"A mid-sized fintech company with 150 engineers rolled out AI coding assistants in early 2025. The productivity gains on greenfield projects hit 40%-better than the vendor's optimistic projections. Engineers building new microservices from scratch reported that AI pair programming felt like having a competent junior developer working alongside them, handling boilerplate, suggesting tests, catching edge cases before they became bugs."
"Feed them code where the architectural decisions live in Slack threads from 2019, where the reason a service retries exactly three times is known only to a senior engineer who joined in 2018, where the edge cases that caused the production outage of June 2021 are handled by conditional logic that nobody dares touch-and the AI produces confident hallucinations. Syntactically correct code that violates constraints the AI cannot see."
A mid-sized fintech with 150 engineers rolled out AI coding assistants in early 2025. Productivity on greenfield projects improved about 40%, exceeding vendor projections. Engineers building new microservices reported AI pair programming felt like a competent junior developer, handling boilerplate, suggesting tests, and catching edge cases. The same engineers saw only a 7% productivity improvement on an eight-year-old, 2.1-million-line core payment system. The tools, engineers, and investment were unchanged. The gap arose from institutional knowledge that AI cannot access: architectural decisions in Slack, undocumented constraints, and historical outages that shaped defensive code. Valuable institutional knowledge remains invisible to AI.
Read at Codegood
Unable to calculate read time
Collection
[
|
...
]