DataPacts

Paper Archives: Scanning and Licensing Physical Records for AI

Checked

At a glance

A paper archive can be licensed for AI training only after it is scanned and converted to text, and in the text read on 2026-11-17 none of the buyer pages asks for scanned paper by name, so the first question is whether any buyer wants a given archive.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

This page covers buyers' own pages, the scan and text recognition step, rights, personal data and retention limits for the wider field of AI training data licensing, and when not licensing is the sensible outcome.

How this page was checked: every figure carries its date and the page it was read on, as described in how we source, date and check every figure.

What Counts as a Paper Archive in an AI License

A paper archive here means records that exist only or mainly on paper: files, ledgers, engineering drawings, case files, correspondence and similar bundles, as distinct from born-digital files. A scan of such a record is a page image, and text recognition (OCR) produces the text layer that a training set would use.

Records that were created on a computer and stored in drives or wikis are a different asset, covered under documents, wikis and drive files.

Which Buyers Name Scanned Paper in Their Own Pages

On 2026-11-17 none of the 26 Mercor listings and none of the seven other program pages read names scanned, paper, handwritten or microfilmed records among what it wants. The table compares the sources read and the wording closest to paper in each.

Source read What the page states closest to paper Scanned or paper wording Source
Mercor, 26 listings (published 2026-10-17 to 2026-11-12) accounting: "Source documents", "Client ledgers and bank statements"; construction: "drawings, specs, RFIs, submittals, change orders and daily logs"; manufacturing nice-to-have: "Asset photos, drone imagery and scans linked to work orders"; research lab requires "Electronic lab notebooks or structured experiment records" none api.data.mercor.com/data-acq/api/v1/listings
Centific, Handshake AI, Scale, micro1 program pages as served; nothing stated none the programs' own pages
Defined.ai, Troveo home or program pages as served; nothing stated none the companies' own pages
Polyshares home page: buys "operational data" (Slack, HubSpot, Gmail named); nothing stated on paper none polyshares.com
Protege, Turing, Human Native, Sieve pages returned little or no readable text, or an error; not read not read n/a

"Nothing stated" means no wording on scanned, paper, handwritten or microfilmed records in the text read, which is not a refusal.

The same Mercor listings state whole-deal ranges as Mercor's own claims, checked 2026-11-17, for the record kinds they list: USD 25,000 to 865,000 for accounting, USD 115,000 to 1,000,000 for construction and USD 50,000 to 1,250,000 for a law firm.

None states a range for paper, and none states a price per page or per box.

A holder can ask a program whether it takes scanned material, and which format and resolution it wants, before any scan is paid for.

The Scan and Text Recognition Step and the Prices Vendors Publish

Turning paper into text takes two steps, imaging each page and recognizing the text, and the only prices on vendors' own pages that could be read are for the second step. The table lists the recognition prices as published by each vendor, checked on 2026-11-17; neither page carries an effective date.

Vendor and product Published price Unit and tier Source
Amazon Web Services, Textract Detect Document Text (US West, Oregon) USD 0.0015 per page; USD 0.0006 per page first 1 million pages a month; pages beyond 1 million a month aws.amazon.com/textract/pricing
Google Cloud, Document AI, Enterprise Document OCR Processor free; USD 1.50 per 1,000 pages; USD 0.60 per 1,000 pages first 1,000; 1,000 to 5,000,000; above 5,000,000 (units printed as "count", read as pages from the page's per-1,000-pages examples; first price column; other columns are discount options) cloud.google.com/products/document-ai/pricing
Microsoft Azure, AI Document Intelligence, Read "$-" placeholders on the page as served no figure recorded azure.microsoft.com

These figures are charges for recognition software only. They exclude handling, boxes, preparation, quality checks and any removal of personal data, and no total cost per archive is stated or computed here.

On the scanning-vendor pages tried on 2026-11-17, none gave a verified price per page or per box for physical scanning, so none is stated.

The court position on digitizing is narrow. In the order of 2025-06-23 (Dkt. 231), the Northern District of California held that digitizing purchased print books was fair use and refused to excuse pirated library copies.

That ruling concerns books Anthropic bought and scanned, not a company's own archive or any licensing deal. The books page covers bought in print and scanned books in full.

Who Holds the Rights to What Is Written on Old Paper

A company can license the content of its old records only where it holds the right to license that content, and holding the paper does not by itself settle that.

Under 17 U.S.C. 201(b), the employer is the author and owner of a work made for hire unless both parties sign a written instrument saying otherwise.

Under the UK Copyright, Designs and Patents Act 1988, s.11(2), the employer is first owner of a literary work made by an employee in the course of employment, subject to agreement.

Records on paper come from three sources. Staff wrote some, and the two provisions above speak to those. Outsiders wrote others (letters received, consultants' drawings, client papers), and those provisions do not make them the company's.

The third group is old or acquired holdings of unknown origin. The seller sorts the boxes into these three groups before scanning, as part of a rights audit of the archive.

Personal Data Inside Files, Ledgers and Correspondence

Old paper records usually hold personal data in the text, the handwriting and the margins, and scanning copies it without removing it. Under GDPR Art. 9(1), health, biometric, trade-union and similar special-category data need an Art. 9(2) exception plus an Art. 6 basis (applies from 2018-05-25).

In the US, 45 CFR 164.514(b) sets two routes to de-identification, expert determination or Safe Harbor with 18 identifier types, and de-identified information is outside the Privacy Rule under 164.502(d).

Identifiers in a scan sit in typed text, handwriting, stamps and file labels, so a text-only filter does not clear an image. The steps are set out under PII removal and de-identification before licensing.

Art. 14(5)(b) GDPR exempts a controller from the notice duty where providing the information proves impossible or involves a disproportionate effort, with appropriate measures; otherwise Art. 14(3)(a) requires notice within a reasonable period and at the latest one month.

The exemption concerns notices, not the basis for processing, and does not remove the Art. 9 question. Case files and medical charts are the clearest special-category cases, and notices are covered under GDPR notices before a data sale.

Destroying Originals After Scanning: What the Retention Rules Leave Open

Retention duties attach to records, and the rules recorded for this site set minimum periods, not whether a scan may replace the paper.

UK private company accounting records must be kept 3 years under Companies Act 2006 s.388(4), and HMRC expects 6. Germany requires 10 years for books and 8 years for booking vouchers under HGB 257(4). The minimum retention periods by country are set out in full on the retention page.

A rule on whether a scanned copy may replace an original is not recorded here, so the page does not say it may or may not.

A license that requires deletion of the buyer's copy is a separate matter from the seller's originals, and the written text of the deletion or return clause decides what it reaches, so the seller checks that clause before anything is discarded.

Scanning and Licensing a Paper Archive: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Scanning an archive turns stored paper into a licensable asset only if a buyer wants it, and today no buyer page read names scanned paper.

Benefits

  • An archive may hold records that have no digital copy anywhere else.
  • The scan is a dataset the company controls.
  • Whole-deal ranges exist for the record kinds Mercor lists (Mercor's claims, 2026-11-17).

Risks

  • Outsiders' papers in the boxes may not belong to the company to license.
  • Health and other special-category data in files and charts need a legal basis and de-identification.
  • Retention duties may apply if originals are discarded after scanning.

Downsides

  • Scanning and recognition cost is real, and no total for an archive is on record.
  • No buyer ask for scanned paper is on record in the text read.
  • A delivered copy usually stays with the buyer, as the documents page explains.

Alternatives

  • Ask a program whether it takes scanned material before scanning anything.
  • License a smaller born-digital set instead, as described on the documents page.
  • Give the archive to a library or archive that licenses collections.
  • Not scanning or licensing at all, keeping the paper under the retention rules; the question is covered under should you sell or license your data.

Not licensing is a valid outcome.

Neighboring Record Types and the Seller Steps for a Paper Archive

Paper records overlap with several other pages: born-digital documents, books, news archives, library collections, legal files, financial records and drawings each have their own.

The seller steps for a paper archive run in this order:

  1. Build a data inventory for an AI data sale and list the boxes.
  2. Sort them in the rights audit, then remove personal data.
  3. Ask the program about format and sample needs.
  4. Sign NDAs before sharing data with an AI buyer and before a sample.
  5. Read the AI data licensing agreements clause by clause.
  6. Check the deletion, return and machine unlearning clauses.

A shut-down or defunct company with boxes of records follows a related path.

This page is information, not legal advice.

Frequently Asked Questions

Do AI companies buy scanned paper archives?

No buyer page read on 2026-11-17 names scanned paper. The listings and program pages state nothing on paper, which is not a refusal, so the holder asks the program first.

How much does it cost to scan and OCR an archive?

Two vendors publish recognition prices, but no scanning price per page or per box was verified. On 2026-11-17 Amazon Web Services listed USD 0.0015 per page for the first million pages a month and Google Cloud USD 1.50 per 1,000 pages from 1,000 to 5,000,000 pages. See what company data is worth to AI companies for valuation methods.

Can handwritten records be used for AI training?

No buyer page read mentions handwriting. Handwriting also carries personal data and identifiers that text filters miss, so the privacy checks above apply before any license.

Do I have to keep the paper after scanning?

Retention minimums apply to the record, and whether a scan can replace the original is not recorded here. The retention page lists the minimum periods, and the contract's deletion clause is checked separately.