Ethan Mollick, a Wharton professor who studies how work changes around new tools, set this down as the first rule of his book Co-Intelligence. It sounds like a cheerleader's line. It is a lab protocol.

The reason is the shape of the thing. Mollick calls it the jagged frontier: AI does some tasks better than you and fails at others that look just as easy, and nobody can tell you in advance which is which.

No vendor, colleague or article you skimmed last month can draw that line for you.

So the map of where the tool helps you has to be drawn by you, one task at a time. Most people draw it once, early, from three bad attempts, and never redraw it. The model improved twice since then. Their map did not.

Look at where you still refuse to try it.

The client email you rewrite four times. The pricing sheet you rebuild every quarter. The contract clause you read at eleven at night because nobody else will.

Those are the tasks you decided AI "can't do" before you tested it. That is a guess, filed as a fact.

So today: pick one task you would never hand to a machine. Give the tool ten minutes on it before you start. Then write one line on where it failed, and keep that line.

You cannot find the edge of the frontier by standing back from it.