A new gap is forming inside organizations, and it has nothing to do with talent.
Artificial intelligence can already carry out far more tasks than companies are actually putting it to work on.
New labor market research from Anthropic reveals a sharp disconnect between what large language models could theoretically accomplish and how they are deployed inside organizations. The findings expose a clear AI adoption gap between what the technology is capable of and how companies choose to use it.
In other words, the speed of AI transformation will depend on how fast organizations rethink the way work gets done.
Why AI Adoption At Work Is Falling Behind
The Anthropic research introduces a concept called “observed exposure,” a metric that combines two dimensions: the tasks AI could theoretically handle and the tasks professionals are actually completing with AI assistance.
When those two dimensions are compared, the gap is significant.
In computer and mathematics occupations, large language models could theoretically support the vast majority of tasks. Yet actual usage today covers roughly a third of them.
The same pattern shows up across many professions. AI technology has moved quickly, but AI adoption inside organizations has not kept pace. The constraint is structural, not technological. Work inside most companies is still built around static roles, fixed responsibilities and tightly scoped processes. Systems that can generate analysis, draft solutions and automate entire task chains cut across those boundaries, making them hard to absorb without rethinking how work is structured.
AI Is Augmenting Work Rather Than Transforming It
Another finding in the research helps explain why the gap persists.
Most AI usage today is augmentative rather than fully automated. In tracking real-world adoption, the researchers draw a distinction between systems that completely execute a task and those that simply help people do it faster.
That distinction reveals how organizations are actually deploying the technology. AI is helping people draft reports, analyze data, summarize documents or generate ideas. But the surrounding process — the approvals, the handoffs, the accountability structures — often stays exactly the same.
As a result, individuals are more productive doing what they have always done, but the work still moves through the same systems, the same checkpoints and the same decision layers that existed before AI arrived. Until those workflows change, organizations will not capture the full value of what is now possible.
Which raises the obvious question: if AI is already capable of performing many workplace tasks, why are companies not using it more broadly?
The research points to several practical reasons. Many tasks still require human verification. In other cases, AI tools are difficult to integrate with existing systems. And in large organizations, most tasks sit inside complex workflows with approvals, policies and dependencies that slow everything down.
These are not technical limitations. They reflect how work is organized inside companies. Over time, organizations accumulate layers of approvals, systems, policies and coordination mechanisms designed to manage risk and align teams.
AI does not simply slot into these structures.
Before AI can automate or augment a task, the surrounding workflow often has to change. Responsibilities need to be reassigned. Decision rights need to shift. Organizational structures need to evolve. Managers need to trust new forms of output.
In many organizations, that redesign work has barely started.
The Real Opportunity In AI Adoption
But employees are not waiting for the organization to redesign work from the top down. Inside many enterprises, the transition is already happening from the ground up.
In a recent conversation on The Future Of Less Work podcast, Bhavin Shah, co-founder of MoveWorks, described what his team is observing across global IT organizations. According to Shah, “91% are saying that a lot of the AI initiatives and innovation is actually happening from the ground up.”
The people closest to the work — finance teams, legal staff, procurement specialists — are often the ones rebuilding processes first, sometimes without waiting for top-down direction.
Shah refers to this change as “shadow innovation,” a deliberate reframing of what used to be called “shadow IT.” Instead of employees bypassing technology governance, they are experimenting with new ways to apply AI inside their workflows.
Organizations that manage that balance — governing without blocking — are likely to move faster than those trying to control every implementation from the top.
The Anthropic research signals something important about the next phase of AI transformation. The technology has arrived. Now organizations have to catch up. Closing the AI adoption gap will demand that companies rethink roles, processes and decision-making so intelligent systems can operate across the full chain of work.
