OCR predates modern AI including deep learning and LLMs. It worked based on heuristics that were hard-coded. These techniques are faster and require less CPU power but are more error-prone while AI-assisted combined heuristics for written and typed content recognition can deal with more variability. Even still, the USPS employs people (or at least it did) to correct addresses manually that its automated pre-modern AI address scanning software cannot determine with enough confidence. Perhaps modern AI (deep learning) could replace most of those jobs but perhaps not all as some could be read by humans that algorithms might not be able to decide with enough confidence and other labels are truly undecipherable even with all the AI techniques in the world and need to be returned to sender (RTS).
