AT&T Scales Agentic AI to 45 Billion Daily Tokens and 5x Returns
AT&T has deployed nearly 500 production agentic workflows, processing 45 billion daily tokens. Here is how the telecom giant structures its architecture and governance for high-ROI AI.
AT&T Corp. is expanding its artificial intelligence strategy beyond basic employee productivity tools, deploying autonomous agents to redesign core business operations from the ground up. Speaking at an Amazon Web Services Inc. event in Boston, Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, detailed how the 151-year-old telecommunications giant now operates nearly 500 agentic workflows in production.
AT&T currently processes approximately 45 billion AI tokens daily across the enterprise, while its dedicated agentic AI platform handles more than 1 million transactions per month. The company plans to triple its number of production agentic workflows over the next year, moving from task-level automation to autonomous business process execution.
Re-architecting enterprise workflows around centralized logic
AT&T’s architectural strategy departs from the common industry trend of building large fleets of independent, specialized agents. Instead, the organization maintains a relatively small set of core agents that execute workflows based on centrally managed logic. According to Corbin, AT&T stores business logic directly in context or memory. In this framework, retrieval agents locate relevant knowledge while planning agents determine the specific tools, actions, and data required to execute a process.
To facilitate this shift, AT&T established the Business Agent Foundry in early 2025 to collaborate directly with individual business units. Rather than gathering isolated automation ideas, the foundry maps end-to-end business processes using industry frameworks to eliminate unnecessary steps, organizational handoffs, and legacy system dependencies.
This structural overhaul has already yielded concrete operational changes. In one deployment, AT&T integrated agentic automation with legacy mainframe systems to handle customer fiber installation orders, replacing manual queue management and accelerating service delivery. In another instance, the company cut the time required to provision dedicated internet connections in fiber-ready buildings by 50%. The platform is also used in sales prospecting, where agents automatically validate customer locations and verify fiber availability before sales teams spend time pursuing opportunities.
Governance, risk boundaries, and financial returns
Governance plays a central role in AT&T's agent platform. While individual employees are free to create personal productivity agents, any workflow that impacts customers or core business operations must pass through the Business Agent Foundry for compliance, privacy, and security evaluations.
Every production agentic workflow operates under human supervision during its initial phase. Autonomous execution is granted only after an agent achieves accuracy in the mid- to high 90% range within an acceptable risk profile. Once deployed, continuous monitoring protocols evaluate agent performance and guard against drift.
Performance is tracked through labor savings, operational costs, customer satisfaction scores, and direct revenue generation. Corbin noted that AT&T’s production AI investments are generating a fivefold return on investment within a single year, largely by offsetting the high ongoing maintenance and training costs associated with legacy applications.
What it means for developers
AT&T's implementation underscores a growing enterprise trend: moving away from fragmented, standalone agents toward unified architectural models built on centralized context, retrieval, and planning layers. For engineering teams, this highlights the necessity of designing robust context-management systems and rigorous governance pipelines before deploying autonomous workflows into production.
As enterprise token consumption reaches massive scales—such as AT&T's 45 billion tokens per day—managing multi-model API access and infrastructure costs becomes a primary consideration. Developers looking to build and test multi-agent systems across leading foundational models can try top AI models cheaply through one API at Apixo, which offers pay-per-token access to Claude, GPT, Gemini, Grok, and DeepSeek via a single API key.
Looking ahead, AT&T is shifting its focus from optimizing legacy infrastructure toward using agentic workflows to enter new markets, launch services, and expand sales channels. Corbin noted that the ultimate goal is to distribute AI transformation responsibility across all business units, eventually rendering central management structures like the Business Agent Foundry unnecessary.
Source: AT&T uses agentic AI to reinvent business processes from the ground up — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
