Instacart Transitions 97% of Engineering Work to AI

At Instacart, a major grocery delivery company, the engineering department is seeing engineers rarely read code directly themselves. According to Anilban Kundu, the company's CTO, speaking at the event "VB Transform 2026," in 97% of cases among current engineers, AI handles code generation and management without humans reading the code directly. As AI agents take on repetitive, high-volume standardized tasks, humans can focus on judgment, intent interpretation, and exception-handling issues that require deliberate attention.

This shift did not happen overnight. In software development, engineers have traditionally spent time on repetitive work such as creating standardized code and maintaining existing code. As generative AI has made code generation automation practical, leading companies have begun revisiting what "humans should do." Instacart is one example, but what sets it apart is that it has restructured its teams down to the actual workflow level, positioning it as a concrete case study within the industry.

In practice, particularly for new projects, the operation has shifted to code being generated and regenerated on a weekly basis. As a result, "technical debt"—the problem of old code and unused processes accumulating—has become almost a non-issue, according to Kundu. Unused code naturally falls away and is rebuilt when needed, resembling a development style closer to when assembly and object code were handled in the past.

Meanwhile, the remaining 3% involves areas where careful human judgment is essential: legacy systems (systems running on older technology foundations), compliance requirements, and delay-sensitive processes. This also includes parts where large amounts of unused code or partially implemented code remain. For these areas, Instacart is methodically addressing them through a project called "Atoms," breaking down monolithic systems into smaller components and rebuilding them in simpler, more manageable forms.

As AI-generated code increases, the meaning of traditional "code review" is also changing. Since AI guarantees syntactic correctness, what humans need to verify is not the appearance of code but the "intent," Kundu explains. Based on this thinking, the company is transitioning to an "intent model" where developers are trained to pose appropriately targeted questions to multiple AI models aligned with their objectives. On the evaluation side, approximately 7,000 automated evaluations run monthly, and a system is in operation responding with approximately 99.9% accuracy to over 8,000 real-time inquiries from developers.

AI utilization is also advancing in reliability management. Instacart has built an AI system using records of actual failures and root cause analysis from its own services as training data. Rather than teaching the system general failure patterns, by providing the company's unique history of failures and countermeasures, the accuracy of problem detection and response in production environments has improved from around 60% to over 90%. An internal tool called "Blueberry" monitors more than 200 Slack channels and detects patterns that humans might easily overlook.

What Instacart's case demonstrates is that AI adoption is not merely adding an efficiency tool but is fundamentally changing the role of engineers themselves. One could argue that rather than the ability to write code, engineers will increasingly need the ability to design what AI should do and to communicate that intent appropriately. With concrete results now emerging—such as the elimination of technical debt and improved defect detection accuracy—the coming focus will be on how far this shift in structure spreads across the software development field and how other companies follow suit.