The machine smells correctly, when the same numbers (or similar when using some norm, e.g. the L2) appear for the same smell (reproducibility) and therefore a mapping (numbers -> smell) can be created. When this starts to exist (practically usable), there can be a database to store the mappings, allowing classification. E.g., the machine says "this tastes like banana". The machines/algorithms/products could itself be rated for precision.

I dont say such machines don't exist, but for my taste (pun intended) the solutions all lack something, either long term stability or having a second source supplier or being able to classify a reasonable amount of tastes or being able to distinguish between two tastes (or lacking all those things together).