Guest post by Former MIT GOV/LAB Practitioner in Residence Luke Jordan and co-authors. 

Authors: Anna Powell-Smith, Luke Jordan, Tiago C Peixoto

Speculation about AI’s impact on jobs is giving way to hard data – and urgent warnings. Recently, Anthropic CEO Dario Amodei predicted that within a year, AI could be writing nearly all new code, potentially eliminating thousands of software development jobs. While other tech leaders, including IBM’s CEO, dispute the timeline, they acknowledge AI’s expanding workplace presence.

Until now, analyses of AI’s labor market impact relied primarily on projections rather than empirical evidence. However, Anthropic’s recently released dataset, based on its Claude language model, provides a revealing glimpse into AI adoption patterns across professions – though a critical question remains: which local labor markets are likely to see the highest occupational AI impact in the short term, and how can policymakers respond?

To explore this question, we combined Anthropic’s empirical data with national labor statistics from 33 European economies, plus local US labor market statistics. Our approach extends previous studies – such as OpenAI’s occupational exposure analysis and the IMF’s global assessment – by leveraging actual usage patterns rather than theoretical models or automation estimates, offering the first geographic comparison of AI’s potential impact on Western labor markets based on usage data.

Anthropic’s Economic Index Data – and Methodology

Our method is straightforward, combining Anthropic’s data on occupational usage with local labor force statistics. We extend the method used by Anthropic to compare occupational representation in Claude data to representation in the US labor market.

We started with Anthropic’s newest data on Claude usage, which maps 1 million real-world Claude conversations to their most relevant task in the US Department of Labor’s O*NET database, and from there to specific occupations.

Anthropic compares these usage patterns to the actual distribution of these occupations in the US workforce. We extend this comparison to the distribution in European and local US workforces, as follows:

• European countries: We map O*NET occupation codes to the occupation codes used in EU labor statistics via official crosswalks, and combine this with the country-level fraction of workers in each high-level occupation category.
• US counties: We map O*NET occupation codes to US census occupation codes via official crosswalks, and combine this with the county-level fraction of workers in each high-level occupation category.

This allowed us to compare occupational representation in Claude data to actual occupational distribution across multiple local labor markets.

Our methods face limitations, being reliant on occupation estimates from surveys and usage data from a single platform, and assuming no local barriers to AI usage. Most critically, our analysis is of imputed local labor market use – based on occupational representation in a sample of Claude usage – not actual use, for which data is not yet available.

Key Findings: Where Is Likely Labor Market Impact of AI Highest?

1. Likely Occupational Impact Doesn’t Always Align with GDP Per Capita

A key finding is that a country’s likely occupational usage of AI does not always correspond with its GDP per capita. For example, Ireland – one of Europe’s richest nations – has a lower proportion of AI-intensive jobs than Denmark. National wealth alone does not determine AI adoption, confirming previous findings based on theoretical AI usage. Instead high-income economies with a strong presence in finance, IT, and consulting may see greater AI impact in the short term, whereas those relying on manufacturing or tourism may see less short-term AI impact.

2. In Europe, Northern Countries Lead in Likely Labor Market AI Impact

Our analysis suggests that the Netherlands, Sweden, Denmark and Finland have the highest levels of jobs likely to see high AI usage in Europe. These economies are characterized by large shares of professionals in knowledge-intensive sectors such as finance, technology, and academia—fields that are a far higher share of AI tools like Claude’s use than other occupations (and, in the Netherlands’ case, the home of Europe’s largest tech champion, ASML). These results, as well as the general finding on GDP per capita, also reinforce that estimates of AI’s potential labour market impact cannot rely on old frameworks based on different technology, which would have predicted the reverse, i.e., higher impact in routine lower-skilled tasks and an inverse alignment with GDP per capita.

3. Southern and Eastern Europe Have Lower Likely Impact

In contrast, Turkey, Romania, Greece, Bosnia and Herzegovina, and Serbia exhibit the lowest levels of likely occupational AI usage. This suggests that their labor markets are dominated by industries less reliant on AI-intensive tasks, such as manufacturing, agriculture, and traditional service sectors. While AI adoption in these regions may lag, it also presents an opportunity for policymakers to proactively introduce AI capabilities in ways that enhance productivity rather than disrupt employment.

However, this uneven geographic distribution of AI’s likely labor market impact aligns with Liu’s (2024) findings on “premature de-professionalization,” suggesting that regions with fewer high-skill service jobs may face a double disadvantage: missing out on early AI-driven productivity gains, while also seeing reduced opportunities to develop these sectors – should additional automation take place, and should competition from early AI adopters intensify.

4. Potential AI Impact Doesn’t Map Neatly Onto Digital Infrastructure

One might expect that countries with the best internet infrastructure would also be those most likely to adopt AI tools like Claude at scale. But the data suggests a more nuanced story. While there’s a mild correlation, it’s far from deterministic. Spain, for example, boast nearly universal broadband access – yet sits relatively low in likely occupational AI usage. Meanwhile, countries like Denmark and Sweden combine relatively high AI usage potential with more modest infrastructure. This suggests that digital readiness alone won’t drive future AI adoption. Instead, sectoral structure, workforce composition, and the prevalence of AI-relevant occupations – like those in finance and consultancy – appear to matter more. The “last mile” to AI adoption may be professional, not physical.

5. In the US, Urban Centers See Greater Likely Impact

The U.S. labor market is already seeing disparities in AI adoption. Our county-level analysis confirms, as have previous estimates by Brookings, that tech and knowledge hubs like San Francisco, Seattle, and Washington, D.C., show more higher likely occupational AI usage than rural counties. The software industry – one of the biggest drivers of this urban skew – may now be at a crossroads.

But even in this high-level analysis, AI’s footprint isn’t limited to the usual suspects. While places like Silicon Valley and D.C. predictably light up the map, some high potential occupational AI use, based on county-level labor estimates, shows up in less expected areas – like pockets of Colorado and Texas. These aren’t your typical tech darlings, yet still have relatively high levels of the computer workers and skilled professionals whose tasks are heavily represented in the Claude usage data. Meanwhile, heavyweight metros like Los Angeles and Chicago register more modest overall levels of potential occupational usage. The result? A labor market map that defies the standard coastal-versus-heartland narrative – and hints at a more complex geography of disruption.

While forecasts differ – some, like Anthropic’s CEO, envision AI handling most coding tasks in the near term; others, like IBM’s Krishna, anticipate a slower shift – what’s clear is that software development is no longer insulated. Google already reports that AI now generates a quarter of its new code. And according to Anthropic’s own usage data, 57% of occupational AI use appears to be augmentative, while 43% involves full or partial automation.

This uncertainty underscores the need to understand not just how AI will affect jobs, but where. Tech hubs may see AI boosting output without eliminating roles – at least in the short run. But the longer-term trajectory remains unclear, especially as companies like Meta pursue AI agents capable of mid-level engineering by 2025. The battle over AI’s labor market effects – between displacement and augmentation – is increasingly unfolding within high-skilled professions and urban economies, rather than along the traditional rural–urban divide.

Policy Implications: How Governments Can Respond

1. Preparing High-Usage Economies for AI’s Workforce Shift

Countries likely to see high usage – such as Luxembourg, Sweden, and the Netherlands – are likely to experience the most significant workplace transformations in the near future. We do not yet know whether this usage will signal future productivity gains, or displacement effects – but policymakers in these markets should be prepared for short-term impact.

To stay competitive, governments should prioritize policies that:

• Monitor employment trends closely, tracking whether AI augments productivity or contributes to job displacement.
• Support lifelong learning and reskilling programs for professionals in AI-intensive sectors.
• Encourage responsible AI adoption in the workplace to ensure AI complements human work rather than replaces it.
• Start thinking about services demand, if AI does augment rather than replace professionals, expanding demand e.g. with more digital public services, will be the difference between maintaining professional employment and not.

2. Bridging the Digital Divide in Lower-Usage Regions

For countries with likely lower occupational AI usage, the challenge is ensuring they don’t fall behind. Policymakers in Turkey, Romania, and Greece should focus on:

• Support and invest in AI integration in sectors like agriculture, logistics, construction and public services, where AI could drive efficiency gains. This should include supporting businesses in rural areas and lower income urban areas to integrate AI tools.
• Investing in digital skills training to prepare their workforce for future AI adoption, including workers outside metropolitan areas.
• Building stronger AI infrastructure – such as broadband access and cloud computing resources – to enable businesses to incorporate AI into their operations, including in rural areas.

Conclusion: Beyond Historical Patterns

When analyzing AI’s potential impact on labor markets, we must recognize a fundamental difference between today’s AI systems and previous waves of automation. It is the high-skilled, professional jobs being impacted first – not, like for previous technologies, the routine ones. This finding suggests that past estimates of a job’s automation exposure may be – at least for the moment – an erroneous way to assess AI’s unique pattern of impact on labor.

The data offers some cause for optimism: much of AI use appears to be augmentative – even in high-risk professions like software development – and the regional economies where disruption is most likely are also those best equipped to adapt, given their resources and institutional readiness.

But the data also lays out the challenge ahead, to ensure that those frontier regions do not enter a high-efficiency low-demand job-loss cycle, and that less exposed regions are able to both invest and support the proactive workforce planning needed to harness AI for inclusive growth.

We also welcome Anthropic’s decision to share this valuable data, and would welcome more detailed data at national level, to help policymakers respond. In particular, more data on augmentation versus automation at local level would be invaluable.

With AI adoption accelerating rapidly, the window for proactive, evidence-based policy is now open – but may not remain so for long. By understanding both potential geographic differences in AI impact and the fundamental differences between AI and previous technologies, we can better prepare for the complex transition ahead.

Photo by The New York Public Library on Unsplash.