A package shipped to me from overseas crossed 7,500 miles in 2 days. Getting it across my city took 7.

DHL kept flagging issues that weren't real: the address is incorrect (it's not), delivery was attempted (I was home, it was not). Customer service could only promise tomorrow, and tomorrow always came with a new excuse. Then one day I'm on the phone with a rep telling me the van has already left my area, nothing they can do, try again tomorrow, and a yellow DHL van drives right past me on the sidewalk (not an exaggeration, life's timing can be funny sometimes).

DHL's problem isn't even an AI problem. It's two systems that don't talk to each other, solvable with technology that's been around for years. AI is going to change a lot, but with problems like this still unsolved, the AI timeline predictions need some humility.

The discourse is all about the flashy stuff. Bring up a "boring" problem and the response is "that's already been solved", "it should be easy," "just use X, connect it to Y, done." But real systems are way gnarlier than that. I currently work on AI that processes business documents. The model can read a messy scan, but making a product that does it reliably, thousands of times a day, across customers with different document types, schemas, and quality bars? Approaches start failing unpredictably at that scale. It's these "should be easy" problems that stand in the way of a business realizing real ROI with AI.

I've spent six years building enterprise products, and the "boring" problems are always where the real ROI lives. The one I'm in right now looks straightforward from the outside. It's some of the most interesting work I've done though, precisely because it's not.