Key Takeaways
- Among the texts reviewed, the EU AI Act names automatic logging for high-risk systems; the US recordkeeping rules reviewed never mention prompts, outputs or model versions.
- US retention follows the record the AI supports: at least three, five or six years under the securities rules reviewed and 25 months for consumer credit applications.
- Broker-dealers must keep electronic records with a time-stamped audit trail or in non-rewriteable, non-erasable storage.
- Log the exact model version, the inputs actually used, the raw output, the human step and the final action for every decision (interpretation).
- From December 2, 2027, EU deployers of in-scope Annex III systems, such as credit scoring, keep logs under their control for at least six months, subject to other law.
Current as of October 30, 2026. This guide draws on SEC Rules 17a-4 and 204-2, FINRA Rule 4511, Regulation B (12 CFR 1002.12) and its official interpretations, FINRA's 2026 oversight report, the EU AI Act (Regulation (EU) 2024/1689) as amended by Regulation (EU) 2026/1744, the interagency model risk guidance SR 26-2, OSFI Guideline E-23, New York's Insurance Circular Letter No. 7 and Colorado Regulation 10-1-1, as published on that date. It is not legal advice.
An examiner reconstructing an AI decision has only what the firm captured at the time. Of the texts reviewed for this guide, the EU AI Act is the one that requires automatic logging by name: "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system." The US recordkeeping rules reviewed (Rules 17a-4 and 204-2, FINRA Rule 4511 and Regulation B section 1002.12) set what must be kept, for how long and in what kind of system, by type of record; none of them mentions prompts, model outputs or model versions.
Unless a step quotes a rule, it is a recommendation, labeled (interpretation). Companion guides cover whether AI meeting transcripts are books and records, drift monitoring and revalidation triggers, and adverse action reasons for machine learning credit models; this one covers the per-decision record.
- Settled (EU, high-risk systems): providers keep the logs their high-risk systems generate (Article 19) and deployers keep them too (Article 26(6)), in each case to the extent the logs are under their control and for at least six months unless other EU or national law, in particular data protection law, provides otherwise. Financial institutions subject to EU financial services governance rules keep them within the documentation those rules require. Under Regulation (EU) 2026/1744 these duties apply from December 2, 2027 for the systems listed in Annex III, which include credit scoring of natural persons; a system already on the market by then is covered only if its design later changes significantly (Article 111(2)).
- Settled (US): retention follows the record, not the technology: at least three years for Rule 17a-4(b) records, the first two in an easily accessible place; five years from the end of the fiscal year of the last entry for most Rule 204-2 records, the first two in an appropriate office; six years under FINRA Rule 4511 for FINRA records with no other specified period; and 25 months (12 for business credit, with exceptions) after notice of action taken for Regulation B application records.
- Settled (broker-dealers): electronic records must sit in a system with a time-stamped audit trail or in non-rewriteable, non-erasable storage, under the SEC's 2022 amendments (effective January 3, 2023; compliance date May 3, 2023).
- Not settled: whether a prompt or a raw model output is itself a required record under the US rules reviewed. FINRA's 2026 report describes storing prompt and output logs as a practice firms may consider, not a requirement. The EU AI Act lists minimum log contents only for remote biometric identification systems; for credit and insurance systems, Article 12(2) sets the purposes logs must serve, not a field list.
On scope: Rule 17a-4 binds broker-dealers and Rule 204-2 binds investment advisers "registered or required to be registered" with the SEC; FINRA Rule 4511 binds FINRA members. Regulation B binds creditors; a Regulation B final rule effective July 21, 2026 amended other sections and left section 1002.12 and its commentary unchanged. FINRA's report (December 2025) is written for member firms, and it "does not create any new legal or regulatory requirements or new interpretations of existing requirements". The EU AI Act binds providers and deployers of high-risk AI systems, a category that includes systems used "to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of AI systems used for the purpose of detecting financial fraud" and systems used "for risk assessment and pricing in relation to natural persons in the case of life and health insurance". SR 26-2 (April 17, 2026), from the Federal Reserve, the OCC and the FDIC, supersedes SR 11-7 and SR 21-8 and is non-binding: non-compliance "will not result in supervisory criticism", though its footnote 1 says "supervisory action may result for any violations of law or unsafe or unsound practices stemming from insufficient management of model risk." It is expected to be most relevant to banking organizations above an asset threshold it sets, and generative and agentic AI models are outside its scope. OSFI's Guideline E-23, issued September 11, 2025, takes effect May 1, 2027 for all federally regulated financial institutions in Canada. New York's Circular Letter No. 7 (July 11, 2024) is guidance for insurers authorized in the state on AI in underwriting and pricing. Colorado Regulation 10-1-1, amended effective October 15, 2025, binds insurers offering individual life, private passenger auto and health benefit plans that use external consumer data and models that use it.
Step 1: Map each AI decision to the recordkeeping rule that governs it
A credit decision falls under Regulation B, which requires creditors to keep, among other things, "any other written or recorded information used in evaluating the application". A broker-dealer communication falls under Rule 17a-4(b)(4), which covers "Originals of all communications received and copies of all communications sent (and any approvals thereof)" relating to the business. An adviser's communication about advice falls under Rule 204-2(a)(7). Whether a prompt or an internal draft is such a record is a judgment for the firm and its counsel (interpretation).
For each AI use case, record the decision it supports, the rule that governs that decision, the retention period and the system of record (interpretation).
Step 2: Log the model version and configuration with every decision
FINRA's report, among monitoring practices firms contemplating generative AI may want to consider, mentions "storing prompt and output logs for accountability and troubleshooting; tracking which model version was used and when; and validation and human-in-the-loop review of model outputs". OSFI's inventory fields for models with non-negligible risk include "model version" and "date of model's deployment into production". Colorado requires a "Documented up-to-date inventory, including version control" of external data sources and the models that use them. New York's circular letter says insurers' documentation may include "a description of the process for tracking changes of an insurer's use of ECDIS and AIS over time".
Those texts address the inventory; tying each decision to it is our step (interpretation). On every decision record, store the inventory ID, the provider's exact model version string, the versions of the prompt template and any retrieval index or rule set, generation settings and the decision threshold in force (interpretation). An alias such as a vendor's "latest" model name cannot be reconstructed later (interpretation).
Step 3: Capture the inputs the model actually received
For mechanical decisions, Regulation B's commentary permits a narrower record: a creditor that decides "based only on the items of information entered into the system, may comply with § 1002.12(b) by retaining the information actually entered."
For a generative system, the equivalent is the full prompt as sent, including the system prompt, conversation history and any retrieved documents or their identifiers and versions (interpretation). For a scoring model, keep the feature values as scored, not as stored upstream (interpretation).
Step 4: Keep the raw output, the final action and the human in between
Store the raw model output, any post-processing or filtering, and the action the firm actually took, as separate fields (interpretation). If a person reviewed, edited or overrode the output, record who, when and what changed (interpretation). For broker-dealers' electronic records, the audit-trail option in Rule 17a-4(f)(2) itself requires capturing "The date and time of actions that create, modify, or delete the record" and, "If applicable, the identity of the individual creating, modifying, or deleting the record".
For AI agents, FINRA's report lists among considerations "how to track agent actions and decisions", and it warns that multi-step agent reasoning "can make outcomes difficult to trace or explain, complicating auditability." Log each tool call, its inputs and result, linked to the originating request (interpretation).
Step 5: Store decision logs where they cannot be quietly changed
A broker-dealer's electronic recordkeeping system must either keep "a complete time-stamped audit trail" that includes "All modifications to and deletions of the record or any part thereof", or "Preserve the records exclusively in a non-rewriteable, non-erasable format". It must also be ready to "immediately provide" records regulators request. FINRA Rule 4511(c) requires FINRA records to be kept "in a format and media that complies with SEA Rule 17a-4." An adviser using electronic storage must "Arrange and index the records in a way that permits easy location, access, and retrieval of any particular record" and keep them "so as to reasonably safeguard them from loss, alteration, or destruction".
Application logs that engineers can edit or rotate do not meet those standards on their own where the log is a required record (interpretation). Write decision records to the firm's compliant archive, or to append-only storage with retention locks; hashing each record and chaining the hashes makes tampering detectable (interpretation).
Step 6: Set retention by the longest applicable rule, and minimize personal data
Where several rules apply, use the longest period (interpretation). Regulation B also requires keeping records beyond 25 months when the creditor "has actual notice that it is under investigation or is subject to an enforcement proceeding", so legal holds should reach AI logs (interpretation). In the EU, Article 26(6) sets a floor of "at least six months" for deployers, subject to other law, "in particular in Union law on the protection of personal data".
Prompt logs also copy the personal data in prompts. Keep identifiers in a separate, access-controlled store and reference them by token, restrict who can read raw prompts, and delete on schedule (interpretation). Do not minimize away what a rule requires you to keep: for a mechanical credit decision, the information actually entered is the least Regulation B's commentary accepts (interpretation).
Step 7: Test that a decision can be reconstructed before an examiner asks
SR 26-2 says "Adequate documentation helps to support effective model risk management", in a passage on model documentation rather than individual decisions. Among the US texts reviewed, Rule 17a-4(e)(8) addresses re-running a system output: broker-dealers keep "reports produced to review for unusual activity in customer accounts" for eighteen months or produce them promptly on request, and if a system change means a report "cannot be reproduced in any format using historical data", the firm "must promptly produce upon request a record of the parameters that were used to generate the report".
Each quarter, have someone outside the team rebuild a sample of decisions from the log alone (interpretation). Do not count on re-running the model: vendors retire versions and generative outputs can vary between runs, so the stored output is the evidence (interpretation). Treat a version change as a monitoring trigger, as the drift companion guide describes (interpretation).
Checklist
Items restate the recommendations above (interpretation).
- Each AI use case is mapped to the decision it supports, the governing record rule and its retention period
- Every decision record carries the inventory ID, exact model version, prompt or template version and thresholds
- Inputs are kept as the model received them, including retrieved context
- Raw output, final action and any human review or override are logged separately
- Agent tool calls are logged and linked to the originating request
- Required records sit in an audit-trail or non-rewriteable archive, indexed for retrieval
- Retention follows the longest applicable rule, with legal holds reaching AI logs
- Personal data in logs is tokenized and access-controlled without dropping required content
- A sample of decisions is reconstructed from logs on a regular schedule
How we sourced this guide: every factual statement in it is drawn from published rules, regulatory guidance and other public documents, and we keep a full set of sources and citations for it. If you would like them, or want to question a point, get in touch through our contact page and we will share them.
Frequently Asked Questions
Does any US rule require firms to log AI prompts and outputs?
Not by name, in the texts reviewed for this guide. SEC Rules 17a-4 and 204-2, FINRA Rule 4511 and Regulation B set retention by type of record and do not mention prompts or model outputs. A prompt or output may still be a required record if it is, for example, a broker-dealer communication or information used to evaluate a credit application (interpretation). FINRA's 2026 report describes prompt and output logs as a practice firms may consider, and says it creates no new requirements.
How long should AI decision logs be kept?
As long as the record they support: at least three years for Rule 17a-4(b) records (the first two easily accessible), five years from the end of the fiscal year of the last entry for most Rule 204-2 records, six years for FINRA records with no other specified period, and 25 months (12 for business credit, with exceptions) for Regulation B application records, longer under investigation. Where several apply, use the longest (interpretation). Under the EU AI Act, providers and deployers of high-risk systems keep logs under their control for at least six months unless other law provides otherwise.
When do the EU AI Act logging duties apply to banks and insurers?
For high-risk systems listed in Annex III, which include creditworthiness assessment of natural persons and life and health insurance risk assessment and pricing, Regulation (EU) 2026/1744 sets December 2, 2027 as the date the high-risk requirements, including logging and log retention, apply. Systems classified as high-risk under Article 6(1) and Annex I follow from August 2, 2028. A system already on the market by then is covered only if its design later changes significantly. Financial institutions subject to EU financial services governance rules keep the logs within the documentation those rules require.
Is it enough to keep the model version and inputs and re-run the model later?
Not on its own, in our view (interpretation). Vendors retire model versions and generative outputs can differ between runs, so a re-run may not reproduce what the firm relied on. Keep the stored output, the final action and any human review alongside the version and inputs, in storage that meets the applicable recordkeeping rule.