AT&T starts every production AI agent workflow under human oversight and removes the human only once accuracy reaches the mid-to-high 90s with an acceptable risk profile. That rule covers nearly 500 agentic workflows in production, which AT&T plans to triple next year, according to SiliconANGLE's account of a talk by Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, at an AWS event in Boston this month.
The number is the easy part to copy. The harder parts are the review lane that every customer-facing workflow has to pass, an architecture with few agents and centrally stored business logic, and a cost model that sends a large share of traffic to cheaper models. Those decide whether your version of the rule survives a year of scale.
What AT&T Actually Runs in Production
AT&T's agent platform handles more than 1 million transactions a month across nearly 500 workflows, and the company processes about 45 billion AI tokens a day, per Corbin. The token figure has grown fast. At Mobile World Congress in March, AT&T said it processed 27 billion tokens daily; by July, chief data and AI officer Andy Markus wrote that the company averaged 45 billion tokens per day.
The workflows Corbin described are unglamorous back-office work:
- Fiber installation order changes used to need someone to pull tickets by hand and update records. Agents now talk to AT&T's mainframes directly. "It was kind of cool having some of the most advanced technology talking to some of the oldest that we have," Corbin said.
- Dedicated internet provisioning time was cut in half in buildings that already have fiber.
- In sales prospecting, agents check a customer's location and fiber availability before a rep pursues the deal.
AT&T says its production AI investments return five times their cost within a year. That is the company's own figure, stated by Markus at MWC as a free cash flow return of five times spending. In January the claim was a "two-times return on investment" expected to grow, so the ratio moved fast and has not been independently audited. Treat it as a vendor-grade claim from a company with a reason to publicize it.
How AT&T Decides an Agent Can Run Without a Human
AT&T removes human oversight from a workflow only when two conditions hold: accuracy in the mid-to-high 90 percent range, and a risk profile the company judges acceptable. After that, continuous monitoring watches performance and checks for drift, according to SiliconANGLE.
An autonomy threshold is the measured accuracy a workflow must sustain before it may act without a person approving its output. AT&T's version is useful because it is explicit and applies to every workflow by default. Most enterprises have the opposite habit. They approve a pilot with a human in the loop and then let the human drift out of the loop as volume grows, without anyone deciding it.
The threshold is also incomplete in ways you should fix before borrowing it.
Accuracy per what? "Mid-to-high 90s" says nothing about what counts as one unit. Scored per step, a five-step workflow with each step right 97% of the time finishes cleanly only about 86% of the time (0.97 to the fifth power; our arithmetic, not AT&T's). Scored per completed transaction, the same label means something much stricter.
What does the miss rate cost? At the platform's volume, a 3% error rate on 1 million monthly transactions would be about 30,000 wrong outcomes a month. That is our arithmetic on Corbin's platform total, which mixes supervised and autonomous workflows, so read it as an illustration of scale. Either way, the same percentage is trivial on a prospecting lookup and expensive on a provisioning order. A single threshold across 500 workflows only works if "acceptable risk profile" carries the weight, and AT&T has not published how it scores that.
Who watches the human? An approval step that a reviewer clicks through hundreds of times a day stops being oversight. We covered how often that happens in our buyer's guide to human-in-the-loop approval gates. AT&T's earlier public descriptions say a human oversees the chain of agents with checkpoints throughout, every agent action is logged, and handoffs between agents enforce data isolation, retention policies and role-based access. The logging is what makes the accuracy number measurable at all.
Two Lanes: Personal Agents and the Business Agent Foundry
AT&T runs two governance lanes. Employees can build personal productivity agents on their own, while any workflow that touches customers has to go through the Business Agent Foundry and pass security, privacy and compliance reviews, Corbin said. "We've democratized agents to the point where people can use them for personal productivity," he said.
The Foundry was set up in early 2025 to work with individual business units. It maps their processes against industry frameworks to find steps, handoffs between teams and legacy application dependencies that can be removed before an agent is built. The order matters because automating a process that has five handoffs produces an agent with five handoffs.
A second layer sits above it. All agentic and generative AI use cases pass through AT&T's Generative AI Transformation Office, which ranks them by value, and each needs a formal business case tied to the CFO's office, with most targeting a free cash flow benefit. The tooling for the self-serve lane is Ask AT&T Workflows, a graphical drag-and-drop agent builder that AT&T announced in November 2025.
The organizational intent is to push ownership out. "I feel privileged to have the team that I do, but in the long run, my team shouldn't exist," Corbin said. A central team that plans its own retirement is unusual, and it is the right instinct for a function whose job is to teach business units to run agents themselves. It also needs an owner for the threshold. If the central team disappears, someone in each business unit has to own the accuracy measurement and the decision to remove the human.
SK Telecom took a different route to the same problem, giving agents employee IDs and assigning supervisors by autonomy level, as we covered in its AI governance report.
Why AT&T Uses Few Agents and Central Business Logic
AT&T runs a relatively small set of agents on top of centrally managed business logic, instead of building a separate specialist agent for each process. "We store the business logic in context or in memory," Corbin said. Retrieval agents find the relevant knowledge, and planning agents decide which actions, tools and data a task needs.
Markus described the same pattern in January: a master or control agent directs smaller agents, which often run on small language models fine-tuned for narrow tasks. "Accuracy is the driver of everything," he said.
The case for this design is maintenance. When a pricing rule or a provisioning policy changes, you update it once, and every workflow that reads it picks it up. Five hundred specialist agents with logic baked into their prompts would mean five hundred places to edit and five hundred places to drift. Independent research points the same way: an Amazon study found one agent is the efficient default for most standard operating procedures, with multi-agent setups using far more tokens. Meta went the other way on one narrow job, using a thin orchestrator over 14 subagents, so the right count still depends on the job.
The cost of central logic is blast radius. One bad edit to a shared rule reaches every workflow at once. That is a strong argument for versioning business logic like code, with review and rollback, and for re-running accuracy checks on affected workflows after each change.
How AT&T Keeps 45 Billion Tokens a Day Affordable
AT&T keeps costs down by routing each request to the cheapest model that can do the job. Markus wrote in July that a proprietary cache-aware AI Gateway weighs speed and cost against expected output quality at each turn, can switch models mid-session, and has cut AI costs by as much as 90%. "Efficiency only matters if accuracy keeps pace," he wrote. AT&T also trains its own telco models; the open telco model OTel 2.0 was post-trained on more than 400 billion tokens on AMD GPUs.
The open-model share is large. Mark Austin, an AT&T vice president overseeing AI, told The New York Times that about 40% of employee AI requests now go to open-source models instead of Claude or GPT, with a target of 60% to 70% within a few years, as summarized by AI News Weekly. On coding workloads, the shift cut costs 56% for a 2% drop in quality, and LiteLLM handles routing of low-complexity work to models such as Nemotron, Llama and Gemma, per the same report.
A 2% quality drop is a trade a cost-led team takes without much thought. It interacts with the autonomy threshold, though. A workflow that cleared the high-90s bar on a frontier model has not cleared it on a cheaper one. If your router can move a workflow to a different model, re-measure accuracy per model, or pin autonomous workflows to the model they were validated on. We covered why routers are better at failover than judgment in our model router buyer's guide, and how Harvey swapped models after agent costs sank its margin.
Where This Model Can Break
The strongest case against copying AT&T is that its numbers are self-reported and its definitions are private. Nobody outside the company knows how accuracy is scored, how the risk profile is weighed, or what "nearly 500 workflows" counts as a workflow. The 5x return comes from AT&T's own executives.
The industry record also argues for caution. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls. Klarna said in early 2024 that its AI assistant did the work of 700 support staff, then reversed in 2025 when CEO Sebastian Siemiatkowski said the AI output was cheaper but "lower quality" and began hiring people again. AT&T's design puts an accuracy number in front of autonomy, which addresses the failure Klarna described, as long as the routing layer is not lowering accuracy underneath it.
AT&T's own legacy also cuts both ways. Corbin described a 151-year-old company with fragmented data after many mergers, and said AT&T had to consolidate data and improve connectivity before agents could work. Most enterprises have that problem. Fewer have spent the money AT&T has to fix it.
Banking shows how often the monitoring half gets skipped: BCG found banks permit AI agents but few have standard agent monitoring.
What to Do Before Your Next Agent Goes Autonomous
This Week:
- List every agent workflow in production and mark which ones run without a human approving output. If nobody decided that, write down who did it by default.
- For each autonomous workflow, write the accuracy number it cleared, how it was measured (per step or per transaction), and the model it was measured on. Blanks are your backlog.
This Month:
- Split governance into two lanes like AT&T's: self-serve for personal productivity agents, and a review lane covering security, privacy and compliance for anything that touches customers or systems of record.
- Set an error budget per workflow in errors per month, using its real volume, instead of one percentage for everything. A 3% miss rate means different things on a sales lookup and on a provisioning order.
- Move shared business rules out of individual prompts into one versioned store, and require a re-run of the accuracy check on every workflow that reads a rule after it changes.
Before Your Next Budget Cycle:
- If you route across models, require each autonomous workflow to be re-validated on any model the router can send it to, or pin it. Ask your platform team how often the router changed models on those workflows last month.
- Name an owner for the autonomy decision inside each business unit, so the threshold survives when the central AI team shrinks.
The Bottom Line
AT&T's contribution is a written promotion rule that applies to every workflow, backed by logging, drift monitoring and a separate review lane for customer-facing agents. Klarna's reversal is the earlier lesson: its chatbot was cheaper than staff, its CEO said the output was lower quality, and the company went back to hiring people. AT&T's threshold is about accuracy and its gateway is built to cut cost, and the two meet on any workflow the router moves to a cheaper model.
Before you copy the number, decide what one unit of accuracy is and what a month of misses costs on each workflow.
Continue Reading
- Human-in-the-Loop for AI Agents: Most Approval Gates Rubber-Stamp
- SK Telecom's AI Governance Report Gives Agents Employee IDs
- Amazon Study Finds One Agent Is the Efficient Default
- Anthropic Monitors 30,000 Agents. Can You Count Yours?
- Model Router Buyer's Guide: Buy Failover, Not Judgment
- Banks Permit AI Agents, but Few Have Standard Monitoring, BCG Finds