Inside Duolingo's AI-Powered Code Review: How Education Unlocked Autonomous Systems

At Duolingo, a dedicated team called DevEx AI has taken a different approach to developer experience work: instead of focusing on traditional tools like CI/CD pipelines, they concentrate on teaching engineers to use artificial intelligence effectively in their daily workflows. That cultural and educational groundwork has proven transformative, enabling the company to deploy autonomous systems—like their AI-driven code review bot—that would have been impossible to trust without first building literacy and buy-in across the organization.

The Barriers to AI Adoption#

Integrating AI into core engineering processes faces real human resistance. Engineers often view foundational practices like code review as "sacred"—best practices refined over years that now risk disruption as AI accelerates code generation. Beyond skepticism, AI adoption creates murky questions around accountability: when an algorithm contributes to a decision, who bears responsibility if something goes wrong? Trust, meanwhile, builds slowly and erodes quickly, making many teams hesitant to hand over authority to automated systems.

Most organizations already have access to AI tooling (Claude, Cursor, copilots). Few, however, have dedicated teams teaching their staff how to use these tools well—or building the institutional confidence needed to deploy AI as an autonomous agent rather than just a productivity add-on.

Building Organizational AI Literacy#

Duolingo's approach centers on four practical levers:

Hands-on workshops. Lab-style, self-paced training on specific AI capabilities—from Model Context Protocol (MCP) servers and IDE rules to batch requests and evaluations—delivered by the engineering organization itself. Ninety-five percent of participants reported learning something new and would recommend the format. Engineers trust pre-vetted internal content far more than generic vendor training.

AI observability dashboards. By tracking daily active users across AI tools, tool adoption by department and team, consumption patterns (languages, IDEs, tokens, costs, model families), the team gained visibility into what the organization was actually doing with AI—and could identify gaps in adoption and training needs. The dashboard itself was built iteratively ("vibe coded"), making it easy to add new charts on demand.

Office hours for support. Fifteen-minute slots where engineers and non-engineering staff (learning designers, QA teams) can get one-on-one help using AI in their workflows. The spillover benefit: these adjacent roles bring AI practices back to their teams, amplifying impact beyond pure engineering.