The Tasks AI Can't Do Yet, and How to Build Your Job Around Them

In 2016, radiologists were the job AI was going to make obsolete. Then the reading part of the work got augmented, the judgment, procedures and accountability stayed with humans, and demand for radiologists grew.

That's why "is my job safe from AI?" is the wrong question, and why job-title safety lists mostly miss. Jobs are bundles of tasks, and AI unbundles them. The useful question is which of your tasks are exposed, and whether you can rebuild your role around the ones that aren't.

Five kinds of work that are holding up

The first is physical dexterity in unstructured environments. Robotics lags language models by years, which protects skilled trades, healthcare procedures and field engineering.

The second is accountability, legal or professional. Somebody has to sign the audit, the prescription, the structural calculation, the deployment to production. AI can draft them, but a human has to own them, and roles built around that sign-off have a structural moat.

Third comes high-stakes judgment with incomplete context, where being wrong carries consequences a model can't bear: incident command, negotiations, clinical edge cases, executive tradeoffs.

Fourth is trust. In enterprise sales, therapy, management and partnerships the relationship is the product, and people demonstrably prefer people.

Fifth is frontier or tacit expertise, where training data is thin: novel research, bleeding-edge engineering, and deep knowledge of how an organization actually works. AI compresses the middle of the expertise distribution. The frontier is much harder for it to reach.

Why parts of law score as exposed and plumbing doesn't

A task is exposed when it's digital, routine and well specified, and when its output is easy to review. Drafting standard documents, producing routine code, summarizing, formatting, first-pass screening, moving data between systems and answering questions that are already documented all fit.

Our test: if you could hand the task to a very fast, well-read intern and check the result quickly, AI will absorb it.

Look at what the test ignores: how skilled the task feels, how long it took you to learn, and what it pays. Exposure comes from how a task is structured, so parts of law, finance and medicine end up exposed while parts of plumbing don't. We know it's a rough test that will misjudge some tasks. We'd still trust it over any list of safe job titles.

Audit your own week

Sort a typical week into those five categories. Most professionals find 30 to 60% of their time goes on exposed tasks, which is less a doom score than a list of things to redesign.

Automate your own exposed tasks first with the AI tools your employer is rolling out anyway and put the hours you get back into the durable work. Take the ambiguous project. Own the sign-off. Build the relationships, and go deeper than the documentation.

People who run that loop on their own get called "AI-forward", and they get promoted. People who wait for someone else to run the audit on them get called "headcount". Same starting point. Timing was the whole difference.