AT&T deploys autonomous agents at scale with supervision and precision criteria, and reports fivefold returns on its projects in production. The operator combines governance, data consolidation and centralized logic to transform its processes, with a plan to expand workflows starting next year.
Thresholds of 90 to 95% and an announced ROI times five
Each production workflow starts with human supervision and only operates autonomously after achieving accuracy between 90% and 95% within a risk profile deemed acceptable. Performance is continuously evaluated in order to identify any possible deviation. The success criteria focus on reducing labor requirements, revenue generation, operating costs and customer satisfaction levels. Luke Corbin says investments in production AI generate returns five times greater over the year.
Centralized governance and supervised use of agents
Workflows that may impact customers go through the Business Agent Foundry and security, privacy, and compliance reviews, while employees can create agents for personal productivity. According to Luke Corbin, the use of agents is democratized for individual efficiency but cannot manage workflows with material impact without control. Created in early 2025, the Business Agent Foundry maps processes with industry frameworks and targets the elimination of unnecessary steps, handoffs and dependencies on legacy systems. Corbin believes that in the long term, transformation responsibilities should be distributed throughout the organization, to the point where there is no longer a need for a dedicated team.
Simple agent architecture and business logic stored in context
At AT&T, the architecture relies on a small number of agents responsible for carrying out workflows, with business logic centralized and persisted in context or memory. Some agents are responsible for retrieving relevant knowledge, while others, dedicated to planning, choose the required actions, tools and data. Feature development required bringing together scattered data and ensuring a stable connection between data centers, clouds and mobile teams in an environment where mergers and acquisitions have been rife.
Reinvented processes: from mainframe to provisioning halved
AT&T automated interactions with its mainframes by removing tasks previously handled manually, reducing delays and speeding up service delivery. Luke Corbin underlines the interest in bringing recent technologies and old systems into dialogue. In another use case, the time to commission dedicated Internet connections was reduced by half in buildings already fiberized. In prospecting, agents validate locations and fiber availability to target promising customers before representatives intervene.
Change of scale and ambition for agentic workflows
Nearly 500 agentic workflows are in production and AT&T plans to triple the number next year. According to Luke Corbin, the company processes approximately 45 billion AI tokens per day and its platform handles more than 1 million transactions per month. Many workflows already operate autonomously and cover tasks like customer provisioning and sales prospecting. The approach favors the reinvention of complete processes, after a rapid transition from experimentation to the systematic identification of redesign opportunities. AT&T is now directing its efforts toward launching services, entering new markets and supporting new business channels. A 151-year-old company relying on legacy systems that are expensive to maintain and train, AT&T believes that simpler generative workflows can offset some of these costs. Involving employees in coming up with ideas helped limit resistance related to fear of replacement and was facilitated by the ability to eliminate inefficient processes that many were familiar with but disliked. Luke Corbin demonstrated this approach at an AWS event in Boston. The whole thing is part of a desire to rethink how the company operates via autonomous agents.
