As @Fildor mentioned in the comment when he did OCR (Invoices) we used to have a multilayered process. If OCR confidence was above a certtain threshold, it went through, directly (very high confidence). If it was less, then it would be validated against several measures depending on the content. For example City names against a Database of all the city names in existence in the respective country. Then we would have a list of say the top 5 most probable hits. If the top hit was still below a certain resemblance indicator, then it would be run by a human to validate / correct it. And the result would be fed back into the AI part of the recognition as additional training data. That way the process is not 100% automatic but we were able to go from 100% human data entry to about 1% Human Data validation and 0.1% Human data correction. With improving numbers in the warmup phase and keeping the training set optimized prevented deteriorating AI performance.

Also mentioned by @Tlaquetzal in the stack link At the moment, it is not possible to do this. I found a Feature Request made to Cloud Vision API to take a PDF file and export it as a searchable PDF which might resolve this issue. I recommend you to subscribe to the Feature Request (click on the star next to the title) so it can get more visibility.

Posting the answer as community wiki for the benefit of the community that might encounter this use case in the future.

Feel free to edit this answer for additional information.