AT&T

Bell Labs (now AT&T Labs) was the company’s research arm founded in 1925 to understand how communication systems work.

Over the next seven decades, its researchers invented foundational machine learning algorithms like Support Vector Machines, AdaBoost, and convolutional neural networks.

Fun Fact. Computer scientist Yann LeCun trained the first practical neural network at Bell Labs in the late 1980s to read handwritten zip codes for the U.S. Postal Service. He later shared the 2018 Turing Award for that work.

We shared some history here because all the stellar work and data collection has become AT&T’s moat for AI applications.

Every call record, signal drop, customer interaction, network fault, and robocall pattern has been accumulating for decades. AT&T processes more than 590 petabytes of data daily.

No competitor, new or old, can replicate that history.

So with all this data, where is AT&T investing in AI?

Most enterprises use AI to improve workflows around their core product. AT&T has made AI the operating system of its core product.

We can look at Geo Modeler, its generative AI system, as an example.

It runs a real-time digital twin of AT&T’s entire U.S. mobile network. Using ray tracing, the same technique that renders realistic lighting in video games, it simultaneously simulates how radio signals travel through buildings, terrain, and weather across the country.

When a cell tower goes down, the Geo Modeler doesn’t wait for a technician to diagnose the problem. It analyzes the situation and determines how neighboring towers should be reconfigured within seconds. AT&T says it runs dramatically faster than prior solutions.

We have studied a few Fortune 500 companies, UPS and BMW, that followed this approach as well.

Next, Gro Modeler is built on the Network Foundation Model (NFM). It’s a model trained on AT&T’s proprietary network data.

It powers two other critical applications:

Humans remain in the loop. But the system reads the problem, proposes the solution, and drafts the fix before a human engineer opens a ticket.

In June 2023, AT&T launched Ask AT&T, a generative AI platform built on Microsoft Azure.

By the end of 2025, it had expanded to 100,000+ employees and was generating an average of 27 billion tokens per day across its operations.

This volume is possible because of how AT&T architected the system. Instead of routing every query to expensive frontier models, AT&T fine-tunes small open-source models (4 to 7 billion parameters) on proprietary data: contracts, network logs, customer call transcripts, compliance documents, etc.

Its orchestration layer routes each task to the right models, resulting in model costs cut by up to 90%.

Employees get all the data and answers they need through simple, natural-language interactions.

We gathered what use cases different roles unlock with Ask AT&T:

…and what tools AT&T professionals use internally:

AT&T handles tens of millions of customer calls a year, manages 50,000 fleet vehicles, and blocks more than 2.5 billion robocalls monthly.

Let’s see how AT&T uses AI in core operational functions.

AT&T runs dozens of fraud-detection models that combine traditional ML and generative AI.

AT&T’s 50,000 vehicles cover 700 million potential routes daily.

The company’s dispatch AI matches technicians to jobs based on skills and proximity. These small language models match the performance of larger models at a fraction of the cost, contributing to multi‑million‑dollar annual savings.

Combined with network-integrated AI, these improvements cut customer service resolution time by 33%.

AT&T partnered with H2O to fine-tune small language models that classify customer-agent interactions into 80+ categories. The models match the performance of expensive large models at 35% of the cost, saving approximately $2 million annually.

Some more results from AT&T’s work with specialist partners:

AT&T’s internal AI capabilities are now becoming external products. The company is positioning itself as the connectivity and AI infrastructure layer for enterprise customers, not just a telco that uses AI well.

Some signals:

Open Telco AI. AT&T released pen-source models trained specifically on telecom data. These models claim to produce more accurate results for telecom-specific tasks. It positions the enterprise as a contributor to the AI ecosystem.

Connected AI for Manufacturing. In collaboration with NVIDIA, Cisco, and Microsoft, AT&T launched a platform combining 5G, IoT, and generative AI for smart factory environments. Early pilot results include a 70% reduction in waste on injection molding lines, 2.5 to 4 hours of pre-failure fault detection, and a 35% improvement in fulfillment center efficiency.

Physical AI for supply chains. Through a partnership with Wiliot, AT&T is deploying battery-free IoT sensors that connect physical goods, retail inventory, food shipments, logistics assets, etc., directly to its network. AT&T handles field execution, device certification, and ongoing operations.

Audit what you already have before buying what’s new. AT&T’s AI works because Bell Labs spent 70 years collecting data before GenAI existed. Most enterprises underestimate the value of their accumulated data. It is a goldmine.

One governed platform scales faster than many ungoverned tools. Ask AT&T is a single gateway for 100,000+ employees. Every new use case, whether in legal, procurement, or engineering, deploys faster because the infrastructure, security, and compliance are already in place.

Match the model size to the task. AT&T fine-tunes small open-source models on internal data rather than defaulting to the largest frontier models available. It results in 90% lower model costs and higher accuracy on domain-specific tasks.