Drafting the narrative is the slowest part of most suspicious activity reports, and it is now the most common place compliance teams try AI. An AI SAR narrative can turn an investigator's case notes and transaction data into a structured draft in minutes. Speed is real. But the narrative is the part of the SAR law enforcement actually reads, the filing decision belongs to the institution, and examiners test SAR quality as closely as SAR timeliness.
Key Takeaways:
- FinCEN's narrative guidance still applies: who, what, when, where, why, and how, in a clear and complete account.
- The decision that activity is suspicious and a SAR must be filed stays with the institution under 31 CFR 1020.320.
- AI drafts must be checked against source data. A fabricated or misstated fact in a SAR is a serious problem, not a typo.
- Keep the investigator's work, the AI draft, the reviewer's edits, and the filing decision in one record retained for five years.
What FinCEN Expects in a SAR Narrative
FinCEN's long-standing Guidance on Preparing a Complete and Sufficient SAR Narrative remains the reference. A sufficient narrative answers:
- Who is conducting the activity, with identifying details, roles, and relationships
- What instruments, products, or mechanisms were used
- When the activity occurred, including the date range and whether it is continuing
- Where it took place, including accounts, locations, and jurisdictions
- Why the institution considers it suspicious, connected to typologies or red flags
- How the activity was carried out, in a chronological account a reader can follow
The guidance stresses clarity and completeness: law enforcement readers should understand the activity without access to your internal systems. Vague narratives, boilerplate, and unexplained jargon are common examination criticisms. See our guide to SAR filing requirements and deadlines for thresholds and timing.
Where AI Helps in SAR Preparation
AI is well suited to the assembly work around a SAR:
- Case summarization. Condensing alerts, KYC data, prior cases, and investigator notes into a structured case file.
- Transaction analysis. Summarizing flows, counterparties, velocity, and patterns across the review period.
- Narrative drafting. Producing a first draft organized around the five Ws and how.
- Consistency checks. Flagging missing elements, date ranges that do not match the data, or amounts that do not reconcile.
- Continuing activity reviews. Summarizing new activity since the prior filing for the 90-day review cycle.
Regulators have encouraged responsible innovation here. The federal banking agencies and FinCEN issued a joint statement in December 2018 encouraging innovative approaches to BSA/AML compliance, including new technology, while making clear that institutions remain responsible for meeting their obligations.
What Stays with Your BSA Officer
AI can assemble and draft. It cannot decide. These remain with people:
- The filing decision. Whether the activity is suspicious and reportable is the institution's determination. For fintechs operating under a sponsor bank, the fintech typically refers and the bank decides and files.
- The no-file decision. Deciding not to file requires a documented rationale. Examiners sample no-file decisions as closely as filings.
- Accuracy of every fact. The person approving the narrative is responsible for its contents. Each name, date, amount, and account must be checked against source data.
- Timeliness. A SAR is generally due within 30 calendar days of initial detection of facts that may constitute a basis for filing, extendable to 60 days if no suspect is identified. Continuing activity filings typically follow a 90-day review with filing within 120 days of the prior report.
- Confidentiality. 31 CFR 1020.320(e) prohibits disclosing a SAR or information revealing its existence. Confirm who at an AI vendor can access SAR drafts and how access is logged.
The Specific Risks of AI-Drafted Narratives
Fabricated or misstated facts. Generative models can produce fluent text that includes details not present in the source data. In a SAR, a wrong account number or invented transaction misleads law enforcement and undermines the filing. Require that every factual statement in a draft link back to the case data it came from.
Boilerplate drift. AI drafts trained on past narratives can converge on generic language. Examiners criticize narratives that read the same across unrelated cases, and law enforcement finds them less useful. The "why" section must describe this customer's activity, not a typology in general.
Rubber-stamp review. If reviewers approve AI drafts at volume without checking them, the institution's SAR quality depends on the model. Track edit rates and review time. A reviewer who never changes anything is a signal, not a success.
Model risk. An AI that drafts SAR content informs a regulatory decision and falls under model risk expectations in SR 11-7. Validate it on past cases, document its limits, and monitor its error rate. See our guide to model risk management for banks.
A Defensible Workflow for AI-Assisted SARs
- Investigator builds the case. Alerts, KYC, transaction data, and notes are gathered in the case record.
- AI drafts. The draft cites the source record for each fact.
- Investigator edits. The investigator corrects, adds context, and confirms the "why."
- Reviewer decides. The BSA officer or delegate makes and documents the file or no-file decision.
- QA samples. A second reviewer samples filed and no-file cases monthly against FinCEN's narrative standard.
- Record retained. The case, AI draft, edits, decision, and filing confirmation are retained for five years with supporting documentation, as 31 CFR 1020.320(d) requires.
The record should let an examiner see what the AI produced, what a person changed, and who decided. That is what makes AI assistance defensible instead of a liability.
How Teams Keep SAR Work Owned and Evidenced
In Canarie, SAR-related obligations, including investigation deadlines, the 90-day continuing activity review, monthly QA sampling, and the annual independent test, run as recurring work with named owners, reviewers, and due dates. Evidence of each review attaches at completion, so when examiners test your SAR program, the timeliness record, QA results, and decision documentation already exist in one place. If you are evaluating products where AI agents prepare SARs, our comparison of agent-first and owned compliance execution and guide to AI compliance agents for fintechs cover the questions to ask.
Keep every SAR deadline owned and evidenced →
Frequently Asked Questions
Can AI write a SAR narrative?
AI can produce a strong first draft from case data, organized around FinCEN's who, what, when, where, why, and how. A qualified investigator must verify every fact against source data, add context, and confirm the reasoning, and the BSA officer or delegate must make the filing decision. The record should show both the AI draft and the human edits.
Is it allowed to use AI for SAR preparation?
Nothing in the BSA prohibits using technology to prepare SARs, and the agencies have encouraged innovative compliance approaches. The institution remains fully responsible for accuracy, timeliness, confidentiality, and the filing decision, and AI tools that inform those decisions should be validated and monitored under model risk management expectations.
What makes a SAR narrative insufficient?
Common deficiencies include missing elements of the five Ws and how, vague or boilerplate descriptions of why activity is suspicious, unexplained internal jargon, missing date ranges or amounts, and failure to describe how funds moved. FinCEN's narrative guidance and examiner feedback consistently emphasize clear, complete, chronological accounts.
Who is responsible for a SAR at a fintech with a sponsor bank?
Typically the fintech monitors activity and refers unusual activity to the bank under the program agreement, and the bank makes the filing decision and files. The agreement should define referral timelines and content. Both sides should keep records showing the referral, the bank's decision, and timeliness.
AI can take hours out of every SAR. It cannot take responsibility for one. Build the workflow so the speed is the model's and the judgment is visibly yours.