AT&T Runs Nearly 500 AI Agent Workflows in Production, Plans to Triple Them Next Year

The company now processes about 45 billion AI tokens a day across its operations, and its agent platform handles more than 1 million transactions a month, Corbin said. Many of the workflows run autonomously, covering work from customer provisioning to sales prospecting.

Corbin said production AI investments are returning five times their cost within a year. The report gave no underlying savings or revenue figures, and the return has not been independently verified. AT&T measures success by labour savings, revenue generated, operating costs and customer satisfaction, he said.

The work is run by Business Agent Foundry, a group set up in early 2025 to work with individual business units. Instead of collecting one-off automation ideas from staff, the team maps business processes against industry frameworks and looks for steps, handoffs between departments and legacy application dependencies that can be removed.

One early project connected agents to AT&T's mainframes. Changes to customer fibre installation orders used to require employees to pull tickets from a queue, update records by hand and send them back into the workflow. Agents took over parts of that process, cutting delays. "It was kind of cool having some of the most advanced technology talking to some of the oldest that we have," Corbin said. In another case, AT&T halved the time needed to provision dedicated internet connections in buildings that already had fibre installed.

Sales is another use. Agents check a prospect's location and whether fibre is available there before a representative spends time on the lead.

AT&T has not built large numbers of specialised agents. It runs a relatively small set that carry out workflows using business logic managed centrally. "We store the business logic in context or in memory," Corbin said. Retrieval agents find the relevant knowledge, and planning agents decide which actions, tools and data a task needs.

Governance is split by risk. Any employee can build agents for personal productivity, but workflows that touch customers must go through the foundry and pass security, privacy and compliance reviews. Every production workflow starts with human oversight, and agents are allowed to run on their own only after accuracy reaches the mid to high 90% range within a permissible risk profile. Monitoring continues afterwards to catch drift.

Scaling up has required more than models. AT&T has had to consolidate fragmented data and provide reliable connectivity across data centres, cloud infrastructure and mobile staff. "We've been through a lot of mergers and acquisitions," Corbin said. "You can imagine what our data infrastructure looks like."

Inviting employees to propose uses for agents has helped limit resistance from staff who worry AI could replace them, he said, as has the chance to get rid of processes they disliked.

The next phase moves from improving existing processes to using agents to launch services, enter markets and support new sales channels. Corbin said responsibility for that work should eventually be spread across the company. "In the long run, my team shouldn't exist," he said.