AI Writes a Lot of Our Code. So We Hired It a Babysitter.

Maksym Kramarenko

Digital Transformation Advisor, Azati

Quick answer: The loudest brag this quarter: "AI writes most of our code now." Ours does a lot too. So we added more review, not less. Here’s why: when a model writes it, the dangerous failure is plausible-but-wrong – it compiles, it looks right, it sails through a glance, and it’s broken in the exact way a tired human ships at 6pm on a Friday. The data isn’t reassuring either: Veracode found 45% of AI-generated code introduces a security flaw – roughly a coin flip you really don’t want to lose. Our rule is one line: if AI wrote it, a senior reads it. That’s AI-assisted development with the supervision turned up.

Plausible-but-wrong is the new bug

Old bugs crashed loudly. This one smiles at you, passes the glance, and detonates in production three weeks later.

Spend the speed you saved on reading

Generation got faster, so we pour the savings straight into senior review – architecture, security, the edge cases. That’s not slowing down; it’s the only "faster" that’s actually true.

AI wrote it. A senior reviewed it. We shipped. We won.

In that order, every time. AI testing and QA do the rest.

Where it paid off

Read the case study

Modernizing a 30M-member insurance accounting system, senior-owned delivery took downtime from 12 days to zero – the kind of number you only hit when a human owns the scary parts.

FAQ

Does AI-generated code need review?

Yes – more than human code. About 45% of it ships with a security flaw (Veracode, 2025), and those flaws pass a casual look.

How does Azati handle AI-assisted development?

Senior oversight on architecture, security, and testing – see Vibe Coding Engineering and our CTO’s The Engineer Is Expanding.

Shipping AI-written code you actually trust?

Let’s talk

Adapted from a LinkedIn post by Maksym Kramarenko, Digital Transformation Advisor at Azati – read the original. Stat: Veracode 2025 GenAI Code Security Report.