Quick answer: We set out to build our internal platform AI-first – the AI at the center, everything arranged around it. Then I asked the team to track deals, and they lovingly rebuilt a plain old CRM: separate database, separate screen, the exact thing we were trying to escape. Clean code. Properly reviewed. Completely wrong. Turns out this is the norm, not my team’s bad day – MIT’s 2025 study found 95% of enterprise GenAI pilots deliver no measurable return, and the culprit is organizational habit, not the model. AI-first is a fight with your own instincts.
Nobody did anything wrong, and that’s the trap
Engineers shipped clean code. A reviewer approved it. Together they built 2009. A working version of the wrong idea is almost impossible to throw away – it’s too tidy to kill.
Twenty years of muscle memory doesn’t care about your vision deck.
The AI-first move was obvious: point an agent at the data we already had. Instinct did the opposite before anyone noticed.
It took us most of three years to stop doing this. Not the AI. Us. Specifically, me.
Proof it works
FAQ
Adapted from a LinkedIn post by Andrew Babkin, Managing Director & CTO at Azati – read the original. Stat: MIT NANDA, State of AI in Business 2025.
