Your AI agent is 90% right at every step. So why does it finish only 2 out of 3 tasks?

Your AI agent can be 90% right at every step and still finish only 2 out of 3 tasks.

We saw this in our own small AI engine: 90.7% right screen by screen, but only 8 of 12 tasks finished.

Then we asked two separate reviewers to argue against us. They made good points: mistakes are linked, step tests show where things break, and checking everything costs time.

So we changed the question. Not "steps or tasks?" but "where can a mistake still be undone?"

Every job has a few points of no return: submit, pay, delete, send. Check there. Test steps on real screens before that. And give the agent a way back.

Where are the points of no return in your work?

The slides