BRENNEN

Industries / Finance and business services / Records management

Scanning

Scanning turns paper records into images and text: first onto microfilm, now into digital files that can be searched and, under some rules, can replace the paper.

5

Milestones achieved, from 1928 to 2023.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1928

    Checks copied on film

    George McCarthy, a vice president of a New York bank, invented a rotary camera that copied large volumes of bank checks onto film automatically. Eastman Kodak bought the invention in 1928, and McCarthy became president of Recordak, a new Kodak division that made and sold microfilm cameras. Banks took up microfilm in the 1930s, and other businesses followed more slowly, by Susan Cady's 1998 history of microfilm.

    webdoc.sub.gwdg.dethisdayintechhistory.com

    Credit

    George McCarthy; Eastman Kodak

  2. 1976

    Reading any typeface

    Kurzweil Computer Products introduced the Kurzweil Reading Machine in 1976. It scanned printed text, recognized each character "regardless of" typestyle, and spoke the words through a speech synthesizer. The Lemelson-MIT Program says it included the first CCD flatbed scanner and the first omni-font OCR software. OCR, optical character recognition, turns pictures of text into text a computer can search, and omni-font OCR reads print in almost any typeface.

    lemelson.mit.edu

    Credit

    Ray Kurzweil and his team

  3. 2004

    Images replace paper checks

    The US Check 21 Act, enacted on 28 October 2003, took effect on 28 October 2004. It created the substitute check, "a paper reproduction of the original check containing an image of the front and back", and made it the legal equivalent of the original. By the Federal Deposit Insurance Corporation's account, its purpose was "to facilitate check truncation and electronic check exchange", so that banks could exchange images instead of carrying paper checks. Banks were not required to create substitute checks.

    fdic.gov

    Credit

    US Congress

  4. 2014

    Paper processed in a mine

    In March 2014 David A. Fahrenthold of the Washington Post reported that 600 employees of the Office of Personnel Management (OPM) processed federal retirement papers by hand, about 230 feet underground in a former limestone mine near Boyers, Pennsylvania, with room for more than 28,000 file cabinets. A retirement took at least 61 days on average, the same as in 1977. Over 30 years, administrations had spent more than $100 million trying to automate the work. A system built by contractors went live in early 2008, broke, and was scrapped after more than $106 million had been spent.

    LessonThe Post found the failure fit a wider pattern in federal IT, where officials "tried to buy a technology they didn't fully understand" or "didn't test what they were getting".

    washingtonpost.comweb.archive.org

    Credit

    David A. Fahrenthold, the Washington Post

  5. 2023

    Digital copies replace originals

    NARA's rule of 4 May 2023, in effect from 5 June 2023, set standards for digitizing permanent federal records so that the paper can be disposed of. Agencies "must not dispose of the source records until they validate" that they met the rule's digitization and quality-control requirements. Once they have, they can use a General Records Schedule as authority to destroy the paper. NARA said it would not directly oversee every digitization project or approve project plans in advance.

    govinfo.gov

    Credit

    NARA, acting under 2014 amendments to the Federal Records Act

Watch for

  1. OCR that reads handwriting

    Mistral AI says Mistral OCR 3 "accurately interprets cursive, mixed-content annotations, and handwritten text layered over printed forms". It reports a 74% overall win rate over its own previous model, Mistral OCR 2, on forms, scanned documents, complex tables and handwriting, measured on "internal benchmarks based on real business use-case examples from customers". VentureBeat reported the same figure as a win rate against competing products.

    Achieved whenan outside group scores it on a published set of office documents with handwriting and forms.

    mistral.aiventurebeat.com

    Claimed by

    Mistral AI, December 2025

  2. Retirement papers go online

    OPM said that from 2 June 2025 new federal retirement applications had to be submitted through its Online Retirement Application, that it would stop accepting paper applications from 15 July 2025, and that the change "is expected to reduce backlogs, improve accuracy, and save taxpayer resources". In October 2025 OPM's director said the system "automates the front-end of the retirement process". OPM's published figures for cases processed in September 2025 put final processing of an immediate retirement "within 76 days", as reported by Government Executive.

    Achieved whenthe Government Accountability Office (GAO) or OPM's Inspector General reports average processing times, from application to final payment, below the 61 days the Post reported in 2014.

    opm.govgovexec.com

    Claimed by

    OPM, May and October 2025

Open

  1. 01

    A checker for digitized pages

    Software that checks every page image from a digitization project against NARA's 2023 standards before the paper is destroyed. It looks for low resolution, cut-off edges, skew, blank or missing pages, pages out of order and unreadable text, and writes a report that a second person can audit.

    The brief
    Credit · unclaimed
  2. 02

    Office paper test set

    A public test set of office paper as records centers hold it: carbon copies, faxes, stamped and initialled forms, typed memos with handwritten notes in the margin, and photocopies of photocopies, each with a field-by-field answer key. Use it to score OCR tools and AI models.

    The brief
    Credit · unclaimed

Open 01 · the brief

A checker for digitized pages

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Software that checks every page image from a digitization project against NARA's 2023 standards before the paper is destroyed. It looks for low resolution, cut-off edges, skew, blank or missing pages, pages out of order and unreadable text, and writes a report that a second person can audit.

Why it is still open

The rule leaves validation to each agency, and NARA said it "will not be conducting direct oversight of all digitization projects in all agencies". Andrew Potter, writing in his newsletter Meta Archivist in July 2026, cites NARA's 2024 annual report: 86% of agencies said they used NARA's digitization standards, and 71% said they had a validation process.

govinfo.govmetaarchivist.substack.com

Built with what exists

OpenCV, an open-source image library, for measurements; the Federal Agencies Digital Guidelines Initiative (FADGI) image-quality guidelines on which the rule draws; a model that reads images to compare each page with its neighbors and spot gaps; boxes of printed test documents.

  1. First step

    Scan one box of printed test pages twice, once properly and once with planted defects, and measure what a simple script catches.

  2. Done looks like

    On 50 test boxes, about 100,000 page images with 2,000 planted defects of 10 kinds, the checker finds at least 98% of the defects, and its error rates for each kind are published.

Open 02 · the brief

Office paper test set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set of office paper as records centers hold it: carbon copies, faxes, stamped and initialled forms, typed memos with handwritten notes in the margin, and photocopies of photocopies, each with a field-by-field answer key. Use it to score OCR tools and AI models.

Why it is still open

olmOCR-bench, an open OCR test from the Allen Institute for AI, holds 1,403 PDF files. Its largest group is 522 arXiv math papers, its old-scans group is 98 documents from the Library of Congress, and none of its groups is for forms or handwriting. Mistral's results for handwriting and forms come from internal benchmarks.

huggingface.comistral.ai

Built with what exists

fictional records written for the test, then printed, aged, photocopied and annotated by hand by volunteers, so that no real person's records are used; flatbed and phone scans; Label Studio for the answer key; olmOCR-bench's scoring code as a model.

  1. First step

    Write and print 50 fictional purchase orders and memos, have three people mark them up by hand, and scan them.

  2. Done looks like

    1,000 pages in 10 document types, scanned at two quality levels, with published field-level accuracy for at least 5 tools, including Mistral OCR 3 and an open-source engine such as Tesseract.