peekgenticBook a call

News analysis · AI agents

AT&T makes its AI agents earn their autonomy

AT&T says it runs nearly 500 AI agent workflows, and every one starts with a person watching. Agents only work on their own after accuracy reaches the mid to high 90s. What that rule means for a 100-person company, what doesn't transfer, and how to copy it without a big team.

By Miguel GutierrezCo-founder, Peekgentic · Dallas
· 4 min read
Preview card: AT&T makes its AI agents earn their autonomy.
Every workflow starts with a person watching. Autonomy comes after the accuracy is measured.Peekgentic

In brief

As reported by SiliconANGLE on Nov. 14, 2026, AT&T says every AI agent workflow it puts into production starts with human oversight, and agents only run on their own after accuracy reaches the mid to high 90% range within an acceptable risk level, with monitoring after that. It also keeps a hard line between personal AI helpers, which any employee can build, and workflows that touch customers, which go through security, privacy and compliance review. A mid-size company can copy both rules without AT&T's scale: start every automation with a person approving, measure accuracy on real work, set the bar before you switch anything to automatic, and keep checking after.

  • AT&T says every agent workflow starts with a person watching; agents run alone only at mid to high 90s accuracy.
  • It separates personal AI helpers from workflows that touch customers, which need security, privacy and compliance review.
  • Its scale and fivefold return are AT&T's own claims, made at a vendor event.
  • At 95% accuracy, 100 items a day means five mistakes a day. Set the bar by what a mistake costs.
  • A mid-size company can copy the rules with one owner per workflow and a weekly sample.

What AT&T said

Luke Corbin, an assistant vice president of digital transformation and AI at AT&T Business, described the company's approach at an Amazon Web Services event in Boston earlier this month. SiliconANGLE reported his remarks on Nov. 14. The numbers, as he gave them:

  • Nearly 500 agent workflows in production, with plans to triple that next year.
  • About 45 billion AI tokens processed a day across the company.
  • More than 1 million transactions a month on its agent platform.
  • Fivefold returns within a year on production AI investments.

Those are AT&T's own figures, given on a cloud vendor's stage. Nobody outside the company has checked them. The return figure especially should be read as a claim, not a result.

The more useful part of the talk isn't the scale. It's the rules.

The two rules worth copying

Rule 1: personal helpers and customer workflows are different things. According to the report, any AT&T employee can build an agent for their own productivity. But a workflow that touches customers, or could materially affect the business, has to go through a central team and pass security, privacy and compliance reviews.

That's a line most mid-size companies haven't drawn. An estimator using ChatGPT to draft an email and an agent that sends order confirmations to customers get treated the same, which usually means both are ignored. Our guide on shadow AI at work covers the first kind. The second kind needs an owner and a sign-off.

Rule 2: agents earn autonomy. Every AT&T production workflow starts with human oversight. Per SiliconANGLE, agents only operate on their own once accuracy reaches the mid to high 90% range within a permissible risk profile, and monitoring continues after that to catch drift. That matches what AT&T's chief data and AI officer, Andy Markus, told CIO.com in January: that people oversee its chains of agents with checkpoints throughout. He put it this way: "accuracy is the driver of everything."

Why the number matters more than it sounds

"Mid to high 90s" sounds close to perfect. It isn't, once you count what gets through.

Wrong items per 100, at each accuracy level

Simple arithmetic. At 100 orders a day, multiply by one; at 400, by four.

  1. 90% accurate10
  2. 95% accurate5
  3. 98% accurate2
  4. 99% accurate1

At 95%, an agent handling 100 orders a day gets five wrong every day. Whether that's fine depends on what "wrong" means. A misfiled email is cheap. A wrong ship-to address or a duplicate vendor payment isn't. That's why AT&T ties the bar to a risk profile instead of one number for everything.

The bar for "runs on its own" should depend on what a mistake costs, not on how impressive the demo was.

Why this is news for everyone else

Most companies don't fail at AI agents because the model is weak. Gartner predicted last year that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value or weak risk controls. That's a forecast, not a count. But two of those three reasons are exactly what AT&T's rules address: you can't show value without measuring accuracy, and you can't control risk without deciding what runs on its own.

The same Gartner release also estimates only about 130 of the thousands of vendors selling "agentic AI" have the real thing. If a vendor tells you their agent runs on its own from day one, ask how they measured that it's right.

What doesn't transfer

To be fair, a 100-person company is not AT&T. Some of this won't fit:

  • You don't need a central agent team. AT&T built one because it has hundreds of workflows. You need one named owner per workflow and a log.
  • You won't have 45 billion tokens a day. Your volume is small enough that a person can review a sample every week, which is actually an advantage.
  • Some steps should never go autonomous. Paying a new vendor, changing bank details or sending a price to a customer can stay behind a person forever. Accuracy in the 90s isn't the point there; one mistake costs too much.
  • The return claim may not repeat. AT&T ties its business case partly to the high cost of keeping old legacy applications running. Your gains depend on how much typing and chasing your office does today.

How to copy it this quarter

  1. List your AI uses in two columns: personal helpers, and anything that touches a customer, a vendor, money or your ERP.
  2. Give every item in the second column an owner and a person who approves its output.
  3. Measure on real work. Run 50 to 100 real items with a person approving each one. Count how many needed no edits.
  4. Set the bar before you look at the results, based on what a mistake costs. Write it down.
  5. Go automatic one step at a time, starting with the cheapest mistakes. Keep sampling after, since accuracy can drift when customers change forms or suppliers change layouts.

For the full approval setup, see human-in-the-loop automation. For what happens when agents run without these rules, see when AI agents went off script.

Common questions

What did AT&T say about AI agent autonomy?

As reported by SiliconANGLE on Nov. 14, 2026, AT&T says every production agent workflow starts with human oversight, and agents only run on their own after accuracy reaches the mid to high 90% range within an acceptable risk level, with monitoring after that.

How many AI agent workflows does AT&T run?

Nearly 500 in production, with plans to triple next year, according to AT&T's Luke Corbin as reported by SiliconANGLE. He also cited about 45 billion tokens a day and more than 1 million transactions a month.

Is AT&T's fivefold AI return verified?

No. It's AT&T's own figure, given at an Amazon Web Services event and reported by SiliconANGLE. Treat it as a claim.

How accurate should an AI agent be before it runs on its own?

It depends on what a mistake costs. At 95% accuracy, 100 items a day means five wrong a day. That may be fine for filing email and not fine for payments or customer prices, which can stay behind a person permanently.

Does a mid-size company need an AI governance team?

Not a team. It needs one named owner per workflow that touches customers, vendors, money or the ERP, a person approving the output at first, and a log of what was checked.

Why do agentic AI projects get canceled?

Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value or inadequate risk controls. It's a forecast, not a count.

Sources

  1. SiliconANGLE, AT&T uses agentic AI to reinvent business processes from the ground up, Nov. 14, 2026
  2. CIO.com, AT&T is all-in on agentic AI, Jan. 28, 2026
  3. Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, June 25, 2025

More from Peekgentic

Published November 15, 2026 · Written by Miguel Gutierrez, Co-Founder, Peekgentic (Dallas, TX). Every statistic links its source.