Shadow AI in BSA/AML Investigations: SAR Confidentiality and Runtime Controls in 2026
Shadow AI in BSA/AML investigations occurs when analysts paste case narratives, customer identifiers, transaction details or draft SAR text into AI tools the institution has not approved. Because SAR confidentiality under 31 CFR 1020.320(e) covers the SAR and any information that would reveal its existence, that content should not reach an unapproved external tool. Policy memos and after-the-fact review do not show what left the institution. Controls enforced in the execution path (allow, redact, block, escalate for human approval) apply at the moment of use and produce evidence an examiner can inspect.
Where unapproved AI use occurs in the analyst workflow
- 1
1. Alert review
Analysts triage alerts and may paste transaction rows or customer details into an assistant to explain a pattern. Exposed data: customer identifiers, account numbers, transaction amounts and dates. Control focus: redact identifiers before content leaves.
- 2
2. Case research
Analysts look up typologies, counterparties and geographies, often with case context in the prompt. Exposed data: relationships among parties, counterparty names, hints that an account is under review. Control focus: allow approved tools, block unapproved destinations.
- 3
3. Narrative drafting
Analysts ask an assistant to structure or polish the SAR narrative. Exposed data: draft SAR text and information that reveals a filing is being prepared. Control focus: block by default, or escalate for human approval.
- 4
4. QA review
Reviewers check consistency and clarity of near-final narratives. Exposed data: complete or near-complete SAR content. Control focus: enforce the same policy as drafting and retain evidence of each interaction.
Questions to evaluate any approach
- Does the control act before data leaves the institution, or only report afterward?
- Can it detect customer identifiers and transaction details in prompts, and redact them?
- Can it treat draft SAR text differently from general research queries?
- Does escalation record who approved or denied each held request?
- Does each interaction produce a retained record of tool, user, policy and action?
- Is AI use on investigation data covered by your model inventory, validation and monitoring under the April 2026 revised guidance?
What shadow AI means for SAR data, and what has changed
In a BSA/AML investigation, shadow AI is any use of an AI tool on case data that the institution has not approved, sized for risk, or placed under control. Typical examples are a public chat assistant, a browser-based writing aid or an unsanctioned extension. The motive is usually productivity: summarizing transaction patterns, researching typologies or tightening narrative wording.
The legal constraint is specific. Under 31 CFR 1020.320(e), a SAR and any information that would reveal the existence of a SAR are confidential and shall not be disclosed except as authorized in that paragraph. For this paragraph, a SAR includes any suspicious activity report filed with FinCEN under any regulation in the chapter. A joint statement dated September 2, 2026, published by NCUA and as OCC Bulletin 2026-43, restates that the BSA prohibits disclosure of a SAR or information that would reveal its existence, including to a customer or other person who is the subject of the SAR. It also notes that unauthorized disclosure could undermine law enforcement investigations, deter institutions from filing SARs and endanger SAR filers.
One point of analysis for counsel to confirm: because the rule reaches information that would reveal a SAR's existence, the data in scope for AI controls plausibly extends past the filed report to case notes, draft narratives and research content indicating a SAR is being considered or was filed.
On model risk management, the OCC and the Federal Reserve issued revised guidance on April 17, 2026 (OCC Bulletin 2026-13 and SR 26-2). Programs that cite SR 11-7 alone should update their references to the current guidance. It covers model development and use, validation and monitoring, and governance and controls, as of November 2026.
Runtime control options and their tradeoffs
These are design options, not requirements set by FinCEN, the FFIEC or the banking agencies. Each applies to an individual AI interaction, evaluated before data leaves the institution. The right mix depends on the tool, the user's role and the content in the request.
- Allow: Permit interactions with approved tools and low-risk content, such as general typology questions with no case data. This preserves analyst productivity.
- Redact: Remove or mask customer identifiers, account numbers and similar elements before the request proceeds. This keeps legitimate research possible, but it depends on reliable detection and does not address narrative text that is itself sensitive.
- Block: Stop requests containing draft SAR content, or any request to an unapproved destination. This is the strongest control and the most disruptive, so it should be reserved for content where redaction is not enough.
- Escalate for human approval: Hold an ambiguous request until a designated reviewer approves or denies it. This handles edge cases but adds latency, so the escalation criteria need to be narrow.
Enforcement in the execution path versus memos, training and review
Policy memos state what analysts should do. Training improves the odds that they do it. After-the-fact review can find some deviations, but only once the content has already left the institution. For SAR-related data, where the concern is disclosure itself, a control that acts after the fact cannot undo the exposure.
Enforcement in the execution path evaluates each interaction before it proceeds. The policy engine sees the destination, the user, and the content, and then allows, redacts, blocks or escalates. The control is therefore applied at the moment of use rather than inferred later.
The revised model risk management guidance emphasizes governance and controls, including clear policies and roles and responsibilities. The guidance snippets do not say controls must be technical, but a reasonable reading is that controls on AI use should be verifiable, not policy-only. Governance should also settle who owns approval of AI tools used on investigation data, and whether general-purpose assistants used in investigations belong in the model inventory and its validation and monitoring.
The evidence an examiner would want to see
A policy memo shows intent. Examiners also need to see what happened. The questions a risk leader should be able to answer are: which AI tools were used on case data, by whom, and what data was sent; which data elements were blocked, redacted or escalated; and who approved any escalated request.
That points to a record for each enforced interaction. Useful fields include the user and role, the tool or destination, the policy that applied, the action taken (allowed, redacted, blocked or escalated), what was redacted, the approver and outcome for escalations, and a timestamp. Records should be durable, retained under a defined schedule and tied to the policy version in force at the time. No examiner format is prescribed in the guidance, so this is a design target rather than a checklist from a regulator.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalate for human approval applied to individual interactions and tool calls. That enforcement generates continuous audit-ready evidence. Deployment is self-managed in the customer's cloud, with configurable retention and data residency. Trussed has run 18+ months in continuous production in a regulated enterprise at 99.95% uptime on live traffic.
Analyst workflow stages and the control point at each
Alert review
Customer identifiers and transaction details; redact before any AI use.
Case research
Case context and counterparty detail; allow approved tools, block others.
Narrative drafting
Draft SAR text; block or escalate for human approval.
QA review
Near-final narratives; enforce policy and retain evidence.
Review how your AI controls would hold up under examiner scrutiny
Talk through your investigation workflow, the data at each stage and the enforcement and evidence you would need.
Talk to an Expert