What is machine learning?

A model is fitted, not programmed. You supply examples with known answers, and an optimisation process adjusts the model's parameters until its predictions on those examples are close enough. The useful property is generalisation: performance on data it has never seen. That is why any honest accuracy figure comes from a held-out set the model was never trained on.

The cost is almost never the modelling. It is getting labelled data that reflects the situation you actually care about, and it is the failure mode too — a model trained on data that doesn't match production quietly degrades instead of erroring. If nobody has the labelled examples yet, the first phase of an ML project is data collection, and it is worth pricing it that way rather than discovering it later.