Build for next month's model, not today's.

Some work is not reliable for AI today, but it is clearly moving into the model’s hands as context, reasoning, and tool use improve. It is gated by capability, not by principle. The mistake is to build a permanent human process around work the curve is about to take.

I have built in a domain where the answers are scattered, conditional, and not yet reliable for a model end to end, and the lesson was about trajectory: a waiting room, not a permanent human domain.

Do not staff forever against work the curve will take. Position for the handover, and watch for the moment it crosses the line.