Industries / Finance and business services / Records management
Retention rules
Retention rules say how long each kind of record must be kept and what happens to it afterwards. An organization sets them out in a retention schedule, a list of record types with a keeping period for each.
7
Milestones achieved, from 2000 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2000
Records software tested
On 27 November 2000 NARA endorsed the certification testing that the Joint Interoperability Test Command (JITC), part of the US Department of Defense, ran for 5015.2-STD, the department's "Design Criteria Standard for Electronic Records Management Software Applications". Software makers could pay to have their products tested, and certifications lasted two years. NARA endorsed the tests, not any product.
LessonCarlin warned that "the move towards electronic recordkeeping is more complicated than simply purchasing a 5015.2-certified records management application".
Credit
John W. Carlin, Archivist of the United States; JITC
2001
A shared international standard
ISO 15489, the international standard for records management, was issued in 2001. Barbara Reed of Recordkeeping Innovation, who was taking part in a global review of the standard in 2013, said it "was written to be technology neutral and to be principles based".
Credit
ISO
2003
Audit papers kept seven years
In January 2003, under section 802 of the Sarbanes-Oxley Act, the SEC adopted a rule requiring accounting firms to keep, for seven years, the workpapers and other records of their audits and reviews of public companies' financial statements. It took effect on 3 March 2003. The SEC's release quotes Senator Leahy on "the apparent massive document destruction by Andersen", Enron's auditor, and its "apparently misleading document retention policy". Section 802 also added sections 1519 and 1520 to the federal criminal code.
Credit
US Congress; SEC
2013
Email kept by role
NARA's Bulletin 2013-02 of 29 August 2013 introduced the Capstone approach. An agency schedules email "based on the work and/or position of the email account owner": the accounts of officials "at or near the top" of an agency are kept as permanent records, and the rest are kept for a set period. In September 2015 NARA issued General Records Schedule 6.1, which gives agencies using Capstone the authority to dispose of email.
Credit
NARA
2018
Storage limitation in law
The EU's General Data Protection Regulation, which applies from 25 May 2018, requires that personal data be "kept in a form which permits identification of data subjects for no longer than is necessary" for the purposes it is processed for, the principle called storage limitation. Archiving in the public interest, research and statistics are allowed longer periods with safeguards.
Credit
European Union
2022
Text messages are records
On 27 September 2022 the SEC charged 15 broker-dealers and one affiliated investment adviser with widespread recordkeeping failures, and they agreed to pay penalties totalling more than $1.1 billion. From January 2018 through September 2021 their employees "routinely communicated about business matters using text messaging applications on their personal devices". The firms admitted the facts.
LessonA recordkeeping rule covers business messages in whatever app employees use, including apps the firm does not control.
Credit
US Securities and Exchange Commission
2026
AI chats can be records
On 21 August 2026 NARA issued guidance on applying the Federal Records Act to AI. Using an AI platform does not by itself create federal records. Prompts, outputs, meeting summaries and audit trails become records depending on their use, such as when they are saved in an agency system and used for official business. AI-related records may be disposed of only under a NARA-approved records schedule.
Credit
NARA
Watch for
AI sorting with human checks
Iron Mountain says agents in its InSight DXP platform "classify documents" and can "Extract metadata and classify complex documents with 99% accuracy using human-in-the-loop validation", meaning that people check the AI's work. It says the figure "is based on internal AI and human-verified benchmarks" and gives no dataset, sample size or definition of an error.
Achieved whena customer or outside evaluator publishes accuracy on a stated set of records, with and without the human check.
Claimed by
Iron Mountain, on its product page as of November 2026
Retention for every Copilot app
Microsoft says that from June 2026, retention policies set for Microsoft Copilot "will automatically include all supported Copilot apps", with the change on by default. Its documentation says Copilot prompts and responses are kept as copies in a hidden folder of each user's Exchange mailbox, and warns that "Messages visible in your AI apps are not an accurate reflection of whether they are retained or permanently deleted".
Achieved whenan outside test confirms that prompts and responses from each Copilot app are captured and then deleted on schedule.
Claimed by
Microsoft, June 2026
Open
01
A scored retention classifier
A public test set of documents, each labelled by two records managers with the item of a published retention schedule it falls under, used to score AI tools that sort records for retention. Base it on the General Records Schedules (GRS) that NARA publishes, with documents that hold no personal data: fictional documents written for the test, and public agency documents such as policies, contracts and reports.
The briefCredit · unclaimed02
Sorting AI chats as records
A tool that applies NARA's August 2026 guidance to an agency's AI use and sorts each prompt, output and AI meeting summary into record, transitory record or non-record, giving its reasons, plus a test set to score it.
The briefCredit · unclaimed
Open 01 · the brief
A scored retention classifier
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 documents, each labelled by two records managers with the item of a published retention schedule it falls under, used to score AI tools that sort records for retention. Base it on the General Records Schedules (GRS) that NARA publishes, with documents that hold no personal data: fictional documents written for the test, and public agency documents such as policies, contracts and reports.
Why it is still open
Vendors publish accuracy without test sets, and Iron Mountain's 99% rests on internal benchmarks. NARA lists its own pilot of automated classification as "Planned for future pilot (TBD)". State Archives and Records NSW, in Australia, announced in 2017 that it would pilot machine learning for "sentencing", meaning assigning records to disposal classes, and searches found no published benchmark from that or any other pilot.
Built with what exists
NARA's General Records Schedules; Gemini Decision 1, which returns a probability for each permitted answer to up to 1,000 fixed-answer questions per request, so a tool can flag records it is unsure of (see the Note "A thousand questions, one pass"); open-weight models that an archive can run on its own servers, such as Mistral Large 4 (see the Note "A trillion open parameters") or a Liquid LFM3 model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune"); Label Studio for labelling.
First step
Pick 10 schedule items, such as travel, procurement and IT records, and write or collect 20 documents for each.
Done looks like
5,000 documents across at least 50 schedule items, each labelled twice with the agreement rate reported, and published accuracy for at least 4 tools, including the share of records each tool placed confidently and wrongly.
Open 02 · the brief
Sorting AI chats as records
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that applies NARA's August 2026 guidance to an agency's AI use and sorts each prompt, output and AI meeting summary into record, transitory record or non-record, giving its reasons, plus a test set to score it.
Why it is still open
The guidance makes the answer depend on "the circumstances surrounding the creation, maintenance, and use of the materials", such as whether the agency relied on an output in a decision or circulated it. It was issued on 21 August 2026, and searches found no tool or test set that applies it.
Built with what exists
synthetic AI conversations written to imitate agency work, such as travel approvals, freedom-of-information searches and policy drafts, with no real people in them; exports of Copilot conversations from a test account; a language model such as Claude given the guidance text and told to quote the passage behind each decision.
First step
Write 50 interactions covering each example in the guidance's list of likely records, and ask two records officers to label them.
Done looks like
1,000 synthetic interactions labelled by at least three federal records officers, with the tool's agreement with their majority published for each category and every disagreement listed.