AI in Economics Research

From Frontier Ideas to a Working Agentic Pipeline

Luka Šikić

A seminar for wiiw — The Vienna Institute for International Economic Studies

2026-07-15

By the end of the hour you leave with three things

The promise, stated on slide one.

  • A mental map of the AI-in-economics literature, in three lenses.
  • A cloneable repository implementing a minimal agentic pipeline for wiiw working-paper production.
  • A realistic cost and infrastructure picture, in euros per month, for running it.

Design philosophy — smallest viable practical. Not 18 agents and 52 skills. The minimum set of components that makes delegating research work to an AI agent both safe and productive. That minimum has five pieces.

Repository: github.com/lusiki/AI_workshop_wiiw — clone it this afternoon.

0 · Hook

The technology is already inside the research production function.

Adoption is happening — the only question is whether it is deliberate

Diffusion, at unprecedented pace

Chatterji et al. (2025) — population-scale usage across asking, doing, expressing, spanning work and personal life.

The canonical causal microevidence

Brynjolfsson, Li & Raymond (2025, QJE) — a customer-support deployment: average productivity +14–15%, gains concentrated among the newest workers, the skill distribution compressed.

These tools already change who produces knowledge work and how fast. For a research institute, the open question is not whether — it is deliberate vs accidental adoption.

A frame economists already own: this is a principal–agent problem

  • You are the principal. The model is a capable, tireless, occasionally overconfident agent.
  • Its output is cheap to produce and costly to verify.
  • Everything today — the plan gate, the permission fence, the quality score — is contract design under asymmetric information.

This single framing carries the whole hour: the binding constraint is your verification, not the machine’s generation.

1 · Three lenses on AI in economics

Tool · Instrument · Object

Lens A — AI as a tool that shifts the research production function

  • Korinek (2023, JEL) — founding taxonomy. LLMs assist across the whole pipeline; usefulness rises as verification cost falls. Verification cost is the organizing variable.
  • Korinek (2025, NBER) — the shift this seminar is about: chatbots → agents. An agent is model + tools + memory + iteration.
  • Dell (2025, JEL) — deep learning as data creation: OCR, embeddings, linkage unlock trapped sources.
  • Bahoo et al. (2025) — the adoption map across subfields.
  • Sant’Anna (2026) — the practitioner’s missing manual. Deployed in Part 2.

For wiiw: historical CESEE statistics, national gazettes and firm registries become tractable.

Lens B — AI as an instrument, with warnings

Ludwig, Mullainathan & Rambachan (2025, NBER) — the discipline paper. LLM outputs enter a pipeline two valid ways: prediction (needs an out-of-training-sample argument you cannot inspect) or estimation (needs a human-labelled gold-standard subsample). Otherwise measurement error contaminates estimates.

One rule for the room: no machine-generated variable enters a regression without a human-labelled validation set.

Lens C — AI as an object of study, and a force on the profession

  • Acemoglu (2025) — sober task-based benchmark: cumulative TFP gains < 1pp over a decade. A defensible lower bound.
  • Jones (2026) — the growth-theory counterweight. Weak links cap gains at the hard human tasks — yet automating nearly everything still delivers large acceleration. The debate is about timing and bottlenecks, not direction.
  • Brynjolfsson, Korinek & Agrawal (2026) — the research agenda if AI is transformative. A menu of wiiw-sized questions: CESEE convergence, labour reallocation, nearshoring.
  • Mullainathan (2025) — algorithms now sit inside the economy as decision-makers. The profession itself is part of the treatment group.

The bridge

Lens A: these tools raise research productivity.

Lens B: they are dangerous inputs unless disciplined. Lens C: the stakes are macroeconomic.

Part 2 shows how to get the Lens A gains while enforcing the Lens B discipline.

2 · Anatomy of an agentic pipeline

The five components — and the failure each one prevents.

The operating loop is the entire rhythm

describe, plan, approve, execute under permissions, verify, commit, then next task

You write plain-language requests and approve plans. The agent does everything between approval and the verification report. You review the diff and say commit.

Context & the plan gate

1 Context — CLAUDE.md

The project constitution the agent reads every session. Lean (~120 lines): principles, folders, commands, current state. Path-scoped rules load only when the agent touches matching files (R rules with any .R, house style with the manuscript).

Prevents: the agent forgetting decisions and reverting to generic defaults.

2 The plan gate

For any nontrivial task the agent plans in read-only mode, saves the plan to disk, and waits for approval. Plans survive context compaction and create an audit trail.

Prevents: solving the wrong problem fast — and losing the reasoning behind choices.

Permissions & verification

3 Permission-scoped execution

A settings file allowlists exactly which commands run without asking — Rscript, quarto render, read-only git. Destructive git, rm -rf, and reads under data/restricted/ are denied.

Prevents: an autonomous process with unbounded ability to touch the machine.

4 Verification

Never “done” without compiling, rendering or testing. Three mechanisms: replication-first, an adversarial critic → fixer loop, and numeric provenance — every quoted number anchored to the script that produced it.

Prevents: confident nonsense reaching the working paper.

Memory & git — and how it all binds

5 Memory & git. Corrections accumulate as tagged one-liners (mistakes made once). Git is the safety net: small commits, an advisory quality score (80 to commit, 90 for a PR, 95 excellence), worktrees for risky experiments.

Context makes the agent competent · the plan gate makes it aligned · permissions make it contained · verification makes it honest · git makes it reversible.

Remove any one and a specific, nameable failure mode returns.

Extensions for later — skills, subagents, MCP servers — are refinements of these five. Start with the constitution and two or three skills.

3 · Live demonstration

A wiiw working-paper section, produced end to end.

The artifact, and the stack

Income convergence in the CESEE EU members since accession — a small staggered event study around the 2004 / 2007 / 2013 waves. Data public, econometrics familiar: all attention on the workflow.

  • eurostat → live data pull
  • fixest (Sun & Abraham) + did
  • modelsummary tables
  • Quarto → wiiw-branded PDF
  • renv pins versions
  • targets = one-command pipeline

The workflow guide and the modern DiD estimator in did share an author — Pedro Sant’Anna.

The catch — verification fails the build on a stale number

One plain-language request, planning first: “Extend the sample through 2025, refresh Figure 2, rerun the event study, and update the results paragraph.”

  1. The agent enters plan mode, writes the plan, and waits for approval. Approval is the moment of delegation.
  2. Execution then runs inside the permission fence: only allowlisted commands (Rscript, quarto render, read-only git) proceed; destructive git and reads under data/restricted/ are blocked.

Verification finds a results number — typed by hand — that no longer matches the refreshed table. The adversarial critic flags it, a separate fixer restores the code-backed reference, and the render passes clean.

The pipeline catches the error a tired researcher on a deadline would ship.

The loop then closes: git diff → advisory quality score → commit → the compiled, wiiw-branded PDF.

4 · Pricing & infrastructure

What it costs, in euros per month.

Any laptop suffices — the model runs in the cloud

Local: a terminal, git, R, Quarto, a LaTeX distribution, Node.js ≥ 18 for Claude Code. No GPU. The repo ships a setup.R that checks all of it.

Two ways to pay (USD; verify at claude.com/pricing before the talk):

Tier Model $ / 1M in $ / 1M out Use it for
Cheap Haiku 4.5 1 5 mechanical: formatting, first-pass scans
Mid Sonnet 5 3 15 review, drafting
High Opus 4.8 5 25 identification, high-judgment steps
Frontier Fable 5 10 50 hardest long-horizon work

Batch processing −50% · prompt caching cuts repeated context to ≈ 0.1× input · Sonnet 5 intro pricing $2 / $10 through 2026-08-31. Subscriptions: Pro $20/mo ($17 annual) · Max $100 & $200 · Team ≈ $25 / $125 per seat.

Cost engineering: route models, cache context, set effort

Route by task — the 70/20/10 pattern

  • ≈70% mechanical → Haiku
  • ≈20% review → Sonnet
  • ≈10% high-judgment → Opus / Fable

Saves half or more vs running everything on the frontier tier.

Two more levers

  • Prompt caching — keep system context stable across turns (most of an agentic session is repeated context).
  • Effort per task — not maximum reasoning everywhere.

Honest order of magnitude: a full iteration of the demo pipeline plausibly costs low single-digit dollars on the mid tier at API rates. A subscription turns a working-paper habit into a flat, predictable line item. Monitor with the built-in cost commands — trust no table, including this one.

Data governance for an institute — three rules

  1. Restricted microdata never enter a prompt. The agent edits code and reads aggregated outputs; raw restricted files live in paths the permission settings deny. The model orchestrates an analysis whose sensitive inputs it never sees.
  2. Prefer Team / Enterprise / API access for institutional work; review the current commercial terms on training-data use before onboarding — not individual consumer accounts with default settings.
  3. For residency needs, the same models are available through cloud providers with EU-region endpoints, at some setup friction.

Plus one process rule: pin the tool version in a shared config so an overnight release cannot silently change behaviour mid-project.

5 · Pitfalls & the division of labour

What never to delegate.

Each pitfall ties back to a paper

  • Machine-labelled variables in a regression without a validation subsample — Ludwig, Mullainathan & Rambachan. Encoded as a check in the data skill.
  • Training contamination — anything on the public internet may sit inside the model; historical text/price exercises need an explicit leakage argument.
  • Fabricated citations — the bibliography is built only from fetched metadata, never the model’s memory of the literature; a claim-verification pass runs before submission.
  • Verification asymmetry — generation is cheap, checking is dear. The five components exist to industrialise the checking.
  • The forking-paths problem gets worse — 200 overnight specifications are a gift to honest robustness and a menace to selective reporting. Pair with preregistration or full specification disclosure.

The division of labour

Agents excel at

mechanical transformation · refactoring · formatting · first-pass literature scans · replication packaging · robustness sweeps

Humans keep

question choice · identification arguments · interpretation · ethical judgement · the final voice of the prose

Jones’s weak link is the researcher’s judgement — the complementary task that protects quality. Brynjolfsson–Li–Raymond: novices gain the most capacity, so senior researchers matter more, not less — review becomes the scarce input.

6 · Close

Three takeaways

  1. The literature gives you three lenses. Tool, instrument, object. Know the flagship paper for each.
  2. The smallest viable pipeline has five parts. Context, plan gate, permissions, verification, git. Start there, not with fifty skills.
  3. The binding constraint is your verification, not the machine’s generation. Design for cheap checking and honest numbers.

Clone it this afternoon: github.com/lusiki/AI_workshop_wiiw — paste the starter prompt, produce a compiled section on your own machine.

The starter prompt

I am starting to work on [PROJECT NAME] in this repository. [Two sentences on the paper, the data, and the tools.] I want our collaboration to be structured and rigorous even when that takes longer. Please read CLAUDE.md and the configuration in .claude/, then propose updates that fit my project. Use the plan-first workflow for every nontrivial task. Once I approve a plan, proceed autonomously and return to me only for decisions or ambiguity. Verify every output by rendering or testing before reporting it complete. For our first sessions, check in more often so I can learn how the workflow behaves.

Reading list, mapped to the seminar

Full verified notes on all 15 papers: reading/PAPERS.md · Repository: github.com/lusiki/AI_workshop_wiiw