AT&T Said It Generated 5x Return on AI Investment
Amdocs said it is launching a new agentic operating system for managing AI agents across carrier networks
BARCELONA, March 6, 2026 — AT&T, the Dallas-based telecommunications carrier, said Tuesday it processes 27 billion tokens daily across its enterprise operations and generates a five-fold return in free cash flow within the same year of investment on every dollar spent on artificial intelligence.
Andy Markus, AT&T's chief data and artificial intelligence officer, said MIT research found 95 percent of AI proof-of-concept projects return zero dollars in investment. AT&T skips extended pilot phases and deploys AI directly into production operations across its network, back office and customer service divisions, he said.
"AT&T is not a pilot company," Markus said, speaking at the Mobile World Congress (MWC) here.
Markus said AT&T replaces large language models with fine-tuned small language models, compact AI systems trained on domain-specific data, for every production use case across the company. Fine-tuned small language models match the accuracy of larger models at 90 percent lower cost, he said.
AT&T collaborated with the GSMA, the Geneva-based association representing mobile network operators globally (and the host of MWC), to build an open-source telecom-specific language model post-trained across 30 architectures, parameter sizes and model families, Markus said. The model outperforms frontier models on telecom tasks including root cause analysis and runs on processors built by Nvidia and AMD, the two dominant U.S. semiconductor companies supplying AI computing hardware, with the model, code and dataset all publicly available, he said.
Markus said accuracy is the central challenge in scaling agentic AI, software systems that take autonomous actions across multiple steps without human intervention. In a six-step agentic workflow where each step operates at 90 percent accuracy, the reliability of the end result is barely better than a coin flip, he said. AT&T's internal data platform, enriched with business and technical metadata, pushes accuracy to 100 percent, he said.
"Clean data, with the right technical and business metadata, is the fuel for driving successful AI," Markus said.
Amdocs, the Israel-based telecommunications software company, announced what it called an “Agentic Operating System” designed to coordinate AI agents across carrier networks, billing systems and customer operations. Gil Rosen, Amdocs' chief marketing officer, said the product gives operators the ability to audit and verify AI decisions across their entire organization rather than managing agents in isolated departments.
"We're not launching a product," Rosen said. "We're launching a different operating model."
Markus said AT&T currently manages hundreds to thousands of AI agents across its operations and expects that number to reach hundreds of thousands within a short timeframe. Agents operating inside a single company are manageable, he said, but agents sent in from outside vendors and partners present a harder governance challenge because their permissions and actions are harder to track and control.
An industry study cited during the session found 20 percent of telecom operators are fully committed to agentic AI deployment, 50 percent are in progress and 30 percent remain undecided, Rosen said.
Member discussion