AT&T Puts Nearly 500 Agentic AI Workflows Into Production
- Cargo
- AI Tools & Apps
- Shipped
- Shipper
- Paul Gillin
- Readers
- 417
AT&T processes 45 billion AI tokens daily and plans to triple its production agentic workflows, with human review and risk based controls.

AT&T is using agentic AI to change how work moves across the business, with nearly 500 workflows already in production and a plan to triple that number next year. The scale is notable: the company processes about 45 billion AI tokens a day, while its agent platform handles more than 1 million transactions per month. These figures point to an effort focused on redesigning operational processes, rather than adding AI assistance to isolated employee tasks.
Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, described the approach at an Amazon Web Services event in Boston earlier this month. A team called Business Agent Foundry, established in early 2025, works with business units to examine how processes run. It maps the steps, handoffs, and applications involved, then looks for changes that can remove unnecessary work or dependence on older systems.
Redesigning workflows around existing systems
One early project connected agents to AT&T’s mainframe environment. For changes to customer fiber installation orders, employees had previously retrieved tickets from queues, edited records manually, and returned the tickets to the workflow. Automating parts of that sequence reduced delays and sped up service delivery.
Corbin described the connection between newer AI systems and older computing infrastructure as “some of the most advanced technology talking to some of the oldest that we have.” That integration matters because a workflow can remain slow even when its decision making improves if people still have to move data between legacy applications by hand.
AT&T also cut the time needed to provision dedicated internet connections by half in buildings where fiber was already installed. The example shows the kind of bounded process where an agent can coordinate work and data that already exist inside a business. The reported reduction applies to that provisioning task; the account doesn’t provide a broader average for time saved across all workflows.
The company’s architecture uses a relatively small set of agents to carry out business workflows, with business logic managed centrally. Retrieval agents find relevant knowledge, and planning agents select the actions, tools, and data needed to complete a task. Corbin said the business logic is stored “in context or in memory.” This design gives teams a common place to maintain process rules while agents handle different tasks using those rules.
Human review and performance thresholds
AT&T distinguishes between personal productivity tools and agents that affect customers or materially change business operations. Employees can create personal agents independently. Customer facing workflows, however, must go through the foundry and receive security, privacy, and compliance reviews.
Production workflows start with human oversight. According to Corbin, an agent can move to autonomous operation only after its accuracy reaches the mid to high 90% range and the task falls within an acceptable risk profile. AT&T also monitors workflows after launch to check performance and detect drift. The accuracy threshold is therefore one part of the release decision, alongside the consequences of an error and ongoing monitoring.
“We’ve democratized agents to the point where people can use them for personal productivity, but not to where you would drive business workflows that could touch your customers or materially have an impact on your business,” Corbin said.
The company tracks labor savings, revenue generation, operating costs, and customer satisfaction as measures of success. Corbin said production AI investments are achieving fivefold returns within a year. He didn’t specify how that figure is calculated or how it varies across projects, so it offers a company reported measure rather than a detailed comparison between individual workflows.
Data and employee adoption shape the rollout
Agents depend on data that is reliable enough to support their decisions. AT&T has had to consolidate fragmented data and make connectivity work across data centers, cloud infrastructure, and employees working on mobile devices. Corbin linked the fragmentation to the company’s history of mergers and acquisitions: “You can imagine what our data infrastructure looks like.”
The work also intersects with decades of business software and process knowledge. Corbin described AT&T as a 151 year old company that has used computers since they were invented, with applications that can be costly to maintain and require employees to learn. In his view, simpler generative workflows can reduce some of that burden. The practical case depends on whether a redesigned process can reduce maintenance and training costs without compromising the service it supports.
Employee input has helped the rollout by surfacing processes that people understand well but have found frustrating. Corbin said inviting staff to bring forward ideas can ease concerns about replacement by giving them a role in changing the work. Sales prospecting is one example of how those changes may affect day to day tasks: agents check prospective customers’ locations and fiber availability before sales representatives spend time pursuing an opportunity.
From process improvement to new services
AT&T is now looking beyond streamlining existing operations. Corbin said the company wants to use agents to launch services, enter markets, and support new sales channels. Those goals would extend agent use from internal process changes into work that can affect how the company reaches customers. The source provides no performance results for these plans yet.
Corbin also sees the foundry as a temporary concentration of expertise. In the long run, he said, responsibility for AI driven transformation should spread across the organization: “What we do should be completely distributed.” That ambition will depend on business teams being able to identify suitable workflows, maintain shared process logic, and operate within the review and monitoring practices now handled by the foundry. The next useful signal will be whether AT&T can expand its workflow count while sustaining the accuracy and service results it reports for production systems.