Velocity Up, Expertise Down: One Engineer's View From the Agent Era

A senior fintech engineer reports AI has roughly doubled delivery speed but is quietly eroding the human expertise and review discipline needed to keep that speed safe.

This brief draws on one senior software engineer working in fintech infrastructure, leading a team of about 20. The operator credits AI coding agents with a 60 to 70 percent velocity gain, turning multi-month features into one or two week deliveries, though testing remains the limiting factor. The picture is double-edged. Leadership now tracks AI credit consumption as a performance signal, pressuring engineers to use tools rather than to verify outcomes. Onboarding has thinned, with managers deflecting questions by saying 'ask the AI,' and engineers shipping agent-written code they cannot debug or explain. Pull requests have ballooned to dozens of files and thousands of lines, breaking the review gate. The operator gives a striking example: an AI-generated handoff document silently dropped open pull requests because of a human label, an error caught only through patient cross-questioning. On governance, this person rates their maturity a four out of five, describing a real control stack: a mandatory orchestrator layer with human feedback, LLM-as-judge, and custom evaluation functions, plus security review of all AI tooling before development. The throughline is a warning. AI is genuinely valuable and used by essentially everyone on the team, but blind reliance, under speed and job-security pressure, risks production defects and lost institutional knowledge. The operator's plea is to preserve documentation, mentorship, and the habit of interrogating AI output.