The question comes up in most manufacturing assessments: can ten years of handwritten maintenance registers be used for prediction? The answer is usually yes, with conditions. Here is what the conditions are.
Where reading works
Dates, machine codes, times and quantities in fixed columns. Even poor handwriting in a box is read reliably when the system knows what kind of value belongs there, and a date in a date column that is not a plausible date is caught at once.
Where it does not
Free-text cause descriptions. Different supervisors, different abbreviations, different pens. Models read them inconsistently, and worse, confidently. We stopped trying to extract these as text and instead classified each one into a dozen cause categories, with a person checking a sample. The categories were what the analysis needed anyway.
The step that mattered
Validation against what the plant already knows. Every extracted row is checked against the machine master, the shift calendar and the production log. A row with a machine code that does not exist, or a breakdown on a day the plant was closed, goes to a queue. A small share of rows fails these checks, and that share holds almost every reading error, which means the rest can be trusted without anyone looking at them.
What the archive was worth
Ten years of registers is the failure history a predictive maintenance pilot needs. Without it a plant has to collect two years of data before starting. Digitisation pays for itself in time, before any model is trained.
If you are considering it
Scan at a decent resolution, keep the pages in order, and find the one retired supervisor who can tell you what the abbreviations meant. Then run a sample of two hundred pages before committing to the archive. The sample tells you the accuracy per field, which is the number to make the decision on.