Ask most organisations whether their historical records are ready for AI and the answer often comes back as a status update on a scanning project. We digitised the client files in 2018. The claims archive went to a service provider two years ago. It’s all in the document management system now.
That’s a real achievement, and it is genuinely a prerequisite. But it isn’t the same thing as being usable by AI. Between a scanned page and information an AI system can reason over, there are several distinct layers, and most digitisation programmes were only ever designed to deliver the first two.
The layers between “scanned” and “usable”
An image. A scan is a picture of a page. It is a faithful record and, for retention purposes, often all that was required. To a machine, it’s pixels.
Text. Optical character recognition converts those pixels into characters. This is where “searchable PDF” comes from, and where most digitisation projects stop.
Structure. A document is not a flat stream of words. It has columns, headings, tables, signature blocks, appendices and a reading order. Strip that away and the words survive while their meaning is damaged.
Identity. What is this document? Whose is it? What date does it carry, what type is it, and which matter, policy or client does it belong to? This is classification and metadata, and it is what lets an AI system find the right document rather than a plausible one.
Context and permissions. How does this document relate to the others around it, is it current, and who is allowed to see it? I’ve written about both of those elsewhere.
An estate can be fully digitised and still stop at layer two. That is the gap this article is about.
OCR is a conversion, not an understanding
It’s worth being precise about what OCR does and doesn’t do, because “we’ve OCR’d it” is doing a lot of load-bearing work in many AI readiness conversations.
Modern OCR is very good on clean, printed, well-scanned text. It is much less reliable on the material that fills historical archives: faxes, carbon copies, microfiche, degraded originals, forms with handwritten entries, stamps, marginal notes and annotations added by hand over the years.
Human readers barely notice moderate OCR errors, because we read the image and the eye repairs the text as it goes. An AI system connected to a document repository usually does not read the image. It reads the extracted text layer. Whatever was lost in extraction is simply absent as far as the AI is concerned.
The research on this is consistent. A study across six natural language processing tasks found a consistent impact from OCR errors, and identified a quality threshold of roughly 70 to 80%, below which performance suffered dramatically, with the authors recommending texts ideally above 90% accuracy[1]. Earlier retrieval work found that an error rate of 5% reduced retrieval effectiveness by around 17%, and an error rate of 20% by around 46%[2].
Note what those numbers describe: not an AI that fails obviously, but one that quietly retrieves the wrong thing, or fails to retrieve the right thing, and answers anyway.
Structure is lost more easily than text
The structural layer is the one most often overlooked, and it matters more than people expect.
Even native digital PDFs are awkward here. PDFs store text as positioned glyphs without inherent semantic markup, so parsers must infer the logical structure – paragraphs, tables, headings – using geometric heuristics or machine learning, which becomes especially challenging for complex layouts, multi-column text and scanned images requiring OCR[3]. Anyone who has copied a table out of a PDF and watched the columns collapse into a single line has seen this first-hand.
For AI systems, that collapse is consequential. A premium figure separated from its policy year, or a fee separated from the schedule it belonged to, doesn’t look wrong to the system. It looks like a fact.
A benchmark published at ICCV 2025 examined this directly, testing how OCR quality cascades into retrieval-augmented generation across 8,561 document images from seven application domains. The researchers evaluated the current OCR solutions and found that none was competent for constructing high-quality knowledge bases for these systems, identifying two distinct problems: semantic noise, where the characters are wrong, and formatting noise, where the structure is wrong[4].
That’s an academic benchmark rather than a verdict on any particular commercial product, and the field is moving quickly. But it makes the wider point clearly enough. The step between a scanned page and reliable machine-readable content is a real engineering problem, not a box that gets ticked when the scanner stops.
What this means in practice
None of this is an argument against digitisation. It’s an argument for digitising with a purpose in mind.
A scanning specification written to satisfy a retention requirement asks: Is the image legible and is the record complete? A specification written for AI readiness asks harder questions. What image quality does the source material need to yield reliable text? Which documents contain handwriting that current recognition will struggle with? What accuracy is acceptable, and how will it be measured rather than assumed? Are tables and forms being captured as structured data or flattened into prose? What metadata is being applied at capture, and is it enough to identify the document later without a human looking at it?
There’s also a sequencing point worth making. Reprocessing an archive is expensive and disruptive, so it pays to decide what the information is for before it is converted, not afterwards. For most organisations, that means selecting the parts of the estate where high-quality capture is genuinely worth it, rather than applying the same specification to everything.
How Dajon approaches it
We’ve been digitising records for organisations for nearly thirty years, which means we’ve also seen what happens when the specification and the eventual use don’t match. Our interest now is in capture that anticipates the second life of a document: image quality chosen for reliable extraction, classification and metadata applied at the point of capture, accuracy measured rather than assumed, and structured content preserved as structure.
If your historical information has already been scanned, that’s a solid foundation, and much of it may be fine as it is. The useful next step is to find out whether what came out of that process is text an AI system can actually rely on.
Having data is not the same as having AI-ready data. Having scanned it isn’t either.
References
- Assessing the Impact of OCR Quality on Downstream NLP Tasks van Strien et al., ICAART 2020[↩]
- How to Do Lexical Quality Estimation of a Large OCRed Historical Finnish Newspaper Collection With Scarce Resources reporting Savoy and Naji, 2011[↩]
- Empirical Evaluation of PDF Parsing and Chunking for Financial Question Answering With RAG ICSE-SEIP 2026[↩]
- OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation Zhang et al., ICCV 2025[↩]