What actually changes when you add AI to a business process.

Most teams I talk to want AI to do one of two things: take everything off their plate, or generate slightly better marketing copy. Neither is where the real change happens.

The middle ground nobody talks about

The interesting AI work sits in the middle — quietly rewriting how a process behaves without touching what it produces. Same output, different shape underneath.

An accounts team I worked with had a six-step invoice triage process. After AI: still six steps, but four of them were now review-and-approve instead of read-and-decide. Same people. Same software. Half the time.

Where it tends to land

The processes that benefit most have three things in common: high volume, low ambiguity per item, and a clear definition of "done." Triage. Classification. First-draft generation. Cross-referencing.

The ones that resist it: anything where the value is in the conversation itself, anything where the output has to be defended in front of a regulator, and anything where the cost of being subtly wrong is high.

What it doesn't change

The org chart. The accountability. The meetings about the meetings. AI doesn't fix process design — it accelerates whatever process you already have. If the process is broken, you now have a broken process running faster.

> The best automation projects start with a week of not building anything. Just watching.

That week is where the value actually gets defined. Everything after is execution.

What to do next

If you're considering an AI project, the first question isn't "what model should we use." It's "what does this process look like today, honestly, including the parts nobody documents." Get that right and the rest is just plumbing.