Industries / Finance and business services / Records management
Finding records
Finding records is filing and indexing them so the right one can be found again, from folders in a cabinet to search engines and AI assistants that answer questions from an organization's files.
4
Milestones achieved, from 1900 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1900
Letters filed on edge
Offices once stored letters lying flat in drawers or folded into pigeonholes. Secondary sources say the vertical file, which stands papers on edge in folders behind guides, was invented in 1892 by a Dr Rosenau and shown at the 1893 World's Fair in Chicago. A business history by Yates says the Library Bureau devised guides and folders for filing letters on edge in 1892 and won a gold medal for them at the fair. The Early Office Museum checked primary sources and found no vertical file at the 1893 fair: the Library Bureau exhibited a "Card-case for records of charitable societies". The first vertical filing cabinet for letters it found in the Library Bureau's catalogs is in the 1900 edition.
Credit
the Library Bureau; Dr Rosenau
1985
Search found a fifth
David C. Blair and M. E. Maron tested IBM's STAIRS full-text search system on about 40,000 documents, some 350,000 pages, gathered for the defense of a $237 million lawsuit over the building of San Francisco's BART rail system. The lawyers said they needed at least 75% of the relevant documents. Their searches found 20% on average.
LessonIn 1996 Blair wrote that one of the study's main arguments was that recall, the share of relevant documents found, "fall[s] off significantly as the document collection grows larger".
Credit
David C. Blair and M. E. Maron
2020
Handwriting made searchable
Transkribus, a platform built by the EU-funded READ project and coordinated by Günter Mühlberger of the University of Innsbruck, recognizes and transcribes handwritten documents after a user trains it on 50 to 100 transcribed pages. By December 2020 it had more than 45,000 users, by the European Commission's account. Finland's National Archives had used it since 2016 and had released about 800,000 transcribed documents.
Credit
the READ project; Günter Mühlberger
2023
Answers from company files
Copilot for Microsoft 365, an AI assistant for work, went on sale to enterprise customers on 1 November 2023. Microsoft said its chat feature "combs across your entire universe of data", meaning a user's "emails, meetings, chats, documents, and more". Microsoft's deployment guidance says Copilot grounds its answers "in the data users already have permission to access", and the first step it lists is "remediating oversharing": fixing files left open to more people than should see them.
LessonBecause the assistant inherits existing permissions, Microsoft's own guidance puts fixing overshared files before rollout.
Credit
Microsoft
Watch for
Meaning-based search at NARA
NARA lists pilots, all in progress, that use Google's Vertex AI platform for semantic search, which matches the meaning of a query instead of its exact words, on the National Archives Catalog and on presidential email archives (the "EOP 42" records). Another pilot auto-fills descriptive metadata for archival records. NARA says the email search is a test of whether such search "can understand context and intent more effectively than keyword matching".
Achieved whenNARA or an outside evaluator publishes how many relevant records the semantic search finds that keyword search misses, on a stated set of queries.
Claimed by
NARA, in its AI use case inventory updated February 2026
Open
01
Find every record, measured
A public test of whether AI search tools find every record that matches a request, as a records officer must for a freedom-of-information request or an audit, not just enough records to answer a question. Generate a synthetic organization's email, chat and documents with planted sets of matching records, write 100 requests, and judge every document against every request so that recall can be counted exactly.
The briefCredit · unclaimed02
Box lists from photographs
A phone app and pipeline that photographs the folder tabs and labels in a records box and writes the box's folder list: each folder's title and date range, and a suggested records series, for a person to check.
The briefCredit · unclaimed
Open 01 · the brief
Find every record, measured
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of whether AI search tools find every record that matches a request, as a records officer must for a freedom-of-information request or an audit, not just enough records to answer a question. Generate a synthetic organization's email, chat and documents with planted sets of matching records, write 100 requests, and judge every document against every request so that recall can be counted exactly.
Why it is still open
The Blair and Maron study measured recall on one lawsuit's documents in 1985. EnterpriseRAG-Bench, released in May 2026, offers about 500,000 synthetic company documents and 500 questions "from simple single-document lookups to multi-document reasoning". Its abstract describes questions to answer, not requests that must return every matching record. NARA lists its own AI pilot for finding records for freedom-of-information requests as "Planned for future pilot (TBD)".
Built with what exists
EnterpriseRAG-Bench's open generation framework, to make the corpus; the search in Microsoft 365 Copilot and in open-source tools such as Elasticsearch; AI agents built on models such as Claude Fable 5.2 (see the Note "Three working devices"); Blair's sampling method for estimating how many relevant documents a search missed.
First step
Generate a 10,000-document corpus with the framework and plant 20 sets of matching records by hand.
Done looks like
100 requests over at least 100,000 synthetic documents with complete relevance judgments, published with recall and precision for at least 4 tools, and a count of the records each tool missed while reporting that it had finished.
Open 02 · the brief
Box lists from photographs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A phone app and pipeline that photographs the folder tabs and labels in a records box and writes the box's folder list: each folder's title and date range, and a suggested records series, for a person to check.
Why it is still open
Boxes in storage are often poorly listed. When the 1973 St. Louis fire struck, no index of the lost files existed, so no complete list of what burned could be made. NARA's pilots that fill in descriptive metadata by AI work on digital records and were still in progress in February 2026.
Built with what exists
phone cameras; a model that reads images, such as Gemini or Claude, for typed and handwritten tabs; Gemini Decision 1, which picks from a fixed list of answers and gives a probability for each, to choose the series and flag doubtful folders (see the Note "A thousand questions, one pass"); Label Studio, an open-source labelling tool, for the answer key.
First step
Build 10 mock boxes from printed public-domain government reports and fictional files, and photograph them.
Done looks like
100 mock boxes holding about 3,000 folders of fictional and public-domain documents, with typed, handwritten and faded labels, and published folder-level accuracy for titles and date ranges against hand-made lists, with the time per box compared with a person typing the list.