Choosing the question
88Deciding what is worth measuring — which policy, which margin, which comparison — is the step before any tool engages, and the one on which entire research programs and policy calls succeed or fail.
Economists are unusually exposed to AI by their own admission: the job is reading, coding, estimating and writing — exactly the tasks large language models handle best. Surveys of the profession and its own automation research place economics well above the average occupation in exposure, and working economists already lean on AI daily for code, literature search and first drafts.
But exposure is not replacement. The parts of the job that carry its value — choosing the question, judging whether variation is truly exogenous, and standing behind advice when a governor or minister acts on it — are judgment and accountability tasks that current systems assist rather than perform. The realistic forecast is a profession that shrinks at the routine end while its senior judgment work, if anything, gains leverage.
Moderate
Share of the work a machine could do
A substantial share of the working day — data preparation, routine estimation, chart production, literature review, first-draft prose, short-horizon forecasting — is already automatable, and much of it is already automated in practice. What resists is the core a client actually pays for: framing questions, judging causal claims, integrating models with institutional reality, and owning the advice. Roughly a third of the job is machine work now; the accountable judgment on top is not.
Scored from the tasks, not the job title. Lower is safer.
Jobs AI cannot take →What machines cannot take
Judgment when the model breaks
84Crises are precisely the moments when fitted relationships fail; 2008 and the pandemic both rewarded economists who could reason beyond their models while the models were wrong.
Accountability for public decisions
80Rate-setting committees, budget scorekeepers and expert witnesses answer by name for their judgments; no institution yet delegates a vote, a fiscal costing or sworn testimony to a model.
Persuasion and political trust
74Advice only matters if a minister, board or public believes it; building that trust — in briefings, testimony and plain language — is relational work automation does not reach.
Institutional and historical context
66Knowing how the data are constructed, how the last crisis actually unfolded, and what a given statistic misses in a given country is tacit knowledge that models consume but do not create.
What they already take
Charts, tables and report production
85Formatting output, drafting boilerplate sections and producing standard chart packs — a real fraction of junior economists' hours — is already handled end-to-end by scripting and AI assistants.
Data cleaning and routine estimation
78Merging datasets, standardizing variables and running well-specified regressions is increasingly automated; AI coding assistants now write serviceable Stata, R and Python for exactly these tasks.
Literature search and first drafts
72Large language models summarize working papers, surface related literatures and produce competent first-draft prose — compressing tasks that consumed weeks of research-assistant time.
Short-horizon forecasting
60Nowcasting GDP and inflation from high-frequency data is largely a machine-learning problem already, and central banks run such models routinely; the human role has shifted to interpreting and overriding them.
How the work is changing
From surveys to administrative big data
The unit of evidence is shifting from thousand-person surveys to entire populations of tax records, transactions and satellite imagery — the approach Raj Chetty's Opportunity Insights team used to map American mobility from anonymized IRS data, now spreading through statistical agencies worldwide.
AI as the research assistant
Coding, literature review and drafting are moving to AI tools, hollowing out the traditional research-assistant apprenticeship; the field is actively debating how the next generation learns craft judgment when the routine work that taught it is automated.
The economist as engineer
Following Alvin Roth's manifesto of the design economist, market design moved from paper to production: kidney-exchange algorithms, spectrum auctions and online ad markets are built and run by economists, shifting prestige from explaining economies toward constructing them.
Communication becomes the job
Central banks now steer economies partly by speech — forward guidance, fan charts, press conferences — and social media has made public explanation a core skill; the profession increasingly promotes economists who can hold a room, not only a regression.
New jobs branching off
Tech economist / market designer
Designing auctions, pricing, and marketplace rules inside technology firms — a career path that barely existed before Google's ad-auction work under Hal Varian and is now among the largest destinations for new economics PhDs.
Behavioral insights specialist
Running nudge units — a role invented when the UK's Behavioural Insights Team was founded inside the Cabinet Office in 2010 and since copied by dozens of governments — applying behavioral economics and trials to tax letters, pensions and public health.
Climate economist
Pricing carbon, modelling transition scenarios and stress-testing financial systems against climate risk, a specialty crowned by William Nordhaus's 2018 Nobel and now demanded by central banks, insurers and ministries alike.
Economic data scientist
Building nowcasting systems and policy dashboards from payments data, job postings and satellite feeds — the boundary role between econometrics and machine learning that statistical agencies and fintech firms are hiring for fastest.
Outlook
The honest reading of the evidence is that economics is a high-exposure, high-judgment profession: more of its daily tasks can be automated than in most fields, and more of its value sits in the residual that cannot. The routine layer — cleaning, coding, drafting, standard forecasts — is being absorbed quickly, and entry-level roles built on that layer will thin.
What expands is the leverage of the judgment layer. An economist with AI assistance can test more specifications, read more literature and produce more scenarios than a whole team could a decade ago — but the choice of question, the causal verdict and the advice still carry a human name. Institutions that stake decisions on economic analysis show no movement toward removing that name.
The precedent worth remembering is that economists have absorbed their own automation before: the electronic computer erased the profession's calculating clerks and made econometrics an industry; cheap regression software democratized estimation and raised the premium on knowing what to estimate. AI is the same trade at larger scale — fewer hands on the data, more weight on the judgment.
Keep exploring
More in Business & Finance
🚀
Entrepreneur
The person who starts the company — spotting the gap, bearing the risk and answering for payroll, from Assyrian caravan financiers to venture-backed founders.
AI-resistant 88 💹Investment Banker
The dealmaker who prices companies and moves capital — a trade running from Medici Florence to today's pitch decks, paid for trust when billions change hands.
AI-resistant 42 🧾Accountant
The keeper of the books: heir to a craft so old it invented writing itself, now negotiating with the software built to automate it.
AI-resistant 35 📣Marketer
The professional who creates demand — from Pompeii's painted walls and P&G's 1931 brand-man memo to the auction-driven feeds of the digital era.
AI-resistant 38 🗺️Product Manager
Decides what a company should build next, and why — turning customer needs, business goals and engineering limits into one shared plan nobody else fully owns.
AI-resistant 50