AT&T’s enterprise AI journey is worth studying not because it processes 45 billion tokens per day, nor because it is shifting a growing share of workloads to open models. Its real significance lies elsewhere: over roughly three years, AT&T has evolved generative AI from an employee productivity tool into a production capability embedded across software engineering, customer service, data analytics, network operations, and supply chain management, while progressively building a unified platform, domain-specific models, agent workflows, and model-routing infrastructure.
This trajectory illustrates a broader lesson for industrial AI transformation: the central question is not which large model an enterprise adopts, but whether it can turn AI into a production system that continuously improves business performance.
Ask AT&T Was the Starting Point, but the Goal Was Never to Build an Enterprise ChatGPT
When AT&T launched Ask AT&T in 2023, its immediate objective was to provide employees with a secure way to use generative AI. Yet the platform was designed from the outset to support multiple models. AT&T explicitly recognized that different models offer different combinations of capability and cost, so Ask AT&T was designed to work not only with OpenAI models but also with alternatives such as Llama and Falcon. (AT&T)
What matters more is how quickly the use cases expanded over the following two years. Ask AT&T moved into code generation, vulnerability remediation, customer-call summarization, product recommendations, natural-language database queries, HR support, and contract management. By 2025, Microsoft reported that AT&T had deployed 71 generative AI solutions serving more than 100,000 employees and processing approximately 9 billion tokens per day. Customer service representatives also reduced the time spent searching for information by roughly 33%.
The software engineering results are even more meaningful from an operating-performance perspective. According to AT&T, AI-assisted development shortened software development cycles by approximately 23% and reduced code-quality issues identified through static analysis by roughly 80%. The value of enterprise AI had therefore moved beyond helping individual developers write code faster; it was beginning to improve development lead time, software quality, and overall delivery capability.
AT&T’s first lesson is therefore clear: enterprise AI should begin with business processes and performance problems, not with a catalog of model capabilities. Customer service teams need shorter handling times, engineering organizations need faster delivery cycles, and network operations teams need lower incident-resolution costs. Models are computational resources used to achieve those outcomes.
From One Model for Every Task to Allocating Intelligence by Task
Once AI moved into production, AT&T encountered another practical reality: there was no reason for every task to use the most capable—and most expensive—model.
By 2026, AT&T’s AI workload had grown to an average of approximately 45 billion tokens per day. The company built a cache-aware AI Gateway that selects models for individual requests according to expected quality, speed, and cost, and can even switch models during a multi-turn conversation. AT&T has said that this mechanism reduced the cost of some workloads by as much as 90%, generating millions of dollars in cumulative savings.
The underlying logic is more sophisticated than simply assigning small models to low-value work and large models to high-value work. AT&T’s practice is closer to task engineering. Highly structured and easily verifiable activities—such as code summarization, classification, and information extraction—can often be handled by lower-cost models. Complex code generation and cross-system reasoning may still require frontier-class models, while highly specialized telecommunications workloads can be handled by domain-specific models.
The broader implication for industrial AI transformation is significant: enterprises ultimately need to manage different tiers of intelligent computing resources rather than manage a single “large model.”
An agent executing a single business task may perform retrieval, classification, planning, summarization, tool invocation, code generation, and verification. If every step invokes a frontier model, deeper automation can actually make the economics progressively worse. Architectures that scale therefore tend toward task decomposition, model portfolios, and dynamic routing.
OTel Shows Why the Long-Term Value of Industrial AI Comes From Domain Intelligence
AT&T’s work on Open Telco AI and OTel is particularly instructive. The company concluded that general-purpose frontier LLMs were not specifically trained for telecommunications standards, network equipment, or operational problems. Simply purchasing increasingly powerful general models therefore does not automatically provide deep telecommunications expertise.
OTel 2.0, introduced in 2026, underwent post-training on more than 400 billion telecommunications-domain tokens and was optimized for tasks including network operations, standards interpretation, retrieval-augmented generation, and agent tool use.
This reveals an important distinction between industrial AI and general-purpose AI.
Knowledge available on the public internet will become increasingly commoditized as foundation models improve. An enterprise’s own network telemetry, customer behavior, incident history, supply-chain rules, operational procedures, and accumulated experience, however, do not automatically become part of a general-purpose model.
Sustainable differentiation therefore comes from turning this proprietary industry knowledge into datasets, knowledge systems, evaluation standards, domain models, and reusable agent capabilities.
For AT&T, the value of open models is consequently not limited to lower inference costs. They also give the enterprise greater freedom to conduct continuous domain post-training, control deployment environments and model lifecycles, and encode proprietary telecommunications knowledge into reusable intelligent assets.
What Enterprises Should Replicate Is AT&T’s AI Production Method
The AT&T case does not prove that every enterprise should train its own large model, nor does it demonstrate that open models are always cheaper than proprietary models. Processing 45 billion tokens per day, training OTel on hundreds of billions of domain tokens, and operating large-scale GPU infrastructure are characteristics of an exceptionally large enterprise.
What is broadly transferable is the sequence in which AT&T built its capabilities:
Start with business-value use cases and use AI to improve real operating processes; consolidate successful use cases onto a shared enterprise platform; establish task-level evaluation and model tiers based on production workloads; introduce model routing and cost governance as usage scales; and eventually encode the enterprise’s most differentiated domain knowledge into specialized models and reusable agent capabilities.
This approach is much closer to the essence of industrial AI transformation than purchasing the strongest available model and encouraging everyone to use it.
From Ask AT&T to Agent Workflows, and then to AI Gateway and OTel, AT&T has followed a clear capability progression: AI evolved from an employee tool into a business capability, from isolated applications into an enterprise platform, and from general-purpose intelligence toward enterprise-owned domain intelligence.
For companies pursuing enterprise AI transformation, the key lesson is therefore not which models AT&T selected, but the operating loop it gradually established:
Business problem → AI use case → task decomposition → data and domain knowledge → models and agents → evaluation → business outcomes → continuous optimization.
Once this loop operates continuously, AI ceases to be a one-time technology purchase and starts becoming a production capability that the enterprise can accumulate, improve, and reuse over time.