THE ESSENTIALS IN 30 SECONDS
- AT&T says nearly 500 agentic workflows are in production and that it aims to roughly triple that number next year.
- The company reports about 45 billion AI tokens processed per day and more than 1 million agent transactions per month.
- Workflows start under human oversight and run autonomously only once accuracy reaches the mid-to-high 90% range, according to AT&T.
AT&T says it now runs close to 500 agentic AI workflows in production and intends to roughly triple that figure in 2027. The numbers come from Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, who spoke at an Amazon Web Services event in Boston earlier this month, as reported by SiliconANGLE on 14 November. None of the figures has been independently audited; they are the company's own.
How the programme is organised
The work is run through what AT&T calls its Business Agent Foundry, set up in early 2025 to partner with business units. Rather than layering agents on top of existing procedures, the team maps each process against industry frameworks to identify steps, handoffs and legacy dependencies that can be removed. Employees are free to build personal productivity agents, but any workflow that touches customers must go through the foundry and clear security, privacy and compliance reviews.
Architecturally, AT&T describes a fairly small pool of agents driven by centrally managed business logic, instead of a large number of narrow, specialised agents. Retrieval agents locate knowledge, while planning agents decide which actions, tools and data a task needs.
According to the company, the platform processes more than 1 million transactions a month. AT&T's chief data and AI officer had already put the group's overall consumption at about 45 billion tokens a day in a July blog post, which also described an in-house AI gateway that routes each request to a model based on speed, cost and expected quality.
Examples and safeguards
Corbin cited agents that automate parts of the manual process for changing customer fibre installation orders, which required working with mainframe systems. AT&T says it has halved the time to provision dedicated internet access in buildings already wired with fibre, and that sales agents now check a prospect's location and fibre availability before a representative follows up.
Every production workflow begins with human oversight. Agents are allowed to act on their own only after accuracy reaches the mid-to-high 90% range within what AT&T considers an acceptable risk profile, and monitoring continues afterwards to catch drift.
What remains unclear
AT&T measures results through labour savings, revenue, operating costs and customer satisfaction, and SiliconANGLE reports that production AI investments return about five times their cost within a year. The company has not published the underlying methodology, the error rates of autonomous workflows or how many roles have changed as a result, so these returns cannot be checked from the outside.
