A team builds a scheduling tool using AI. It works. Two weeks later they’ve got a reporting dashboard too, built the same way. A month after that, a third tool handles customer intake.
Each one solved its own problem well.
Three months in, none of them share a data model. Someone’s now spending six hours a week manually reconciling numbers between two of them. The tools that were supposed to save time have become the thing eating it.
The healthier version of that story starts with one extra question, asked early: same three tools, same team, same AI. This time, someone asked what this needs to talk to before building the first one. Tool two reused the data model from tool one. Tool three plugged into both. Same starting point. Completely different trajectory.
Nothing about the AI changed between those two stories. What changed is whether anyone was thinking about the system before they hit generate.
Access Isn't the Bottleneck. Judgment Is.
For years, a lot of ideas died at “we don’t have the engineering time.” That’s less true than it used to be. More people can now build something that works without waiting on a dev team.
But a tool that works in isolation and a tool that works inside a business aren’t the same thing, and AI doesn’t know the difference. It’ll build you exactly what you described, disconnected from everything else you already have, because nobody asked it to do otherwise.
AI didn't remove the need for systems thinking. It removed the excuse for skipping it.
That’s systems intelligence: the discipline of knowing how a new piece fits with everything else before you build it, not after you’re stuck maintaining five things that don’t talk to each other.
Two Ways This Breaks
We see this go wrong in two directions, and they’re different enough to be worth separating.
The first is horizontal. Tools that don’t share data with each other. This is the sprawl problem: everything works, nothing connects, and someone becomes the human API stitching it all together by hand.
The second is vertical. A single tool that works fine today and can’t survive its own next version. It’s the dead end problem: someone prompts a working v1, and adding the obvious next feature six months later means half-rebuilding it, because nobody thought about where it needed to go.
Both come from the same root cause. Both are avoidable with the same habit: think about the system, not just the task in front of you.