Is an AI prompt a Rule 17a-4 communication?

The answer depends on the prompt's content, business purpose and the firm's applicable recordkeeping analysis; Rule 17a-4 preserves specified communications relating to the broker-dealer's business, but a technical gateway should avoid making the legal classification by itself. Record the identity, business reference, destination, data class and policy outcome. The records team can then map the interaction or its payload to the correct category and retention schedule.

Can AI perform the principal review required by FINRA Rule 3110?

AI can sort a population, identify terms or propose exceptions. Rule 3110 assigns supervisory duties to the member and requires written evidence for specified reviews by a registered principal. Preserve the model interaction and its result, then record the principal's disposition in the supervisory system. The model helps process material. It does not hold the registration, authority or accountability assigned to the principal.

Which AI events should be linked to order records?

Link an event when model output affects extraction, entry, modification, cancellation, routing or review of an order or instruction. Use the order memorandum identifier supplied by the application and keep timestamps for receipt, AI processing and entry. For an October 2, 2026 sample, a reviewer should be able to move from the Rule 17a-3 memorandum to the exact request and then to the associated person's acceptance or correction.

How should a firm evidence blocked AI requests?

Keep the authenticated identity, event time, business context, destination, detected data class, policy version and block reason. Preserve enough information to investigate without duplicating sensitive content into a general log. A block event proves that the enforcement point acted before transmission on the routed path. It says nothing about personal devices, local models or vendor-native inference, so the supervisory procedure should state those exclusions.

Capital Markets AI Audit Trails Must Join Supervision to Books and Records

SEC Rule 17a-3 requires broker-dealers to make and keep current specified records, including order memoranda with account, time, price and responsible-person details. Rule 17a-4 sets preservation duties for those records and for business communications. FINRA Rule 3110 requires a supervisory system, written procedures and documented reviews.

An LLM can now draft or classify information between those control points. AI audit trail capital markets design has to reconnect the model interaction to the order, communication or review that the firm already supervises. A transcript sitting in a vendor console is a weak substitute for that link.

TL;DR

SEC Rules 17a-3 and 17a-4 govern creation and preservation of specified broker-dealer records.
FINRA Rule 3110 requires written supervision and written evidence of specified reviews.
AI request records should bind the user, business context, data class, model and policy outcome to the authoritative record.
Gateway coverage excludes local models, vendor-native inference and traffic outside routed HTTP model calls.

Rule 17a-3 starts with the business event

The SEC record-creation requirements in Rule 17a-3 require broker-dealers to make and keep current books and records relating to their business. The rule specifies itemized daily transaction records and detailed memoranda for brokerage orders or instructions. Depending on the record, fields include the account, terms, receipt time, entry time, execution price and identities of associated people.

An AI assistant can summarize an instruction, extract order terms from an email or flag an exception. The authoritative order memorandum still belongs in the broker-dealer's books. The model interaction is supporting evidence that explains an intermediate step.

Suppose an associated person receives a bond instruction at 8:57 a.m. and asks an assistant to extract the terms. The order system records the entered instruction at 9:01. Without a request reference, compliance cannot see the omitted call condition that appeared in the email but vanished during summarization. The gap lasts four minutes and one model response.

Rule 17a-4 makes retention a design decision

The SEC preservation requirements in Rule 17a-4 assign periods to specified broker-dealer records. The rule preserves certain records for six years and others for three years, with special accessibility requirements during the first two years. Paragraph (b)(4) covers originals of communications received and copies of communications sent relating to the broker-dealer's business, including electronic communications within scope.

AI records need a documented relationship to those categories. A prompt that asks an approved assistant to draft a customer response may be part of the production history, while the sent response remains the authoritative communication. An internal research query may support a separate review record. The firm should classify each use case and map retention before rollout.

Keeping all AI traffic for one generic period is administratively neat and evidentially careless. The schedule should identify the linked business record, legal basis, owner and disposal trigger. AI governance for capital markets covers approval and inventory at the use-case level.

FINRA supervision needs evidence of review

FINRA Rule 3110 requires each member to establish and maintain a supervisory system reasonably designed for compliance. Its written procedures cover the firm's business and associated-person activities. The rule also requires principal review, evidenced in writing, for specified transactions, correspondence and internal communications.

A model can assist that review without becoming the principal. If AI classifies correspondence for complaint indicators, the evidence should show the population submitted, model and policy used, exceptions returned and principal's disposition. The request log establishes which analysis occurred. The supervisory record establishes who reviewed the result and what action followed.

I would never let a green "review complete" badge stand in for a registered principal's judgment. It compresses a regulated act into a user-interface state controlled by the same application that wants the workflow to look finished.

AI vendor risk in capital markets addresses the due-diligence record when a third party supplies that classification feature.

Six fields create usable lineage

The first field is the authenticated person or service identity supplied by the application. The second is a business reference, such as an order memorandum ID, communication ID, account-review case or research approval. Third comes the data classification, which can identify material nonpublic information, customer information or public market data.

The record also needs the resolved model and tenant, the policy version and enforcement result, plus a timestamp tied to a consistent clock source. Use the gateway as a source of request evidence. Put the review disposition in the supervisory system and link it back to the interaction.

A hash or controlled reference can protect full payloads that contain customer data or deal information. Complete prompt retention in a broadly accessible observability tool creates another sensitive-data store. Signed audit logs for AI requests explains how signature verification and a separate write path protect the event metadata against silent alteration.

Material nonpublic information needs pre-transmission policy

A research analyst may use the same assistant for a public 10-K summary and a draft note containing material nonpublic information. An allowlist of model domains treats both requests alike. Policy at the request boundary can distinguish them when the application supplies identity and content classification.

A request with public issuer data may proceed to an approved endpoint. A request carrying deal code names, restricted-list context or customer account data can be blocked, redacted or routed to an approved private environment. Each decision should record the rule and version applied. This preserves evidence even when the request never reaches the model.

The event population also supports surveillance. Compliance can group blocked requests by desk, use case and destination, then investigate changes after a policy release. Shadow AI in capital markets covers discovery of unapproved services; request-level records address the approved routes the firm controls.

Reconciliation catches missing events

A control test should reconcile AI interactions with the relevant business population. Compare customer correspondence marked as AI-assisted with gateway records and principal review evidence. Order-workflow samples should trace interaction references to Rule 17a-3 records. In research, confirm that model use followed the approved information-barrier policy.

Sequence numbers and signed daily manifests can expose gaps. If the application received a model response but no event reached the evidence store, the exception should appear before an examiner asks for the record. A separate enforcement point reduces the self-attestation problem because the business application lacks custody of the event write path.

The review result needs a date and owner. "AI logging enabled" is a configuration statement. "October 1, 2026 reconciliation found two missing application references across 4,218 routed requests, both closed by Market Supervision" is test evidence, provided the count is drawn from the firm's actual review rather than a marketing claim.

The HTTP boundary belongs in the written procedures

An HTTP proxy covers model calls routed through it by authenticated applications, users or agents. It misses a local model running on an analyst workstation, inference performed inside a market-data terminal and vendor-native AI inside a communications platform. It also misses a personal device outside firm management.

Written supervisory procedures should name covered channels and assign controls to excluded ones. Routed research and correspondence assistants may use gateway evidence. A terminal's embedded model may depend on vendor exports and application review. Local inference may be prohibited and monitored through endpoint controls.

AI audit trail requirements by regulation describes why a system inventory and request trail answer separate questions. For a broker-dealer, the procedures should connect both to the Rule 3110 owner and state the October 2026 test used to validate coverage.

DeepInspect

DeepInspect is a stateless proxy for authenticated HTTP traffic between capital-markets users or agents and LLM endpoints. It evaluates application-supplied identity, request classification, approved destination and policy before forwarding. Every permit, redaction, reroute or block creates a signed per-decision record outside the calling application's write path.

For routed workflows, those records can connect an order, communication or review reference to the model and policy used. DeepInspect does not maintain the broker-dealer's Rule 17a-3 books, decide Rule 17a-4 retention, perform principal review or cover embedded vendor inference. Book a demo today.