How We Analyze AI Job Impact

Most "will AI replace my job?" tools give you a single number and call it a day. That is not useful. AI does not replace entire jobs at once — it replaces specific tasks within jobs, often the most routine ones first.

A radiologist does not disappear overnight. But AI can already read certain scan types faster and more accurately than humans in controlled studies. That changes the proportions of the job — less time on routine reads, more time on complex diagnoses. Understanding that shift is far more useful than a yes-or-no prediction.

Our analysis works the way labor economists and automation researchers model job impact: at the task level, with honest framing about timelines and uncertainty.

Every job is a collection of recurring tasks. An accountant does data entry, prepares tax returns, performs financial analysis, advises clients, and makes regulatory compliance judgments. Each of these tasks has a different exposure to AI automation.

We decompose each job into its atomic sub-work-tasks — typically 8 to 20 per role — using occupational databases, job description analysis, and expert review. Each task is then scored independently against current AI capabilities.

The job-level score is a weighted aggregation of all task scores, where weights reflect the approximate proportion of time spent on each task. This means a job where 80% of time is spent on high-risk tasks scores very differently from one where 80% of time is spent on low-risk tasks — even if both contain some automatable work.

There is an important distinction between cognitive/digital AI and physical automation (robotics). Current AI — large language models, vision models, code generation tools — excels at information processing: reading, writing, analyzing data, generating content, and pattern recognition. These are the capabilities that are advancing fastest.

Physical tasks — construction, plumbing, nursing care, electrical work, cooking — require dexterous manipulation, spatial reasoning in unstructured environments, and real-time physical adaptation. Robotics capable of performing these tasks reliably and affordably at scale is a decade or more away in most fields.

Our analysis reflects this honestly. A data entry clerk (mostly cognitive tasks) scores very differently from an electrician (mostly physical tasks), even though both are legitimate jobs. We do not inflate risk for physical roles just because AI is in the headlines. When robotics does become relevant to a role, we update the scores.

AI capabilities are changing rapidly. An analysis from 12 months ago may already be outdated. Our research pipeline runs monthly and incorporates three types of data:

Published AI capability benchmarks — we track major model releases and what they can demonstrably do, filtering hype from real-world performance. Peer-reviewed automation exposure research — academic studies from labor economists that quantify which tasks and occupations are affected by specific technology changes. Deployed AI systems — tracking which AI tools are actually being used in workplaces, not just demonstrated in labs.

When a significant development happens between scheduled updates — a major new model release that changes the ceiling for specific tasks, or a large-scale deployment in a particular industry — we bring that update forward rather than waiting for the next cycle.

Every job analysis page shows when that job was last reviewed, so you can see how recent the data is.