Much of the current discussion around artificial intelligence focuses on how quickly AI is transforming knowledge work. Developers are coding faster. Analysts are summarizing data in seconds. Legal teams are drafting documents with AI assistance.
But a recent research paper from Anthropic highlights a far more interesting reality.
Across the labor market, there is a significant gap between what AI could theoretically do and what it is actually being used for today.
In many occupations where large language models could potentially perform a majority of tasks, real-world usage remains surprisingly low. The technology is advancing quickly, but the workflows that need to absorb it are evolving far more slowly.
Nowhere is that gap more visible than in the industries responsible for building the physical world.
The Pattern of AI Adoption
Anthropic’s research shows a clear pattern in how AI is spreading across different professions.
Adoption is highest in areas where work is already digital and information-heavy. Software engineering, finance, legal analysis, and office administration have seen rapid uptake. These roles revolve around text, data, and structured digital artifacts, exactly the type of information large language models are designed to process.
In these environments, AI functions primarily as a productivity tool. It helps people write faster, summarize complex documents, generate code, and analyze information.
But when we look at sectors like architecture, engineering, construction, infrastructure, and energy projects, the picture changes dramatically.
These industries show very low levels of observed AI usage, even though many of their tasks appear highly suitable for automation.
At first glance, that seems puzzling. After all, these are complex industries that generate enormous amounts of information — designs, specifications, schedules, procurement data, engineering reports, and meeting discussions.
So why hasn’t AI taken hold here yet?
The Real Challenge: Modeling Work, Not Documents
The answer lies in the nature of the work itself.
In knowledge professions, most tasks revolve around interpreting and producing information. AI can read text, summarize ideas, and generate responses with impressive accuracy.
Capital projects and engineering delivery are fundamentally different.
They are not primarily about documents. They are about systems of interdependent decisions.
Design choices affect procurement timelines. Procurement delays affect construction sequencing. Construction sequencing affects commissioning schedules. A change in one discipline can cascade across an entire project.
This type of work requires reasoning about dependencies, constraints, and coordination across dozens of stakeholders.
And that is not something AI can do by simply reading documents.
It requires a model of how the system itself operates.
The Missing Layer in Enterprise Software
Most organizations already run sophisticated systems of record. Tools like Primavera, SAP, Procore, and various engineering document platforms capture enormous volumes of data about what is happening in a project.
But these systems were designed to store transactions and records.
They capture what happened, not how work flows through the system.
For AI to truly support execution in complex industries, it needs something different: a structured representation of how decisions interact across the lifecycle of a project.
In other words, AI needs context.
Without that operational context, AI can summarize meetings or retrieve documents. But it cannot reason about how a design change might affect procurement, or how a delay in fabrication could propagate through a schedule.
