Scanned, Filed and Forgotten: Getting Value Out of Twenty Years of Documents

A lot of businesses ran a scanning project at some point. Boxes went in, PDFs came out, the filing cabinets went to the tip, and everyone agreed it was a good thing.

Then nobody ever opened any of them again, because a scanned PDF is a photograph of information rather than information. You can find it if you know its filename. You cannot ask it a question.

This article is about what it takes to change that, what it costs, and when it is not worth doing.

This article is part of our guide to the data you already own.

The four states a document can be in

Cost and usefulness both depend entirely on which of these you are starting from.

Paper. Not digitised. Findable only by someone who knows where it is.

Image. Scanned, but the text is a picture. Searchable by filename and metadata only. Most legacy scanning projects produced this.

Text. The words are extractable, either because it was born digital or because optical character recognition has been run. Full-text searchable.

Structured. The meaningful facts have been pulled out into fields. Contract start date, party names, value, renewal terms. This is the state where a document archive becomes genuinely useful, and almost nobody's is here.

The jump from image to text is cheap and largely automated. The jump from text to structured is where the work and the value both sit.

What it costs

Rough UK figures, per thousand documents.

Paper to image. £40 to £200 per thousand pages for bulk scanning services, more for anything fragile, bound, or unusually sized. Add handling and transport.

Image to text. Effectively pennies per document now. Optical character recognition on clean printed text is a solved problem. Handwriting is not solved, and is considerably more expensive and less reliable.

Text to structured. £2,000 to £15,000 per document type, depending on how variable the format is. A standard contract template across ten thousand documents is cheap. Twenty years of correspondence in no consistent format is not.

Ongoing. Whatever you spend keeping it current. An archive that stops being added to stops being useful within about two years.

The economics have changed substantially in the last couple of years, mostly at the text-to-structured stage, which used to require either manual extraction or brittle template matching.

The question that decides whether to do it

Not "is the archive valuable". Almost any archive is interesting. The question is:

What decision would we make differently if we could query this?

Finish the sentence or do not start. Real examples that pass the test:

• "We could tell which of our contracts have uncapped liability, which we currently cannot."
• "We could answer a client's question about a job from 2014 in a minute rather than a day."
• "We could find every drawing that used the superseded tolerance."
• "We could tell how our prices have moved against this client over fifteen years." See your pricing history is a dataset.

Examples that fail it, and that we hear regularly: "it would be good to have it all searchable", "we might find something useful", "we want to use AI on it".

The slice test

Before committing to an archive project, take one year or one document type and do the whole thing on that alone.

Three to six weeks, usually £3,000 to £8,000, and it answers the three questions that determine whether the full project is viable. Is the source material good enough to extract from? Does the extracted information actually answer the question? And is the answer worth what the full extraction would cost?

Suppliers who skip straight to the full archive are skipping the only stage that could tell you to stop.

When not to do it

When the facts are already in a system. Frequently the useful information, dates, values, parties, was keyed into a database at the time and the document is just the evidence. Check before scanning anything. This is the most common wasted archive project.

When it is ballast. If you keep it because you must and would never query it, keep it cheaply. Move it to archive-tier storage and leave it alone. See dark data: how to audit what your business is already storing.

When it is mostly handwriting. Extraction accuracy on handwritten material at scale is still poor enough that you will need human verification, which changes the economics entirely.

When you should have deleted it. Some archives are a retention liability rather than an asset. Extracting personal data from twenty-year-old files you have no basis to hold makes the problem worse, not better.

Key Takeaways

• A scanned PDF is a photograph of information. Searchable by filename, not by content.
• Four states: paper, image, text, structured. The jump from image to text is now nearly free. Text to structured is where the work and the value are.
• Finish the sentence "we would decide ___ differently" before starting. Wanting it searchable is not a reason.
• Run a slice test on one year or one document type first, at £3,000 to £8,000. It is the only stage that can tell you to stop.
• Check whether the facts are already in a system. The most common wasted archive project extracts information that was keyed in at the time.

Frequently Asked Questions

Can AI just read our whole archive?

It can extract and summarise at a scale and cost that was not available three years ago, which has genuinely changed what is viable. It cannot tell you whether the source documents are reliable, which is the subject of data quality before AI, nor which question is worth asking. Those two remain the determining factors.

How accurate is the extraction?

On clean printed text, high enough to rely on with spot checks. On poor scans, unusual layouts or handwriting, considerably lower, and the honest answer from any supplier should include a verification step and a stated accuracy target. Anyone quoting without seeing samples of your actual documents is guessing.

Should we scan the paper we still have?

Only if it passes the decision test above, and only after checking whether the facts are already recorded elsewhere. Scanning as insurance is a reasonable separate argument, and it is a much cheaper project because you do not need extraction at all.

---

Have an archive nobody can query and are not sure it is worth the money? Talk to Halo Technology Lab. Our strategy and scoping service starts with a slice test, and we will tell you when the answer is already in a system you own.