s other cost of revenues fell 1.4% and its SG&A rose 3.2%.

Until then the only AI operating figure in any of these operators’ annual reports is Telstra’s share of enquiries. Not one of the eleven has published an AI-specific saving with a period, a base and a filing behind it, and we would like to see one.

The terms, briefly

• Run-rate saving. The annual saving a programme delivers once it is fully in place. A company can reach it before a full year of it shows in the accounts.
• Core Adjusted EBITDA. T-Mobile’s earnings measure before interest, tax, depreciation and amortisation and its stated adjustments. A contribution to it can come from cost or from revenue, which is why we read its $3 billion with care.
• Containment. The share of customer contacts an automated channel resolves without a person. No operator here has published its definition.