What AT&T said about its agentic AI use cases
AT&T has nearly 500 agentic processes in production and plans to triple that number next year, SiliconANGLE reported on 14 November 2026.
The figures come from Luke Corbin, assistant vice president for digital transformation and AI at AT&T Business. He spoke at an Amazon Web Services event in Boston in early November 2026. SiliconANGLE (14 November 2026) quotes him: AT&T processes about 45 billion AI tokenstoken. A token is a piece of text that a language model works with: a word, part of a word, or a symbol. a day, a token being a small piece of text that a model reads or writes. The carrier’s agentic platform handles more than 1 million transactions a month, he said.
The main shift, as Corbin describes it: AI at AT&T no longer just speeds up employees, it changes how the process itself runs. AT&T is 151 years old, SiliconANGLE notes in the same piece. After many mergers its data is scattered, and it still runs a lot of old systems, mainframes included.
Rebuild the whole process, not single steps
AT&T rebuilds whole processes instead of automating single steps, SiliconANGLE writes. Business Agent Foundry, an internal team set up in early 2025, does the rebuilding. It maps each process end to end against industry frameworks and removes extra steps, handoffs between departments and dependence on legacy systems. It does not collect one-off automation ideas from staff.
The carrier keeps the number of agents small and stores business logic in one place, in shared context or memory, SiliconANGLE writes. Retrieval agents find the knowledge a task needs, and planning agents choose the actions, tools and data. For a manager, I think the point is simple: a process rule changes once, not inside every agent.
The examples Corbin gave in Boston, as SiliconANGLE reported on 14 November 2026, come from the daily operations of a telecom operator:
- agents work with mainframes when a customer changes a fiber installation order;
- dedicated internet setup in buildings that already have fiber is now twice as fast, according to Corbin;
- agents check the customer’s address and fiber availability before a salesperson takes the deal.
I see the main lesson of the case here: an agent is likely to pay off soonest where old applications are expensive to keep and to train people on. Corbin himself, SiliconANGLE writes, believes his team should eventually not need to exist.
When AT&T lets an agent work without a human

An AT&T agent runs on its own only after its accuracy reaches the mid- to high-90s percent range, SiliconANGLE reported on 14 November 2026. The risk of the process must also stay within an acceptable profile. Until then, a human oversees every process. After that, the system monitors quality and drift, meaning whether the agent starts to make more mistakes over time.
Processes that touch customers first pass a review for security, privacy and compliance. Employees can build personal agents for their own tasks themselves.
What the AT&T figures leave out
AT&T did not disclose how it measured its figures: all of them come from Corbin alone, at a vendor event hosted by Amazon Web Services in November 2026. None of them has been checked independently.
The boldest one: Corbin said AI investments in production return five times their cost within a year, SiliconANGLE reported on 14 November 2026. The carrier did not say how it measured this or against what base. The same goes for the twice-as-fast setup. Four more questions stay open:
- how many of the nearly 500 processes run fully without a human: the article only says many;
- what tripling next year means: about 1,500 processes by the end of 2027, or another timeline;
- which models and platforms the agents run on: the article does not say whether the platform runs on Amazon Web Services;
- how the agents have affected headcount.
So I read the fivefold return as a company claim, not as a benchmarkAI benchmark. To benchmark AI models means to compare them on one standard set of tasks with known answers. That set is called an AI benchmark. for your own business case.
What other large companies can take from AT&T
In my view, the AT&T case suggests a workable order for rolling out agents in large companies with old systems. I think it fits telecoms, banks, retailers and logistics firms most closely, where old systems and scattered data are common. I led software development for Walmart, Disney, Apple and Microsoft for 20 years, and these are the six steps I take from the case:
- Plan for scattered data before the first agents: AT&T copes with data scattered by mergers, and a company built through mergers is likely to face the same.
- Pick a whole process, from request to result, and remove extra handoffs between departments instead of speeding up each step on its own.
- Set up a dedicated team that leads the rollout together with business units, and keep business logic in one place, not inside every agent.
- Write into your internal policy the accuracy threshold after which an agent runs without a human, and who watches quality after that.
- Look for the money where legacy support, manual work on request queues and sales time on weak deals cost the most.
- If an agent sees customer data, check where that data is stored and whether it leaves the country under the data protection law that applies to you. In Russia, for example, the personal data law (Federal Law No. 152-FZ of 27 July 2006) requires data on Russian citizens to be collected into databases inside Russia (Article 18, part 5). It also sets rules for transfers abroad (Article 12).
I would start with a process where an agent’s mistake is cheap to fix and the result is easy to count in hours and money. That gives you your own numbers, and you no longer need someone else’s claim of a fivefold return.


