■ HIGH RISK ■ Science & Research
A good chunk of it, yes — particularly the applied forecasting and analysis work that makes up most economist employment. Policy judgment and institutional credibility survive; producing another quarterly outlook does not.
“AI can model economies. But economists were already wrong most of the time, so who'd notice?”
Our AI replacement risk score — how we score jobs
Most economists don't theorize; they do applied work. Bank and consultancy economists forecast growth, inflation, and rates, write commentary, and brief clients. Government economists appraise policy, run impact assessments, and produce statistical analysis. Academic economists design identification strategies, clean data, run regressions, and write papers. Across all three, the daily material is data wrangling, econometric modelling, literature review, chart production, and writing — a great deal of writing.
Nearly every one of those is a strong AI use case. Models handle data cleaning and merging, specify and run econometric estimations, generate the robustness checks a referee will demand, summarize literatures faster than any research assistant, produce publication-quality charts, and draft the prose around results. Machine learning forecasting methods routinely compete with traditional structural approaches on short-horizon prediction. Firms that once retained a house economist to produce a quarterly outlook can now generate a competent one automatically — and the honest observation, which our 60 reflects, is that consensus macro forecasts have a poor accuracy record anyway, making them hard to defend as expensive human output.
What survives is narrower but real. Choosing which question matters and constructing a credible causal identification strategy remains genuinely hard intellectual work — the difference between a correlation and a policy-relevant finding is a research design, not a regression. Institutional roles carry weight because of who says it: a central bank economist's judgment shapes markets partly through the institution's authority. Advising ministers involves political feasibility as much as welfare analysis. Novel theory, and interpreting structural breaks where historical data stops being a guide, also resist. But an applied economist whose output is a forecast and a chart pack is competing directly with software that never asks for a bonus.
Automatability: our editorial assessment of current and near-term AI capability
Pressure is building now and should be serious by around 2030. Routine forecasting, commentary, and modelling roles in banks and consultancies are the first to compress, since their output is precisely what generative tools produce cheaply. Central banks, treasuries, and international institutions retain economists longer because their value includes institutional authority and accountability, but even there team sizes are likely to shrink as analytical work automates.
On short horizons, machine learning approaches are competitive with and sometimes better than traditional models, and they update continuously. That's less impressive than it sounds, because human macro forecasting has never been very accurate either. The honest framing is that AI matches a mediocre baseline at near-zero cost, which is enough to displace a lot of paid forecasting work.
It depends heavily on the target. If the goal is producing standard applied analysis, the return is weakening — that work is automating. If the goal is research design, policy leadership, central banking, or academic theory, the training still buys something a model can't replicate: the judgment to know which question is worth asking and whether an answer is credible.
Applied roles whose output is recurring analysis: bank and consultancy economists producing quarterly outlooks, commentary, and client chart packs. Research assistant and junior analyst positions are also heavily exposed since data work and estimation were their core. Least exposed are senior policy advisers, regulators, and institutional economists whose authority is part of the value.
Causal inference and experimental design first — identifying what actually causes what is the discipline's most defensible contribution. Then machine learning methods, so you're using the tools rather than being replaced by them. Finally, communication and domain depth: an economist who understands a specific industry and can advise decision-makers is far harder to substitute than one who runs regressions.