I lead AI work inside a globally distributed team. The work I spend most time on is getting systems people will actually use: keeping judgment in the loop, building agent workflows other teams can reuse, and the trust conditions that decide whether anyone will rely on the output.
Synthetic data and evaluation harnesses still show up — usually when the organisation cannot safely test against production records, or when a skeptical reviewer needs something firmer than a demo. They are part of the toolkit, not the headline.
Most of what I've learned doesn't generalize as a technique. It generalizes as a posture toward where the failure modes hide and who will own the output once it ships.