AT&T runs 500 agentic AI workflows. Copy the rules.
AT&T runs nearly 500 agentic AI workflows and about 45 billion tokens a day. The useful part is the rule set: few agents, central logic, human oversight first.
AT&T says it has nearly 500 agentic AI workflows in production and processes about 45 billion tokens a day. You will not operate at that scale. But the rules that AT&T uses to put an agent into production fit a 10-person company as well as a 151-year-old carrier. They cost nothing to copy.
What actually happened
SiliconANGLE published a case study on November 14. The source is Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, who spoke at an Amazon Web Services event in Boston. These are AT&T's own numbers, with no outside audit:
- Volume. Nearly 500 agentic workflows, about 45 billion tokens a day across the company, and more than 1 million transactions a month on the agent platform. AT&T plans to triple the workflow count next year.
- Results. The time to provision a dedicated internet connection, in buildings that already have fiber, went down by half. Corbin says production AI investments return five times their cost in one year.
- Legacy work. Agents now make changes to fiber installation orders on AT&T's mainframes. Before, employees pulled tickets from a queue and updated the records by hand.
- Sales. Agents validate a customer location and check fiber availability before a sales rep spends time on the lead.
The article names no model and no agent framework.
How AT&T controls its agentic AI workflows
This is the part to copy.
- Few agents, central logic. AT&T does not build one special agent for each task. A small set of agents runs the workflows, and the business rules are in one managed place. Retrieval agents find the applicable knowledge. Planning agents select the actions, tools and data.
- One gate for customer work. Employees can make personal productivity agents freely. A workflow that touches customers must go through a central team, the Business Agent Foundry, with security, privacy and compliance reviews.
- Human oversight first. Each production workflow starts with a person in the loop. The agent runs alone only after its accuracy is in the mid-to-high 90% range, inside an agreed risk level.
- Monitoring after launch. AT&T continues to measure performance and looks for drift.
The Foundry also maps each process and removes steps, handoffs and legacy dependencies before it automates anything.
Why it matters for your business
Most small businesses do the opposite. They buy five tools, each with its own agent and its own copy of the rules. Then the refund policy in the support bot does not agree with the refund policy in the email bot.
What to take from AT&T:
- Write the rules once. Pricing, refunds, service area, approval limits. One source that each agent reads.
- Fix the process before you automate it. A bad process with an agent is a faster bad process.
- Start with approval on. Measure how often the human changes the agent's output. Remove the approval step when that number is small and stable.
- Give old systems to agents last, but do give them. If AT&T can put an agent in front of a mainframe, your 2014 inventory system is not a blocker.
- Measure money. AT&T measures labor savings, revenue, operating cost and customer satisfaction.
Key takeaways
- AT&T reports nearly 500 agentic workflows, about 45 billion tokens a day and more than 1 million transactions a month
- Provisioning time for dedicated internet in fiber-ready buildings went down by half
- A small set of agents runs on centrally managed business logic
- Customer-facing workflows go through one review gate; personal agents do not
- Each workflow starts with human oversight and runs alone only at mid-to-high 90% accuracy
You do not need 500 workflows. You need one that pays. We map the process, put the business rules in one place and launch the agent with approval on until the numbers say it can run alone. Estimate what one automated workflow returns, or tell us the process that costs you the most hours.
Sources: SiliconANGLE.
- #agentic-ai
- #ai-agents
- #att
- #workflow-automation
- #human-in-the-loop
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Qwen4-Max released: $1.50 input, cost per task down 62%
Alibaba released Qwen4-Max with a 1M-token context and API access at $1.50 input and $4.50 output. It is not open weights. Measure cost per task before you switch.
Read itAI accounting software: Blue J $176M, Sumary $25M
AI accounting software raised two large rounds: Blue J $176M for tax research and Sumary $25M for firm workflows. Ask your accountant which AI reads your books.
Read it