What resolution is best for OCR scanning?

300 dpi is the standard for business documents and gives the best balance of recognition accuracy and file size. Use 400 dpi for small print, poor-quality photocopies or degraded originals. Scanning above 600 dpi rarely improves OCR results on ordinary office paper and significantly increases storage requirements.

Should scanned business documents be saved as PDF or PDF/A?

PDF/A for anything you need to keep long term. It is the ISO 19005 archival standard, embeds fonts and removes features that break over time, and is listed by The National Archives among preferred preservation formats. Standard PDF is fine for short-lived working copies, but records held for six years or more should be PDF/A.

How accurate is OCR on business documents?

Clean machine-printed originals typically achieve 99% or better word accuracy. Photocopies and faxes usually fall to 90-96%, and cursive handwriting remains unreliable. Key index fields should be specified at 99.5% or better, achieved through double-key entry or validation against reference data rather than relying on the OCR engine alone.

How many index fields should a scanning project use?

Three to five is usually right: document type from a fixed list, date in YYYY-MM-DD format, an entity reference such as client or employee, a unique document reference, and a retention or destruction date. Capturing a small set accurately is far more useful than capturing a large set inconsistently.

Can I destroy the original paper once documents are OCR scanned?

In most cases yes, provided the digitisation process is documented to a recognised standard such as BS 10008 so the digital copy can be shown to be a trustworthy record. Some categories - original wills, deeds, share certificates and certain signed instruments - should be retained in physical form regardless, so agree a retention policy before any destruction takes place.

Best OCR Setup for Searchable Business Documents

The best OCR setup for searchable business documents is 300 dpi scanning, output to searchable PDF/A with the image layer retained, full-text OCR across every page, and a small set of indexed metadata fields captured per document — typically five or fewer. That combination gives you text search inside the file, structured search across the collection, and a format that will still open in twenty years. Anything less than 300 dpi hurts recognition accuracy; anything much more inflates file sizes without measurably improving results on ordinary business paper.

The scan settings that actually change OCR accuracy

OCR software is only as good as the image it reads. Most accuracy problems in UK digitisation projects trace back to capture settings, not the recognition engine. The settings worth getting right:

Resolution — 300 dpi. This is the working standard for text documents and the level BS 10008 audit trails are usually built around. Small print, carbon copies and 6pt footnotes benefit from 400 dpi; going to 600 dpi roughly quadruples file size against typical office paper for negligible gain.
Colour mode — greyscale or 24-bit colour, not bitonal. Pure black-and-white capture throws away the tonal information OCR engines use to separate faint characters from background. Greyscale at 300 dpi is the safe default; use colour where highlighter, coloured stamps or signature ink carry meaning.
Deskew and despeckle on. A page rotated 3 degrees can drop recognition accuracy several percentage points. Automatic deskew, border removal and speckle filtering are standard on production scanners and cost nothing in throughput.
Blank page and duplicate detection. Prevents thousands of empty reverse sides bloating the archive and confusing search relevance.
No aggressive compression. Heavy JPEG compression creates artefacts around character edges that OCR misreads. Keep compression moderate and let the searchable text layer do the work.

Document preparation matters as much as settings. Staples, folded corners, sticky notes covering text and torn edges all produce unreadable regions no software can recover. A proper document scanning service prepares each file before capture for exactly this reason.

Output format: searchable PDF/A is the right default

A searchable PDF holds two layers — the scanned image everyone sees, and an invisible text layer sitting behind it that search engines read. You get the visual fidelity of the original plus full-text search, in a file anyone can open without special software.

PDF/A is the archival subset of PDF, standardised as ISO 19005. It embeds fonts, forbids external dependencies and rules out features that break over time, which is why The National Archives lists it among its preferred formats for long-term preservation. For records you must retain for six years (VAT and general accounting under the Companies Act), six years after employment ends (HR files), or fifteen years and beyond (pension and some legal records), PDF/A-2b or PDF/A-3b is the sensible target.

When to keep something else alongside it

TIFF masters — worth retaining for documents likely to be produced as evidence, where an uncompressed master image is easier to defend.
Separate text or XML output — useful if you plan to feed the content into a database, reporting tool or AI extraction pipeline later.
Image-only PDF — only for material where OCR would return meaningless text anyway, such as engineering drawings or heavily handwritten notes.

Indexing: the layer that makes OCR genuinely useful

Full-text OCR answers “which document contains this phrase”. Indexed metadata answers “show me every invoice from this supplier in Q3 2024”. You need both, and the second is what most projects underinvest in.

Keep the index field set small and consistent. Three to five fields captured accurately beats twelve captured erratically:

Document type — from a fixed, closed list. Free text here destroys filtering within a month.
Date — always ISO format (YYYY-MM-DD) so files sort chronologically.
Entity — client, employee, supplier or matter reference, ideally matched against your existing system IDs.
Unique reference — invoice number, case number, payroll number.
Retention or destruction date — the field almost nobody captures and everybody later wishes they had, because it turns GDPR storage-limitation compliance into a scheduled task rather than a manual review.

Zonal OCR extracts these automatically from fixed positions on structured forms — invoices, application forms, timesheets — while unstructured documents are usually indexed by an operator with the image on screen. A well-defined convention here pays for itself; see our guide on how bad file naming and indexing ruin a scanning project.

What accuracy figure should you actually expect?

On clean, machine-printed modern documents, a production OCR engine will typically reach 98–99%+ character accuracy. That sounds close to perfect until you do the arithmetic. A page holds roughly 2,000 characters, so 99% character accuracy means around 20 wrong characters per page — enough to break several words. Across a 10,000-page archive that is a lot of imperfect text.

This is why accuracy should be specified at word or field level, not character level, and why the numbers vary by material:

Clean laser-printed originals — expect 99%+ word accuracy, minimal manual correction.
Photocopies and faxes — commonly 90–96%; degraded contrast and dot patterns cost you.
Older typewritten or carbon-copy material — highly variable, often needing a verification pass.
Handwriting — ICR handles neat block capitals reasonably; cursive remains unreliable and should be indexed manually where it matters.

For index fields specifically, insist on a verified accuracy target — 99.5% or better on key fields, achieved through double-key entry or automated validation against a reference list. Full-text OCR can afford the odd error; a wrong invoice number cannot.

Building the setup end to end

A workable specification for most UK businesses looks like this:

Capture: 300 dpi greyscale, colour where relevant, deskew and despeckle enabled, blank pages removed.
Recognition: full-page OCR on every page, English language pack, plus zonal extraction on structured document types.
Quality control: 100% page-count reconciliation against the original file, plus a sampled visual QC pass and field verification on indexed data.
Output: searchable PDF/A-2b, one file per logical document rather than per box or per batch.
Delivery: a structured folder tree mirroring your index fields, with a CSV or XML manifest so the files can be loaded into a document management system later without rework.
Originals: a documented decision to securely shred or to retain in off-site storage, recorded against the retention schedule.

Two compliance points deserve attention. Under UK GDPR and the Data Protection Act 2018, a searchable archive makes subject access requests far easier to answer within the one-month statutory deadline — but it also means personal data is easier to over-retain, so the retention field matters. And if you intend to destroy originals, BS 10008 gives you the framework for demonstrating that the digital copy is a trustworthy record, which is the standard most solicitors and auditors will look for.

Run a pilot before committing the full archive. A few hundred pages across your worst-quality material will tell you more about realistic accuracy than any specification sheet, and it lets you refine the index fields while changing them is still cheap. More guidance is available in our resources library.