Technically it should be possible (not easy but doable), but from my experience ICR always* requires human verification anyway, so a pure digitizing solution might be better anyway (most of them have ICR interpretation before any human sees them anyway).
Robots are great for accompanying those processes, but personally I view handwritten the same as unstructured input - a struggle between accuracy and %coverage.
While you can get very good results from ICR alone, it’s still risky (“JOHN SMITH” -> “JOHN SHITH”… that customer probably wouldn’t be too happy, and it’s a quite common interpretation error if the M is written with one smooth hand movement).
Biggest risk is always that something is read wrong, but falls within allowed values - there’s no way for a program to know it’s wrong if the value is plausible.
I’d love for someone to pull it off, but let’s remember that even scanned books aren’t recognized 100% (see reCaptcha).
*unless it’s dictionary based data and/or low risk/impact stuff or it’s really good capital handwriting with high DPI scan on an OCR-optimized form… a lot of if’s there unfortunately, while it still can be wrong.
Disclaimer: I may be biased due to previous digitzing experience.