The number of possible drug-like molecules is enormous — far beyond any laboratory's capacity to make and test one at a time. So the search is framed as screening: start with a very large collection of candidates, apply a test that estimates how well each one would act on the target, and keep only the few that pass.
Screening in the laboratory is real but slow and expensive. Each candidate has to be made or obtained, then tested in an assay — a controlled measurement of whether the molecule binds the target and how strongly. Testing millions of molecules this way is not practical.
AI changes the order of operations. Instead of testing every candidate, a model trained on known molecules and their measured activity predicts which candidates are likely to bind the target well. The model's output is a ranking, not a measurement, so the top candidates are still made and tested — but the laboratory now spends its effort on a shortlist rather than on the whole collection. The value is concentration: the same laboratory budget covers far more of the search space, because most of the obviously unpromising candidates are filtered out before anyone picks up a pipette.