Quick answer: Everyone’s building AI in-house, or about to, so here’s what year three actually feels like, including the figures I’m not proud of. People kept asking what our AI-first platform cost, and for a while the honest internal answer was "nobody’s really sure." So: six figures over three years (hundreds of thousands, not the millions everyone assumes). 24 of those 30 months pointed the wrong way; only the last six clicked. That’s an ~80% wrong-direction rate before it worked, which, funnily enough, is almost exactly the industry’s AI failure rate. And we more than doubled while building it, 110 developers to 250.
One view: don’t build, buy
MIT found 95% of enterprise GenAI pilots return nothing, and external tools often beat internal builds. By that math, our 24 wrong-direction months are Exhibit A for buying.
The other view: the walls were the point
Most of that money didn’t buy software – it bought a head start. Doing it a second time is far cheaper because we already walked into every wall. That’s showing up in client work now.
The part that still stings
The one genuinely AI-first thing we built, call it CV-Tinder, scores any candidate against any role 0–100 with a rationale, and it worked. And it still lost to a 30-second message to a team lead, because that’s the habit. We also shipped it too raw, too early, partly because leadership (me) wanted to show momentum. Not all the walls were the market’s fault. Some were mine, which is why I asked one of our engineers to publish his unedited version, warts and my name included.
What it taught us
We run this discipline as a service now: Managed AI and an honest build-vs-buy assessment, because the head start is only worth it if someone’s already survived the walls.
FAQ
Adapted from a LinkedIn post by Andrew Babkin, Managing Director & CTO at Azati – read the original. Industry stat: MIT NANDA, State of AI in Business 2025.
