The Frontier Gap Is Widening

In August 2026, OpenAI reported a striking shift in enterprise AI adoption: frontier firms, defined as the top 10% of AI usage each month, now generate 8.3× as many output tokens per active user as typical firms. In January, that number was just 2.6×. That is not a small increase. It is a widening gap.

It also helps explain a paradox many companies are feeling right now. On one hand, GPU prices continue to rise and infrastructure demand keeps climbing. On the other hand, many business leaders still feel underwhelmed by their own internal AI results. If AI is supposedly transforming work, why are so many organizations seeing only marginal gains?

The answer is that most enterprises are still using AI at the surface level, while the most advanced firms have moved deeper into agentic workflows.

Chatbots are opt-in; agents are opt-out

Most companies are still using AI like a chatbot

For many organizations, enterprise AI adoption still looks like this: employees occasionally open Microsoft Copilot or Enterprise ChatGPT, type in a prompt, get a response, and move on. The interaction is helpful, but limited. It depends on the employee remembering the tool exists, knowing when to use it, and taking the extra step to engage with it.

That is opt-in AI.

Opt-in systems can improve productivity around the edges, but they rarely transform the way work gets done. They live outside the company’s actual operating processes. They are separate from the workflows, databases, approvals, systems, and routines that define how the business runs day to day.

As a result, usage remains shallow. Output stays modest. And leaders conclude that AI is interesting, but not yet mission-critical.

Frontier firms are using agents that execute

The firms pulling away are doing something fundamentally different. They are not just giving employees better chatbot interfaces. They are deploying agents that can actually execute work inside the business. These agents connect directly to the databases and information systems that matter. They are designed around the company’s specific internal processes. And instead of simply suggesting what a person should do next, they do the heavy lifting themselves.

That is a major leap. (edited)

An agent that can read context, work across systems, take action, and escalate only when human judgment is required will naturally consume far more tokens than a single chatbot exchange. In fact, these workflows can consume 3 to 4 orders of magnitude more tokens per task than traditional chatbot usage.

This is why that tiny line next to “Chatbot Query” in token consumption charts is not a rendering glitch. It is drawn to scale.

The difference is real. And it reflects a completely different category of value creation.

The real breakthrough is opt-out, not opt-in

The most important distinction is not better prompting. It is not that frontier firms have somehow trained every employee to become a world-class prompt engineer. It is not that they have run more workshops or distributed better internal playbooks. The real breakthrough is that their agents operate on an opt-out basis rather than opt-in.

That changes everything.

In an opt-in model, AI only creates value when someone consciously decides to use it. In an opt-out model, AI is embedded into the workflow by default. Employees do not need to remember to activate it. They do not need specialized training to get started. They do not need to stop what they are doing and ask, “Could AI help with this?”

The work just gets done.

An agent handles the process in the background, and when human input is actually needed, the relevant person gets pinged. Instead of asking people to adapt themselves around the tool, the tool adapts itself around the organization.

That is what makes agentic systems powerful. They do not rely on ideal user behavior. They are built for real operating environments.

Why companies are willing to pay for higher AI costs

From the outside, rising AI infrastructure costs can seem irrational. Why would firms willingly absorb much larger token bills?

Because when agents execute effectively, the economics look very different.

A chatbot that occasionally helps someone draft an email or summarize notes is useful, but easy to deprioritize when budgets tighten. An agent that actually moves work forward across real business systems is much harder to dismiss. It saves time in a more direct way. It reduces friction in core processes. It increases throughput without requiring every employee to become an AI specialist.

In other words, the spend rises because the value rises.

Frontier firms are not paying more for novelty. They are paying more because they have found a model of AI usage that actually works at operational scale.

That is the divide the market is starting to reveal.

AI workflows in 2026 are no longer experimental

The state of AI workflows in 2026 looks very different from where things stood even a year ago.

The conversation is no longer just about whether employees should have access to a chatbot. It is about whether businesses are redesigning their processes around systems that can reason, integrate, and act.

The companies making that shift are building a very different relationship with AI. They are not treating it as a side utility. They are treating it as part of the execution layer of the organization.

And that is why the frontier gap is widening.

The firms that remain stuck at the chatbot stage may still describe themselves as “using AI,” but they are not participating in the same game as the firms building agentic workflows into the fabric of their operations.

That difference will only become more visible over time.

Conclusion

If your own experience with AI at work has felt underwhelming, it may not be because the technology is overhyped. It may be because your organization is still approaching AI as an optional productivity assistant rather than an embedded execution system.

The companies pulling ahead are not simply prompting better. They are building agents that operate across their real workflows, integrate with the systems they already depend on, and run by default instead of by exception.

That is what frontier adoption looks like now.

And if businesses want to keep pace, they cannot afford to stay at the chatbot layer much longer.

‍