On March 11, Anthropic announced the formation of the Anthropic Institute, a dedicated research body focused on understanding what powerful AI actually does to the economy, the legal system, public safety, and broader society. The announcement coincided with the publication of new labor market research that is worth every business owner reading, even if most headlines buried it under shinier product news.

The Institute is led by Anthropic co-founder Jack Clark, who has taken on a new role as Head of Public Benefit. It brings together three existing Anthropic research teams: the Frontier Red Team, which stress-tests AI systems to understand their limits; Societal Impacts, which studies how Claude is actually being used across the real world; and Economic Research, which tracks what AI is doing to jobs and economic output. Clark has also hired economists and legal scholars from Stanford, UVA, and Google DeepMind to staff out the research agenda.

The launch matters not because it is unusual for a big AI company to create a research institute. It matters because the research it published alongside the announcement is unusually honest about what is and is not happening to the workforce.

What the data actually shows

Alongside the Institute announcement, Anthropic published its most detailed labor market analysis to date, using a privacy-preserving dataset drawn from actual Claude usage patterns. The headline finding is one that will surprise people on both sides of the AI-will-destroy-jobs debate.

Coding and computer science tasks account for 35 percent of conversations on Claude’s platform. For occupations in the “Computer and Mathematical” category, theoretical models suggest that 94 percent of tasks are “exposed” to AI tools like Claude. The number sounds alarming until you look at what is actually happening in practice.

Actual coverage by Claude currently sits at around 33 percent of those theoretical tasks. There is a 60-point gap between what AI could theoretically do to knowledge work and what it is measurably doing right now.

The researchers call this “observed exposure” versus “theoretical exposure.” It is a more honest framing than most of what gets published on AI and jobs, because it distinguishes between what a model can do in a controlled demonstration and what people are actually using it for at work.

The ‘Great Recession for white-collar workers’ scenario

The research names the risk scenario everyone in the knowledge economy should be thinking about, which is why it is worth understanding clearly.

During the 2007 to 2009 financial crisis, the US unemployment rate doubled from roughly five percent to ten percent. Anthropic’s researchers note that a comparable doubling in unemployment for the top quartile of AI-exposed knowledge workers would be clearly detectable in their framework. That would mean unemployment in the most AI-affected occupations rising from around three percent today to around six percent.

It has not happened yet. The researchers found no evidence of a systematic rise in unemployment for workers in highly AI-exposed occupations. What they do see is slower projected growth in those occupations and reduced hiring for entry-level roles, which suggests the effect may be appearing at the point of entry into the workforce rather than in mass layoffs of existing workers.

That distinction matters for how businesses and individuals should respond. Workers already in mid-career roles have more time than the doomscrolling headlines suggest. People entering the workforce or early in their careers face a more compressed window to develop skills that are hard to automate.

What businesses should take from this

The gap between theoretical exposure and observed exposure is not permanent. As AI models get more capable and as organisations get better at integrating them into workflows, that 33 percent will rise. The question is how fast, and whether the businesses and workers inside those organisations are ready when it does.

The Anthropic Institute’s monitoring framework gives researchers and policymakers early warning. But businesses cannot wait for a report to tell them the gap is closing. They need to be ahead of that curve, not catching up to it.

For data professionals and knowledge workers, the lesson from this research is not that AI is overhyped and nothing will change. It is that the change is happening selectively and at a speed that creates a window, not permanent safety. The occupations showing slower growth and reduced entry-level hiring now will look different in 18 to 36 months if observed exposure continues climbing.

The most defensible position for any professional in an AI-exposed role is to be competent with the tools that are changing their work, not just aware of them.

The EDNA perspective

Enterprise DNA has spent a decade building data skills across 220,000 professionals in 50 countries. The research coming out of the Anthropic Institute validates something the data education community has known for a while: the businesses and individuals who invest in genuine competency rather than surface-level familiarity with AI tools are the ones who maintain leverage as automation spreads.

The 33 percent observed exposure figure is not a ceiling. It is an early reading. The organisations that will navigate the next phase well are the ones building data literacy at every level of their teams, not just in their technical departments.

If you are a business leader thinking about where to invest in your team, this research points clearly toward capability building over tool acquisition. The tools are commoditising fast. The judgment to use them well is not.

What This Means for Business

For executives and business owners, the Anthropic Institute’s research provides three actionable signals:

First, you have more time than the headlines suggest, but not unlimited time. The observed exposure gap is real today. It will narrow. Use the window.

Second, the risk is concentrated at entry-level hiring, not existing workforce. If you are worried about AI disruption to your team, the near-term story is more about who you hire and what you invest in developing, less about the near-term displacement of experienced staff.

Third, data literacy is now a defensive skill, not an aspirational one. Organisations where employees understand how to work with data and AI tools are less exposed to the disruption risk than those where those skills are concentrated in a small technical team.

For individuals, the research is a strong argument for investing in hands-on AI and data skills now, before observed exposure closes the gap with theoretical exposure. Enterprise DNA Learn was built for exactly this: practical data skills that remain valuable when AI handles the routine parts of knowledge work.