Home / Research

Outsourcing Algorithm Development: Evidence from Contractors and LLMs

Daniel Ershov, Elizabeth Lyons. 2025. CEPR Discussion Paper No. 20901.

Abstract

Algorithmic pricing is widely deployed across many markets, but firms rarely write their own algorithms; they commission them from third-party developers or potentially generate them through large language models. We study pricing algorithms commissioned from Upwork programmers and generated by two LLMs dominant through mid-2025. Across 225 generated algorithms, none uses reinforcement learning. Most are supervised-learning algorithms that predict prices directly from observables. An economic-fundamentals prompt improves the efficiency of contractor algorithms but raises LLM prices above the competitive benchmark by pushing LLMs toward misspecified demand estimation.

Main Finding

Most commissioned and LLM-generated pricing algorithms are supervised-learning routines; economic prompts improve contractor algorithms but can push LLMs toward pricing mistakes that raise prices.

Policy Relevance

Firms' delegation of algorithm design affects competitive outcomes, so algorithmic-pricing policy should account for third-party contractors and LLM-generated code.