← gunjan.mehta — back to site

case study 02 · ai-assisted engineering

Rolling out AI-assisted engineering

Gunjan Kumar Mehta · Concirrus · 2026

Every engineering team is asking the same question: what does a great team look like when AI writes a lot of the code? The tempting answer is adopt the tools and let speed happen. The honest answer is slower: adoption without discipline creates security risk, quality drift, and engineers who do not know what they merged.

100%
adoption across squads
~3x
delivery velocity
90%+
retention held
0
security incidents

Policy before tools

Before rollout, I wrote the approved-tools policy and the norms for AI-generated code: review standards, attribution, and what AI-generated work requires before it counts as done. Only sanctioned tools with clear use-case boundaries. No use whatever AI, which creates shadow IT and security risk.

The shift: specify, review, verify

The core change was moving every squad from write to specify, review, verify. AI does the writing; humans own the judgement.

What changed for people

Hiring stopped optimising for typing speed and started optimising for judgement: decompose a problem into a spec, review a diff you did not write, tell the difference between looks right and is right. Junior engineers got the biggest upgrade: the AI acts as a patient tutor, and ramp-up time collapsed, provided the review bar stayed human.

The result

Full adoption across every squad, roughly 3x delivery velocity, and retention above 90% through the reshape. The discipline is the differentiator: the policy and the norms are what make AI-assisted delivery safe enough to scale, and they are the part anyone can copy and most teams skip.

Engineers moved from writing code to specifying, reviewing and verifying it. Velocity roughly tripled, retention stayed above 90%, and no security incident came from AI-generated code.