The harness beat the model
Two AI labs hit the same top benchmark score the same day. Neither of them won on the model.
ReadIt worked in the demo
A live demo that works flawlessly is not evidence the system will survive a normal Tuesday with a normal file.
ReadThe approval bottleneck
A team speeds up its approval queue and the wait never shortens, because the bottleneck was never the queue. It was the gate.
ReadClean the data first
The data cleanup project that never finishes is a design problem, not a size problem: clean in the path of one real use, not in advance of it.
ReadThe dashboard nobody opens
A dashboard built to impress in the review meeting rarely survives contact with a Tuesday. Here is what changed when we stopped adding charts and asked what one number would move.
ReadThe tool bought to fix the last two tools
You bought something to close the gap between your last two tools. Now you have three tools and the same gap. The fix is not another integration.
ReadThe pilot that never leaves pilot
The demo impressed everyone eight months ago and it is still a demo. Nobody owns the decision it was supposed to change.
ReadThe automation that saved no time
You automated the boring half, and the hours moved into reviewing badly shaped output. The fix is the handoff, not the robot.
Read