Picking the first thing to do. The instinct is to choose the most impressive problem; the better choice is one with high frequency, tolerable error cost, available data, a clear measure of success and a willing team. Impressive but unmeasurable is how projects end up quietly discontinued.
Technical view
Assess candidates on value, defined as frequency multiplied by time or cost per instance, feasibility given data and integration requirements, error tolerance and reversibility, measurability of the outcome, and organisational readiness. Prefer augmentation over full automation initially, since the human check bounds the downside. Maintain a register with expected and realised benefit, because the discipline of recording realised benefit is what prevents a portfolio of unfalsifiable pilots.
Related
Part of AI Terminology Index, section: Product and Adoption.