AI for Pricing: What to Entrust to an Agent and Where Causal Analysis Is Unavoidable

Artem Duplinsky23 May 2026

Problem

A pricing manager gives a sales table to an LLM agent (Claude Code in the talk), which fits a log-log regression and recommends a price. On observational data this can fail without any warning. Prices are cut when demand fades (clearance, end of a collection), so price and demand move together through hidden drivers, and the elasticity estimate comes out positive. The optimizer then says "raise the price without limit", and the agent's own text can still sound reasonable.

Example from the article: on 30 weeks of a luxury SKU, the agent estimated β=+1.35 (R2=0.31) and recommended p→∞. An IV estimate on three clothing SKUs gave −1.55 to −1.95.

Solution in brief

Let the agent do the mechanical work: EDA, prototypes, optimizer code, test monitoring. Estimate elasticity with an identified causal model (DAG, instrument, fixed effects), not a plain regression. Decide what to delegate by two axes: build vs optimize, reversible vs irreversible decision.

Questions this case answers
  • How to estimate the response of demand to a controllable price-like variable from observational data when that variable was changed for reasons linked to demand
  • How to decide which parts of a causal analysis an LLM agent can run alone and which need an identified causal model reviewed by a human
  • How to estimate price elasticity from observational sales data without a positive-elasticity artifact
  • How to decide which pricing tasks an LLM agent can run alone and which need a human with causal analysis

Tools and methods

  • Instrumental variables with SKU x week fixed effects — price elasticity with promo as the instrument
  • Causal graph (DAG) — where confounders and reverse causality sit
  • Double machine learning, CCE, IPW, uplift — the causal engine layer
  • Geo-A/B, switchback, per-SKU rolling, synthetic control — measuring effects without randomization
  • Claude Code — agent for EDA, optimizer, dashboard, post-mortem
Continue reading — it's free

Google or a one-time email code. No password.