Source: AI in the Enterprise (OpenAI)

OpenAI's 2025 guide distilling seven enterprise adoption lessons from Morgan Stanley, Indeed, Klarna, Lowe's, BBVA, Mercado Libre and OpenAI itself, starting with evals.

Contents

Summary

Seven lessons for enterprise AI adoption, each illustrated with a customer story. Also describes OpenAI’s org split into Research, Applied and Deployment teams and its “iterative deployment” approach.

Key claims

  • Lessons: start with evals; embed AI in products; start now and invest early; customise and fine-tune; put AI in the hands of experts; unblock developers; set bold automation goals.
  • Morgan Stanley ran evals for every proposed application (translation, summarisation, comparison against expert advisors) before production; 98% of advisors now use it daily and document access rose from 20% to 80%.
  • Indeed’s GPT-powered “why this job” explanations raised applications started by 20% and downstream success (employers more likely to hire) by 13% when tested against the previous matching engine; a fine-tuned smaller model then matched results with 60% fewer tokens.
  • Klarna’s assistant handled two-thirds of service chats within months, cutting resolution time from 11 to 2 minutes, with a projected $40M profit improvement and satisfaction scores on par with human support; results came from continuous testing and refinement.
  • Lowe’s fine-tuning improved product-tagging accuracy by 20% and error detection by 60%.
  • BBVA rolled out ChatGPT Enterprise with Legal, Compliance and IT Security involved, and employees built 2,900+ custom GPTs in five months.
  • Mercado Libre built an internal platform (Verdi) with security, guardrails and routing built in for 17,000 developers.
  • Successful companies align on high-return, low-effort use cases, iterate, and back the work with rigorous evals and guardrails.

Feeds

AI use-case discovery · POC to production

Suggest changes →