"AI is the new electricity. Just as electricity transformed almost everything 100 years ago, today I have a hard time thinking of an industry that I don't think AI will transform in the next several years."
Two years ago I wrote a piece on this site about how to start with AI. I’d just left TMB after the Jukin Media acquisition. I hadn’t yet built anything I’d consider production. The post was useful for what it was, a frame for getting unstuck. I’d write almost none of it the same way today.
The version of the question that mattered in 2024 was: how do we begin. Identify a project. Pair business experts with technologists. Demonstrate traction in 6 to 12 months. Andrew Ng’s framework was sound advice for a world where most companies were still trying to figure out whether AI was real.
That world is gone.
Today every company has started. Most are in some flavor of experimentation. The interesting question is no longer how to begin. It’s how to do it well, and what separates the companies actually shipping production AI from the ones still running pilots that never go anywhere.
I’ve spent the last year answering that question by building.
At LOST iN, the travel media and AI-concierge company I co-founded, I’ve personally built the AI infrastructure end-to-end. A production RAG concierge running on Pinecone and Supabase pgvector, multi-model routing across Claude and GPT-4o, document ingestion through LandingAI ADE. Eval scores iterated from a 48 percent code-deflect rate down to 3 percent. I write production code daily in Claude Code and Cursor. I ship features, iterate evals, and operate at the same engineering velocity as the AI-native teams now defining how technology companies build.
That experience changed what I think the right framework looks like.
Here’s what I’d tell a CEO today.
Stop running pilots. Build the system.
The pilot framing made sense when no one knew if AI worked. Now we know it works. The companies winning are the ones that stopped piloting and started building infrastructure. Not “an AI feature” bolted on, but the data ingestion layer, the retrieval layer, the eval harness, the model routing, the agent orchestration. Pilots without infrastructure are theater. Infrastructure without pilots is product.
Build evals before features.
The single biggest mistake I see is companies shipping AI features without an eval system underneath them. Then nothing improves systematically. Quality becomes vibes-based. Engineers tweak prompts based on the loudest customer complaint and break two other things in the process.
When I rebuilt the LOST iN concierge, the eval harness came first. I scored every response, tracked deflect rate, tracked accuracy, tracked tone. The system told me what to fix and gave me a number that I could move. That’s how you go from 48 percent to 3 percent. Not through brilliance. Through measurement.
Hire operators who can code. Or coders who can operate.
The old framework said pair business experts with technologists. Still true in theory. In practice the loop is too slow. The person closest to the customer needs to be able to build, or the person building needs to understand the customer deeply enough that the round trips disappear.
I write production code now. Not because I’m a great engineer. Because the bottleneck in shipping AI is the gap between someone who understands the problem and someone who can build the solution. When those are the same person, things move at a different speed.
Stop calling it transformation.
“AI transformation” is consulting language. It implies an end state, a moment where you’ve successfully transformed. There isn’t one. The companies that will win are the ones that internalize that AI is now infrastructure, like the cloud became infrastructure 15 years ago. Nobody talks about “cloud transformation” anymore. Cloud is just how you build software. AI is becoming the same thing. The right verb is “build with,” not “transform to.”
The hardest part isn’t technical.
A working RAG concierge with strong evals is now a couple of weeks of focused work for someone who knows what they’re doing. The hard part is everything around it. The data hygiene before the retrieval layer is good enough. The product decisions about when to defer to AI and when to escalate. The org redesign when AI starts doing work humans used to do. The question of what your company even is when 60 percent of the operational load has shifted.
This is where most companies are stuck. They think they’re solving an AI problem. They’re solving an organizational problem in AI clothing.
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Two years ago I wrote that the first AI project should be one with a high chance of success, more important than picking the highest value project. That’s still right. But it’s also incomplete. The bigger question, the one I couldn’t have written then, is what you do after the first project works. Most companies don’t have a good answer.
The companies that figure it out aren’t going to transform. They’re going to operate differently.