Which regression check would you automate with an 18-hour implementation budget?

Two stable checks have equal business risk and equal defect-detection effectiveness whether run manually or automatically. Each runs 5 times a week for 12 weeks: 60 runs. Candidate A takes 20 minutes manually; automation costs 8 hours to build, 0.5 hour of maintenance per week, and 2 minutes of human review per run. Candidate B takes 45 minutes manually; automation costs 18 hours to build, 1 hour of maintenance per week, and 3 minutes of review per run. Implementation capacity is 18 hours, so both cannot be built. Assume the automated checks remain useful for the whole horizon, all stated costs are incremental, and ignore machine cost.