Anthropic's Own Data Reveals Which Jobs AI Is Already Replacing, Not Predicting, Measuring

Every job displacement forecast comes with a built-in escape hatch: "this is speculative." Anthropic just closed that door by publishing data from millions of real Claude interactions to measure which jobs are already being automated today, not in 2035. The result is cleaner and more unsettling than any prediction model.

What Anthropic Actually Did

Rather than running surveys or building speculative models, Anthropic analyzed actual Claude usage patterns across real workflows. They call this "observed exposure," a methodological shift from previous AI labor studies that modeled vulnerability by task similarity. Anthropic instead measured how Claude is currently being used to partially or fully replace human work.

The data comes from millions of real interactions. Claude customers aren't self-reporting; Anthropic is observing where the model shows up in job workflows at scale. If your role involves tasks that Claude regularly handles, you show up in the exposure map.

This is fundamentally different from McKinsey or Goldman Sachs predicting which jobs AI might disrupt. Those studies project. Anthropic measured.

The Jobs Most Exposed

Anthropic's data points to five categories experiencing the highest observed exposure:

Computer Programmers. The most exposed group, and unsurprising. Claude handles code generation, debugging, refactoring, and documentation. The shift isn't theoretical; it's happening in IDEs and CI/CD pipelines right now. A programmer's job hasn't disappeared, but large portions of the task stack have been reassigned to a model.

Customer Service Representatives. Support workflows already run on chatbots, and Claude is increasingly the backbone. Intake, routing, escalation, and first-response handling all show high exposure. The gap between customer-facing support and Claude-generated support is narrowing faster than any other category.

Data Entry Clerks. This role was already fragile, but Claude accelerates the end. Document parsing, information extraction, and structured data creation are core Claude use cases. Automation that would have required custom ETL pipelines now runs on a general-purpose model.

Medical Record Specialists. Healthcare is using Claude for chart documentation, clinical note synthesis, and coding compliance. The exposure here is high because large portions of the role are document processing and structured transcription, both Claude-native tasks.

Market Research Analysts. Report generation, data synthesis, competitive analysis, and insight summarization are all Claude workflows. The analyst role doesn't disappear, but the labor-intensive data-to-document pipeline is collapsing.

Why "Observed" Matters More Than "Predicted"

Previous studies asked "could AI do this?" Anthropic asked "is Claude doing this, and at what scale?" The answer can't be argued with the way forecasts can. If Claude is already handling 40% of a programmer's daily tasks, that's not a prediction. That's data.

This shifts the conversation. Predictions are debatable. Observation is fact. When Anthropic says customer service has high observed exposure, they're not modeling, they're reporting. The exposure is already here.

The methodological shift also removes the false comfort of timeline. Studies that predict job disruption in 2040 give time to adapt. Observed exposure says this is happening now, within current workflows, today.

What to Do If You're in One of These Roles

The practical response isn't panic, but action. If your role shows high observed exposure, you have options:

Shift upward in the stack. Stop defending the tasks Claude is better at. Own the work that requires judgment, client relationships, or domain context. Customer service representatives should move toward strategy and escalation. Programmers should move toward architecture and tooling decisions that machines can't make. The work exists; it's more valuable and harder to automate.

Specialize in oversight. If Claude handles 40% of your job, become the person who audits Claude's work. Quality control, validation, and human approval of model output is increasingly valuable. You move from execution to verification.

Build on top of the model. Use the freed-up time to learn. If you're a programmer spending less time on boilerplate, learn the infrastructure or product strategy side. If you're in customer service, understand the systems you're supporting and move into product management or operations.

Retrain now. The window for this is closing. Roles with lower exposure still exist, and transition is easier before your current role becomes untenable. This isn't a guess, it's data.

Cleoops Take

The Anthropic study matters because it removes the permission to wait. We can argue about whether jobs will change in the future, but Anthropic just showed us they're already changing. The five roles above aren't at risk in 2030. They're restructuring right now. Workers in these roles have an advantage: the market hasn't fully reacted yet. That window closes fast.