AT&T is all-in on agentic AI

AT&T is no stranger to AI. It has invested heavily in its traditional forms over the years, leveraging it for analytics, workflow automation, and voice assistants. But Andy Markus, the company’s chief data and AI officer, says full commitment to agentic AI is the key to move AI from the information economy into the action economy.

Even though earlier uses of AI have affected operations throughout the company’s business, he adds, gen AI and agentic AI will be even more transformative. Agents, or autonomous assistants, as AT&T prefers to call them, don’t just generate content, they can plan and execute tasks from beginning to end. While agents still turn top human guidance and intervention when needed, they combine goal understanding, multi-step planning, and tool orchestration into a single package.

“There’s not a single person across AT&T that’s not impacted by this,” Markus says.

An early use case, first deployed in a limited fashion last September, is the AT&T digital receptionist, a network-based agentic AI to fight spam calls and fraudsters. While the company already provides spam filters like AT&T ActiveArmor, the AT&T digital receptionist can engage with callers directly to determine whether they’re suspicious. It can then disconnect suspicious calls or take messages if needed. Customers can also watch a live transcript of the digital receptionist’s interactions with callers and pick up at any time.

Agents as colleagues

AT&T has also begun using agentic AI tools to assist its employees. In 2023, working closely with Microsoft, it introduced a version of ChatGPT called Ask AT&T, a gen AI platform for everything from answering employees’ HR questions and translating customer and employee documentation, to helping software developers write code.

In November, AT&T upped the ante when it worked closely with Microsoft using Azure to create Ask AT&T Workflows, a graphical drag-and-drop agent builder to help teams build agents to automate time-consuming tasks.

The first in-production tool built with Ask AT&T Workflows was created jointly by the AT&T business and technology development teams. These agents take customer service update requests, synchronize data across systems, and auto-install information in real time. Markus says the tool has already proved its value at scale, improving the customer experience while reducing wait times, and allowing employees to shift their focus to higher priority actions.

In addition, AT&T network engineers can now use agents to help them resolve network alerts and get customers reconnected after an incident. Agents can correlate telemetry to identify where an alert was issued, pull recent change logs, check for known issues, and open a trouble ticket. Another agent could then propose a resolution and write new code for a patch. After the issue is resolved, a third agent could compile artifacts and a summary that the engineer could use to plan preventative measures to keep that incident from happening again.

Sequencing events

AT&T funnels all potential agentic AI and gen AI use cases through a program office called the Generative AI Transformation Office, which considers the potential value of those use cases for customers and the company. That helps it prioritize its efforts, and each use case gets a formal business case tied to the CFO’s office.

“Those use cases are what’s driving the value, and that value is a two-times return on investment,” Markus says. “We expect it to grow significantly this year, and we’re seeing it already.”

He’s also looking for a free cash-flow impacting benefit in most cases, though he says there are certain use cases done because they serve the customer better. “If there’s a direct customer benefit, that’s going to get something approved,” he says.

Markus adds that as the company’s work around gen AI and agentic AI has matured, the use cases have become more complex.

“Now we’re looking at entire workflows, entire processes where decisions have to be made throughout that process,” he says. “That’s where agentic AI shines.”

The value of fine tuning

His team designed Ask AT&T Workflows with both pro-code and no-code paths. The no-code path offers the drag-and-drop capabilities, while the pro-code path provides much finer control.

“Typically in a workflow, what we have is a master agent or control agent, and smaller agents that are making the decisions in the process,” Markus says. “In general, we like to have the smaller agents use a small language model because their tasks are very defined. At the scale we’re operating in, we’ve been very successful at fine tuning small language models to be as accurate as possible, and accuracy is the driver of everything.”

He adds that a human always oversees the chain reaction of agents, which has always required human checkpoints throughout any process. The agents are able to pull information and data from Ask Docs and Ask Data, AT&T’s proprietary gen AI tools for housing the company’s information. Each action taken by an agent is logged, and data isolation, retention policies, and role-based access are enforced whenever one agent hands off a workload to another.

Markus says one of the keys to the success of AT&T’s agentic AI efforts has been getting the business excited by showing them what’s possible.

“We have hundreds of business cases that are waiting for us to prioritize,” he says. “That’s the result of the enthusiasm and passion of the business because they’re seeing the value here.”