If you ask me what made AI implementation work at the places I've led it for, I'll tell you it was the process. The decision rules, the SOPs, the version control for prompts, the eval protocols. The boring operational layer that nobody puts on a slide.

I think most AI failures look the same up close. A team picks a tool. The pilot looks impressive. Leadership signs off. A quarter goes by, then another. The tool is technically still in the stack, but most of the team has stopped using it. They've gone back to doing things the way they did them before the AI initiative, sometimes with the AI cost still on the budget. Ask anyone on that team why, and you'll hear answers like "it's faster to just do it myself" or "I never know if I can trust the output for this kind of customer." Both of those are process problems. The model works. The team is missing a shared rule for when the model is the right call.

This is what process work actually looks like in practice. It's a document, often a few pages, that sits in a shared drive nobody opens for fun. It says things like "for outbound sales emails to existing customers, AI drafts go to manager review before send. For internal newsletters, AI drafts go straight to send. For anything in legal or compliance categories, AI stays out of the workflow." It's mundane. Writing it takes a week. Nobody wants to be the one writing it. Once it exists, every member of the team makes the same decision the same way, and a manager can stay out of the loop for individual calls. That's where the efficiency gain comes from. At Trane I'd put 40 to 50 percent of the gain on documents like this. The model was a smaller piece of the story than most people assume.

There's also a cultural reason this layer gets skipped. The language is unsexy. Governance, decision criteria, eval protocols, prompt registries, content policies. It sounds like compliance. It sounds like the part of the company that slows things down. So when AI strategy gets discussed, the conversation tends to drift toward tools and capabilities, because those are the parts you can demo. Process is harder to demo. It's just a doc.

If I were starting an AI program tomorrow, I'd budget more time for process design than felt reasonable. I'd write the SOP first, before I picked a tool. I'd be willing to look slow for a quarter to be fast for the next two years. Most teams skip this work, which is why most AI initiatives stall out.

For more info or opportunities, let's talk: leethepolymath@gmail.com