Sickness insurance guarantees employees the right to take leave from work when they are sick,but is vulnerable to excessive use because monitoring of recipients’ health is difficult and costly.In terms of costs, it would be preferable to focus monitoring on individuals whose sickness absenceit strongly affects. This paper studies targeted monitoring in the setting of a large-scale randomisedexperiment where medical certificate requirements were relaxed for some workers. I employ amachine learning method, the generalised random forest, to identify heterogeneous effects on theduration of workers’ sickness absence spells. This allows me to compute treatment effect estimatesbased on an extensive set of worker characteristics and their potentially complex relationships witheach other and with sickness absence duration. The individuals who are most sensitive tomonitoring are characterised by a history of extensive sick leave uptake, low socioeconomic status,and male gender. The results suggest that a targeted policy can achieve the same reduction inmonitoring costs as took place during the experiment at a 51 percent smaller loss in terms ofincreased sickness absence. Monitoring all insured individuals is estimated to be inefficient, butthe benefits of targeted monitoring are estimated to exceed the costs.