What makes a useful AI coach inside a task

I keep seeing teams start with the model and end with the interface. In workflow coaching, that sequence should be reversed.

The first question is not “How smart is the assistant?” It is “What is the user trying to do right now, and what would make that easier without slowing them down?”

In practice, three behaviours matter most:

• Timing: suggestions appear at decision points, not continuously.
• Compression: guidance is concise, with depth available on demand.
• Escalation: uncertainty is visible, and handoff paths are obvious.

Confidence handling is especially important. Overconfident wrong suggestions are more dangerous than missing suggestions. A useful coach shows what it knows, what it might know, and what it does not know.

The systems that feel trustworthy are usually the ones that make their limits legible.