The Enterprise AI Ownership Threshold

For large enterprises, the question is no longer whether artificial intelligence will become part of everyday work. The question is how much of that AI usage should depend on frontier-model providers, and how much should be owned and operated internally.

Mark Austin, who leads AI for AT&T’s 100,000 employees, described an organization processing approximately 45 billion tokens every day. Open-source models currently handle around 40% of those prompts, with an ambition to increase that share to between 60% and 70%.

At frontier-model list prices, that level of usage could represent hundreds of millions of dollars annually. Instead, AT&T is running a growing share of its workloads on its own Nvidia and AMD servers.

The strategy reflects a broader shift in enterprise AI: use frontier models where their capabilities are essential, but stop paying frontier prices for work that no longer requires them.

Not every AI task requires the most powerful model

The assumption that every AI task should run on the most capable available model is becoming increasingly difficult to justify.

Many enterprise workflows involve tasks such as summarizing a code change, retrieving an internal HR policy, or cleaning up meeting notes. These tasks can be useful and high-volume, but they are not necessarily complex. They often require consistency, speed, and access to the right context more than they require the absolute leading edge of model intelligence.

That distinction matters when AI usage reaches enterprise scale.

A small difference in cost per request may seem unimportant when evaluating an individual interaction. Across billions of tokens, however, the economics change dramatically. If a capable open-source model can handle routine workloads adequately, using it for those tasks can reduce costs while preserving frontier models for situations where their additional capabilities matter.

This is not an argument that open-source models have surpassed frontier labs. Austin acknowledges that open weights may still trail frontier models by six to ten months. The point is that many enterprise tasks do not need to operate at the frontier in the first place.

The three-part enterprise AI strategy

The approach described by AT&T can be summarized in three principles.

1) Own the harness and swap the models

The “harness” surrounding a model (the application logic, routing, evaluation, monitoring, and workflow infrastructure) can be more durable than any individual model choice.

When an organization owns that layer, it can change models as requirements evolve. A task can be routed to an open-source model, a specialized model, or a frontier model depending on its complexity, sensitivity, and cost.

This creates flexibility: the enterprise is not locked into one provider or one model family. Instead, models become interchangeable components inside a broader system.

2) Run the boring volume yourself

Routine AI workloads are often the most numerous. They may not be the most visible or technically impressive, but they represent a large share of total usage.

Running this volume internally can make economic sense when an organization has the infrastructure, engineering capacity, and operational maturity to support it. The goal is not to build everything from scratch; it is to avoid using the most expensive option by default for tasks that can be handled elsewhere.

The “boring” work is precisely where scale matters most. A single summary is inexpensive. Millions or billions of summaries become a strategic cost consideration.

3) Use frontier models where necessary

Frontier models still have an important role. Some problems require more advanced reasoning, broader capabilities, or stronger performance on difficult and ambiguous tasks.

The alternative to universal frontier-model usage is not abandoning frontier models. It is reserving them for work that benefits from them most.

This creates a tiered architecture: routine tasks are handled by lower-cost models, while complex work is escalated to the frontier when necessary.

Enterprise AI is becoming a routing problem

The deeper change is that AI deployment is becoming less about selecting a single “best” model and more about designing an effective routing system.

A mature enterprise setup may need to answer several questions for every request:

• How difficult is this task?
• What level of accuracy is required?
• Does the request contain sensitive information?
• Is latency important?
• Which model can handle the task at an acceptable cost?
• When should the request be escalated to a more capable model?

This approach treats model selection as an operational decision rather than a permanent commitment.

It also changes how enterprises think about competitive advantage. The value may not come only from access to a particular model. It may come from the systems built around those models: the data, workflows, evaluations, routing logic, and infrastructure that determine how AI is used across the organization.

The cost of staying on the frontier by default

Organizations that send every request to a frontier model may be paying for capabilities their workloads do not need.

That cost is not limited to the model invoice. It can also influence infrastructure planning, vendor dependence, data-handling decisions, and the ability to adapt when models or pricing change.

Enterprise AI adopters therefore face a choice. They can continue treating every prompt as a frontier-model problem, or they can separate routine volume from genuinely complex work.

The organizations that make this distinction may gain more control over both their costs and their technology strategy. Frontier models will remain essential for difficult tasks, but not every task needs to be difficult to be valuable.

The emerging lesson is simple: own the system, run the routine workload efficiently, and use the frontier when it truly adds value.