Why OCR Accuracy Claims Deserve a Second Look
If you are choosing a scanning provider, treat the accuracy figures in their proposals as a starting point, not a conclusion. The recognition technology behind those numbers has changed more in the last few years than in the previous twenty, and not every provider has kept up.
For organisations digitising records, the quality of the text layer decides whether an archive is searchable, discoverable and usable. A figure such as "99% accuracy" sounds reassuring. What matters is what it was measured on, and whether it reflects documents like yours.
Where traditional OCR falls short
Traditional optical character recognition was built for machine-printed text. On a clean, typed page at good resolution, mature engines perform very well. That is where most headline accuracy figures come from.
Real archives are rarely that tidy. Accuracy drops sharply on:
- Handwriting, especially cursive or rushed notes
- Margin annotations, stamps and signatures over text
- Faded, stained or low-contrast originals
- Carbon copies, faxes and photocopies of photocopies
- Mixed layouts such as forms, tables and handwritten entries on printed pages
On handwriting in particular, classic engines often produce little usable text. Some handle neat block capitals in form fields, but flowing handwriting was never what they were designed for. For a file containing handwritten file notes, correspondence or annotated documents, that gap can mean a significant share of the content is effectively unsearchable.
What has changed
Advances in computer vision have drastically increased the accuracy now within reach. Modern AI vision models do not just match character shapes. They read a page much as a person does, using the surrounding words, layout and context to interpret what is written.
The practical effect is greatest on exactly the material traditional OCR struggles with. Handwritten notes, annotated pages and poor-quality originals that once produced unusable text can now be transcribed with accuracy that would have been out of reach only a few years ago. These models can also describe structure, such as identifying a letter, an invoice or a form, which helps with indexing and classification.
This technology is widely available. A provider still relying solely on older OCR engines may be quoting results that set a lower bar than you need to accept.
A different kind of error
Greater accuracy does not mean zero errors, and the errors look different. When traditional OCR fails, it usually fails visibly: garbled characters that anyone reviewing the output will spot.
An AI model is more likely to fail plausibly. If it cannot make out a word, it may substitute one that fits the sentence but was not what was written. In legal, regulatory or evidential work, a confident wrong word can matter more than an obvious garble.
A good provider designs for this. Their process should flag low-confidence or illegible text for human review rather than quietly filling the gap, and should keep the original image alongside the text so any reading can be checked.
Questions to ask a scanning provider
- What recognition technology do you use, and when was it last updated? Ask whether it is a traditional OCR engine, an AI vision model, or a combination.
- What was your accuracy figure measured on? Clean typed pages and mixed real-world files give very different results.
- How do you handle handwriting? Ask for examples on material similar to yours.
- How are uncertain readings flagged? Look for confidence marking and a clear human review step.
- Where is the processing done? For sensitive records, confirm where data is processed and stored, and what happens to it afterwards.
Run your own test
The most reliable evidence is a small test batch of your own documents. Include the difficult ones: handwritten notes, annotated pages, faded copies. Transcribe a few pages manually as a reference, then compare each provider's output against it. Differences that are invisible in a proposal become obvious within a few pages.
Raise the bar
Digitisation is usually done once, and the archive it produces is used for decades. The accuracy achievable today is substantially higher than many established benchmarks suggest, particularly for handwritten and difficult material. When comparing providers, ask hard questions, test with your own documents, and expect results that reflect current technology rather than yesterday's.