Designing a molecule and predicting what it will do are different tasks, and the second is the more tractable one. Given a structure, a model estimates a number: how strongly it will bind the target, how quickly the body will clear it, whether it is likely to interfere with the heart's electrical rhythm. This is a regression problem, and it is where AI has been most reliable, because the training data are measurements — thousands of compounds whose activity against a target was actually measured in an assay.
A model trained this way learns a mapping from structure to number. Show it a new structure and it returns an estimate with an uncertainty range. The estimate is useful precisely because it is cheap: it costs seconds instead of the weeks and grams of material a physical assay requires, so it can be applied to the entire shortlist rather than to a handful.
What the prediction is not is a measurement. The model has never seen the molecule, and it has never seen the target in a living tumor. Its accuracy depends on how close the new molecule is to the training set; for a genuinely new chemical class the estimate can be badly wrong, and the model usually cannot tell you when it is out of its depth. Predicted potency also ignores everything that happens after binding — whether the molecule reaches the tumor, whether the cell responds, whether the body tolerates it.
This is why prediction and simulation are used together, and why both are followed by experiment. Physics-based simulation computes how a candidate actually fits the target pocket, atom by atom, using the forces between them. It is slower and needs a plausible pocket structure, but it does not depend on having seen similar molecules before. AI models can then be trained on large sets of such simulations to make fast approximate versions of the same calculation. The two answer slightly different questions, and a candidate that survives both is a better bet for the bench.