Arun Rajkumar, CTO of a UK-based payment platform, reports that 54% of code is now AI-generated, an increase from 28% last year. While these AI tools excel at routine tasks like scaffolding APIs, generating boilerplate, and managing environment configurations, they struggle with critical 20% of logic involving domain-specific edge cases and negative outcomes. Rajkumar notes that AI agents often prioritize feature completion over correctness, frequently failing to account for illegal state transitions or idempotency requirements in financial transactions. Unlike human developers who possess operational experience, AI models lack the judgment to identify scenarios where system failures occur, such as handling duplicate webhooks or complex retry logic.

The reliance on AI for sensitive financial operations has led to a pattern where agents unintentionally create maintainability issues by reinventing logic instead of utilizing existing, trusted utility packages. Furthermore, because agents are optimized to reduce token usage and interaction turns, they often skip complex error handling. Rajkumar argues that viewing AI as a replacement for senior engineers leads to speed at the cost of silent, high-risk defects. Instead, he advocates for an 'AI-enabled' approach where senior staff oversee the judgment-heavy aspects of product development, while AI performs the repetitive coding labor.

To manage these risks, the author’s team has implemented a system where architectural rules and design patterns are made machine-readable, forcing agents to adhere to established boundaries. They have also invested in rigorous automated testing suites that specifically target negative cases to catch omissions before deployment. Senior engineers continue to perform manual reviews on all code impacting financial logic, focusing specifically on judgment rather than syntax. Rajkumar also introduced an open-source framework called Bodhi Orchard, which aims to improve AI-driven development by providing agents with comprehensive context, including existing test plans and architectural constraints, to reduce blind spots in generated code.

Arun Rajkumar, CTO of a UK-based payment platform, reports that 54% of code is now AI-generated, an increase from 28% last year. While these AI tools excel at routine tasks like scaffolding APIs, generating boilerplate, and managing environment configurations, they struggle with critical 20% of logic involving domain-specific edge cases and negative outcomes. Rajkumar notes that AI agents often prioritize feature completion over correctness, frequently failing to account for illegal state transitions or idempotency requirements in financial transactions. Unlike human developers who possess operational experience, AI models lack the judgment to identify scenarios where system failures occur, such as handling duplicate webhooks or complex retry logic.

The reliance on AI for sensitive financial operations has led to a pattern where agents unintentionally create maintainability issues by reinventing logic instead of utilizing existing, trusted utility packages. Furthermore, because agents are optimized to reduce token usage and interaction turns, they often skip complex error handling. Rajkumar argues that viewing AI as a replacement for senior engineers leads to speed at the cost of silent, high-risk defects. Instead, he advocates for an 'AI-enabled' approach where senior staff oversee the judgment-heavy aspects of product development, while AI performs the repetitive coding labor.

To manage these risks, the author’s team has implemented a system where architectural rules and design patterns are made machine-readable, forcing agents to adhere to established boundaries. They have also invested in rigorous automated testing suites that specifically target negative cases to catch omissions before deployment. Senior engineers continue to perform manual reviews on all code impacting financial logic, focusing specifically on judgment rather than syntax. Rajkumar also introduced an open-source framework called Bodhi Orchard, which aims to improve AI-driven development by providing agents with comprehensive context, including existing test plans and architectural constraints, to reduce blind spots in generated code.