AI Won't Replace Your Experts. It Will Need One to Check Its Work.

Will AI replace economists? The honest answer is no, and it isn't close. Task-based estimates put an economist's automation risk near 19%, low for a desk job, because the part that matters is judgment and explaining findings to people, which is exactly what AI is worst at. The job shifts toward interpretation and advising. It doesn't disappear. (Same pattern as bank tellers and ATMs: the machines took the counting, the people moved to higher-value work, and headcount went up, not down.)

But one finding buried in that question should change how any organization uses AI for analysis, in economics or anywhere else.

The model has opinions. IMF research found that GPT-3, asked about wealth taxes, argued against them, not from a neutral reading of the evidence but from the leanings baked into its training data. The researcher's phrase for it is "Keynesian machines versus neoclassical ones." Whatever the training corpus leaned toward, the model leans too, and it says so with the same confidence either way.

That flips the expert's job around. The value isn't producing the analysis anymore. It's catching where the analysis is quietly wrong. "The model says X" is cheap. "The model says X, and here's why that's off in this case" is the work.

Which is the part employers should sit with. AI is a tool for your experts, not a replacement for them, and that only holds if you can see how it reached an answer. Run it as a black box and you haven't bought a faster analyst. You've bought a confident one that's sometimes wrong and never flags it. The fix isn't a smarter model. It's a setup where the answer is grounded in sources you control and traceable enough that a person can check it.

We pulled the full data on the economist question, the risk scores, the task breakdown, and the subfields most and least exposed, into one piece in the Knowledge hub: Will Economists Be Replaced by AI? What the Data Says.

If the part you care about is keeping AI analysis governed and checkable inside your own network, that's what we build.