NewsIndustry AnalysisAI & DevelopmentTech Business

AT&T: 500 AI Agent Workflows, 45 Billion Tokens Per Day

AT&T is running nearly 500 agentic AI workflows in production, processing 45 billion tokens per day, and handling over a million transactions per month. The company says its production AI investments are returning fivefold within a year. These are not pilot numbers. This is what enterprise AI agents at scale actually look like — and AT&T’s architectural choices carry lessons that most teams building agents today tend to skip.

What the Business Agent Foundry Is

AT&T established the Business Agent Foundry in early 2025 to work with individual business units on process redesign — not process automation. The distinction matters. The Foundry maps existing workflows using industry frameworks, identifies unnecessary steps and organizational handoffs, and removes legacy application dependencies before any agent is introduced. The agents come last, not first.

The platform handles over 1 million transactions per month across customer provisioning, network operations, sales prospecting, and knowledge retrieval. An AI gateway routes tasks to the right model — cheaper models for lower-complexity work, more capable ones for the rest. The platform is model-agnostic by design, so AT&T is not locked into any single vendor as the market shifts.

The Architectural Choice That Makes This Maintainable

AT&T chose a centralized architecture over the alternative most teams reach for — a large fleet of specialized agents, each carrying its own business logic in its prompts. The reason is operational: 500 specialist agents with logic baked into their prompts means 500 places to update and 500 places to drift. The centralized model keeps business logic in one place, maintained consistently, audited against a single source of truth.

In practice, the stack uses a small set of agent types: retrieval agents that locate applicable knowledge, planning agents that determine which actions and tools are needed, and a master agent that coordinates them. Smaller agents in the stack run on small language models fine-tuned for narrow tasks — which keeps inference costs manageable at this scale.

This is not how most teams build. The default pattern is to create a new agent for every new capability, let prompts grow in complexity, and then discover six months later that no one knows which agent owns which logic anymore. AT&T’s choice to centralize is a deliberate constraint that becomes a competitive advantage at 500 workflows.

The Governance Rule That Gets Agents to Production

Every production workflow at AT&T starts under human oversight. Human oversight is removed only when two conditions are met: accuracy in the mid-to-high 90 percent range, and a risk profile the company judges acceptable. This applies to all 500 workflows. Customer-facing workflows require a separate security, privacy, and compliance review before going live.

Industry data puts this in perspective. Forrester 2026 data shows that 88 percent of agent pilots never reach production. The gap between an impressive demo and a production-ready system is exactly where governance lives — security reviews, compliance requirements, identity management, and exception-heavy workflows that break agents that only saw clean test data. AT&T has cleared that gap 500 times.

Every action is logged. When one agent hands work to the next, data isolation, retention policies, and role-based access are enforced at the transition. Human checkpoints remain embedded in complex workflows even after the accuracy threshold is met for lower-risk steps.

The Results

The production numbers are specific enough to be credible. Network fault resolution time dropped from 47 minutes to 28 minutes — a 40 percent improvement — as AI agents eliminated false-positive queue time and pre-assembled context for engineers before they looked at an alert. iPhone fraud, an approximately $1 billion annual problem in the US, has been reduced by more than 80 percent. Call center analytics work has decreased 84 percent. And the company’s AI investments are returning fivefold within a year, according to Luke Corbin, AT&T’s AVP of Digital Transformation and AI.

AT&T plans to triple its workflow count — from roughly 500 today to approximately 1,500 by 2027. At 45 billion tokens per day processed now, the infrastructure is already under real load.

What to Take From This

The enterprises actually running AI agents in production at scale share a few patterns: they redesigned processes before they automated them, they built governance in from the start rather than bolting it on after a production incident, and they made deliberate architectural choices about where logic lives and how it is maintained. AT&T’s results across fraud, network ops, and customer service make a simple argument: get the architecture and governance right, and the results follow.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News