AI Compliance Assistant

The broker-dealer setting makes the boundary strict. FINRA Rule 3110 expects firms to maintain a supervisory system and written supervisory procedures that are reasonably designed for the firm's business. FINRA's AI materials also emphasize that AI-based tools introduce supervision and governance questions across broker-dealer functions. That means an AI compliance assistant should be governed as part of the supervisory environment, not treated as an informal search bar.

The strongest use case is evidence assembly. A supervisor may need to review a trade correction, exception queue, communication item, complaint response, financial reporting support file, or WSP control test. The assistant can gather the relevant procedure section, source records, dates, actors, prior approvals, unresolved gaps, and downstream outcomes into one review package. When multiple alerts share the same root cause, it should cluster them into a master case file for human review instead of scattering the same issue across disconnected queues. The reviewer should still see the citations and make the conclusion.

Procedure comparison is useful when it is concrete. The assistant can compare an activity record against the current WSP section that describes who should review it, what evidence should exist, when escalation is required, and how the decision should be documented. If the WSP and operating workflow disagree, the output should be a control gap for review, not an automatic judgment.

Retention is part of the design. Broker-dealer books-and-records rules require firms to make and preserve records under FINRA, Exchange Act, and related recordkeeping obligations. An assistant workflow may create its own evidence: prompt, retrieved records, cited answer, draft note, user edits, supervisor decision, approval timestamp, and final outcome. Those records should be preserved according to the firm's retention policy and the applicable recordkeeping framework.

Scoping runs before retrieval. A compliance officer, supervisory principal, FINOP, operations analyst, and technology user may have different access rights. The assistant should not retrieve customer, trading, financial, or supervisory records outside the user's entitlement and then rely on final-answer filtering to hide restricted data.

Retrieval should use source records, not loose memory. Relevant inputs may include trade blotters, order tickets, account records, customer complaints, communications, confirmations, reconciliation breaks, financial-reporting support, prior review notes, and WSP versions. Each retrieved item should carry source system, record ID, version where available, timestamp, and access result.

Summarization should separate fact from inference. The assistant can explain the activity, identify missing documents, group similar exceptions, detect stale procedure references, and draft a review note. It should use objective operational statuses such as "evidence pack staged," "data gap detected," "out of WSP timeline scope," or "review required." It should avoid binary legal labels such as "compliant," "violation," or "resolved" when the evidence does not prove that state and the designated reviewer has not made the decision.

Procedure comparison turns WSPs into operational checkpoints. The workflow can map a record to the relevant procedure, responsible supervisor, review cadence, required evidence, escalation rule, and retention output. This helps expose gaps between policy text and actual workflow without implying that software has made the legal or supervisory decision.

Review routing preserves accountability. If evidence is complete and the user has authority, the assistant can prepare the review package for a supervisor. If evidence is missing, stale, ambiguous, or outside scope, the workflow should hold the item, ask for clarification, or route escalation. Overrides should require a stated reason and retain the original warning.

Recording closes the loop. The final file should show the question, scope, sources, WSP version, model output, reviewer edits, decision, approver, timestamp, attachments, downstream state, and any exported report. That file is the practical bridge between AI assistance and an examinable supervisory record.