AT&T AI Strategy: Small Models at Scale and a Smaller Workforce

AT&T uses AI mainly to run a leaner company, while it builds the fiber network it says the AI era will need. More than 100,000 employees have its in-house AI tools, its agents increasingly run on small, cheap models, and AI is part of a $4 billion cost savings commitment. Its technology chief says headcount will not be the same in five years. AT&T does not say how much of its savings come from AI.

AT&T has no separate AI strategy. Its plan centers on fiber and 5G, and CEO John Stankey says: “Our asset base positions us well for the AI era, with a leading position in fiber – the best technology available” (October 2026). He wants AT&T to be seen as a “dramatically different” company by the end of the decade (Wired, September 2026).

Inside, AI is a cost lever. AT&T says it expects AI to contribute to its $4 billion cost savings commitment by the end of 2028 (responsible AI page, March 2026), and its annual report lists AI-driven efficiencies in network design and operations, software development, sales, marketing, customer support and general costs (2025 10-K).

The focus has four parts.

• Network: AI to choose sites for new cell towers, find maintenance issues, and adjust network settings during extreme weather through a generative AI system called GeoModeler (Wired)
• Customer care: AI agents that handle customer service requests across back-end systems (PYMNTS, February 2026)
• Fraud and robocalls: an AI receptionist in the network that screens unknown callers, in testing since mid-2025 (February 2026)
• Internal work: Ask AT&T, its own generative AI platform, and Ask AT&T Workflows, a no-code agent builder (VentureBeat, February 2026)

A central data and AI team owns the platform. AT&T’s Chief Data and AI Officer reports to the Chief Technology Officer and sets the company’s standards for data and AI (AT&T). Andy Markus, chief data officer, leads that work (VentureBeat, February 2026). Chief Technology Officer Jeremy Legg oversees the technology platforms (Wired, September 2026).

AT&T bets on many small models. Markus said: “I believe the future of agentic AI is many, many, many small language models.” After its AI use reached 8 billion tokens a day, AT&T rebuilt how its agents work together around small language models, cutting costs by up to 90%. Use has since grown to 27 billion tokens a day, and Markus keeps a person in charge: “Things do happen autonomously, but the human on the loop still provides a check and balance” (VentureBeat).

Training comes before access.

• Required: every employee must complete an AI ethics and policy course before using internal generative AI tools (AT&T, March 2026)
• Completed: more than 50,000 employees had finished formal AI training by July 2025 (PhoneArena, citing Forbes)
• Tools: 20,000 Microsoft Copilot licenses in 2025, an Ask AT&T Academy with three levels, and an ambassador network, The GenAI Collective (AT&T)
• Use: Ask AT&T Workflows reaches more than 100,000 employees, and more than half use it every day (VentureBeat)

The workforce keeps shrinking. Headcount fell from about 160,700 at the end of 2022 to 133,030 at the end of 2025 (2025 10-K), and to nearly 131,000 by June 2026. Legg said: “We’re not going to have the same headcount in five years as we do today.” A person familiar with the matter told Wired the internal target is about 85,000 by 2030, which AT&T called inaccurate. Middle management and junior developer roles are the most exposed (Wired), the pattern behind how AI is breaking middle management. AT&T is also hiring: thousands of technicians in 2026 for its fiber and 5G build (Reuters via Allwork, March 2026), and people to build and govern AI agents (Wired). In its third-quarter 2026 results, it reported a roughly $800 million employee separation charge (Q3 2026 call).

Governance has a review board. A Data and AI Governance Council and Review Board, with leaders from technology, legal, security, privacy, compliance and the business, reviews high-risk AI use cases before they go live. Every production use case is logged in a central registry, and the board’s committees receive regular reports on AI. Three principles guide the work: by people, for people; responsible; and secure and ethical (AT&T, March 2026).

AI fluency, coded against the Lead with AI Fluency Matrix (September 2026 edition):

• Fluency floor: Level 1 (Chatting), training required before access. Every employee must complete an AI ethics course before using internal generative AI tools, and more than half of the 100,000-plus employees with Ask AT&T Workflows use it daily. Training and use show a baseline, not a skill standard.
• Highest level evidenced: Level 3 (Integrating), company-claimed. Teams of agents find network faults, open tickets and propose fixes, and agents handle customer service requests. A person still checks the results, so Level 4 is not evidenced.
• Portfolio design: Platform-led and broad. A central data and AI team runs one platform for most employees, with a no-code agent builder, and builds network and customer care agents on small models.

AT&T folds AI into its cost savings. All figures below are company-claimed:

• Savings: generative AI “helped to contribute” to $1 billion of total cost savings in 2025, and AT&T expects AI to contribute to a $4 billion commitment by the end of 2028 (AT&T, March 2026)
• Return: every dollar invested in generative AI returns two, according to Markus (July 2025)
• Productivity: active users of Ask AT&T Workflows report productivity gains of up to 90%, and a data product that would have taken six weeks was built in 20 minutes (February 2026)
• Cost of AI itself: up to 90% lower after moving to small models (February 2026)

The AI share of savings is not reported. AT&T gives no figure for how much of its savings come from AI rather than from retiring copper, vendor deals or job cuts. Separating the two is the step from AI adoption to AI value.

Next markers: progress on the $4 billion cost commitment, headcount at the end of 2026, and whether AT&T lets customers’ own AI agents manage their accounts, which Stankey said it is “spending a lot of effort and time on” (October 2026).

Change log

• December 2024: AT&T sets a target of more than $3 billion in run-rate cost savings by the end of 2027 (AT&T release).
• July 2025: More than 50,000 employees have completed formal AI training.
• Late 2025: Ask AT&T Workflows, a no-code agent builder, launches.
• February 2026: AT&T reports up to 90% lower AI costs from small models.
• March 2026: AT&T says AI will contribute to a $4 billion cost commitment by 2028 and plans to hire thousands of technicians.
• September 2026: The CTO says headcount will not be the same in five years.
• October 2026: AT&T reports a roughly $800 million employee separation charge.

• Disclosure: 2 of 4. AT&T answers where it will focus and what it will take, with unusual detail on its platform, training and governance. Where it is going is answered only in part, since AI sits inside a fiber and cost plan, and how it will know is answered in part, since AI is folded into total cost savings.
• AI is a cost lever first. The customer story is the fiber network for the AI era; the internal story is savings.
• Its engineering choice stands out: many small models instead of a few large ones, which it says cut its AI costs by up to 90%.
• Training is a condition of access, and more than half of the 100,000-plus employees with its agent builder use it daily.
• The say-do gap is the workforce. AT&T talks about AI helping people work better, while its technology chief says headcount will not be the same in five years, and it does not say how much AI contributes to the cuts.
• Watch the AI share of savings. The test is whether AT&T separates AI from the rest of its $4 billion cost commitment.

What the numbers show: Revenue rose 4% from 2022 to 2025 while headcount fell 17%, so revenue per employee rose 26%, to about $945k. AT&T has cut jobs for years as it moves from copper to fiber, and says generative AI helped deliver $1 billion of cost savings in 2025, without saying how much came from AI. These figures do not show that AI caused any change.

Revenue, headcount and revenue per employee, FY2022 to FY2025

• FY2022: $120.7bn revenue, 160,700 employees, $751k per employee
• FY2023: $122.4bn revenue, 149,900 employees, $817k per employee
• FY2024: $122.3bn revenue, 140,990 employees, $868k per employee
• FY2025: $125.6bn revenue, 133,030 employees, $945k per employee

Change, FY2022 to FY2025: revenue up 4%, headcount down 17%, revenue per employee up 26%.

Notes: Revenue is total operating revenues, as reported. Headcount is approximate, at year-end. AT&T employed nearly 131,000 people at the end of June 2026, after about 2,100 more job cuts in the first half of the year, according to Wired.

These figures give context on scale and productivity. They do not show that AI caused any change. Sources: AT&T 10-K filings.

What is AT&T’s AI strategy?

AT&T has no separate AI strategy. It uses AI to cut costs in its network, customer care and back office, and it expects AI to contribute to a $4 billion cost savings commitment by 2028. Externally, it positions its fiber network for the AI era.

‍

What is Ask AT&T?

Ask AT&T is AT&T’s internal generative AI platform. Its no-code agent builder, Ask AT&T Workflows, reaches more than 100,000 employees, and more than half use it every day.

‍

Is AT&T cutting jobs because of AI?

AT&T’s headcount fell from about 160,700 in 2022 to about 131,000 in mid-2026, and its CTO says it will not be the same in five years. AT&T has not said how much of that comes from AI rather than from retiring copper and other changes. It called a reported internal target of 85,000 by 2030 inaccurate.

‍

Why does AT&T use small language models?

AT&T says small models are about as accurate as large ones for its tasks and much cheaper to run. Moving its agents to small models cut its AI costs by up to 90%, according to chief data officer Andy Markus.

‍

How is this profile put together?

Lead with AI profiles use public sources only, such as annual reports, earnings materials, filings and on-record interviews, dated at the time of writing. AI fluency is coded against the Lead with AI Fluency Matrix. Company figures are marked as company-claimed unless they have been independently verified.