What Makes a SAR Narrative Defensible
The narrative is the heart of a Suspicious Activity Report — the free-text section where the institution explains, in FinCEN’s framing, the who, what, when, where, why, and how of the suspicious activity. It is the only part of the filing law enforcement reads as prose, and the part BSA examiners use to judge whether your monitoring program actually works. As LLMs take over narrative drafting — and they are, across the industry — the failure mode shifts from narratives that are thin to narratives that are fluent but ungrounded. A defensible narrative has three properties: every fact traces to the case file, the suspicion is articulated as reasoning rather than asserted, and the account is complete against the underlying alerts. None of the three is a writing skill. All three are architecture.
Why the narrative carries the filing
A SAR is mostly structured fields — subjects, amounts, dates, activity codes. Software populates those competently. The narrative is different in kind. It is where the institution converts data into an account of why this activity is suspicious: the pattern that emerged, the expected behavior it deviated from, the explanations that were considered and ruled out, the connection between the transactions and the suspicion. FinCEN’s longstanding guidance frames it as the five W’s and how, and is blunt that a SAR with an inadequate narrative is of limited use — law enforcement triages on it, and a filing whose narrative doesn’t establish the suspicion is functionally a null filing regardless of how accurate its fields are.
Examiners read narratives for a second purpose: as a window into the program. A narrative that misstates dates, contradicts its own transaction summary, or asserts suspicion without articulating a basis tells the examiner something about the investigation process that produced it. Weak narratives are treated as symptoms. Systemically weak narratives become findings about the BSA/AML program itself — and timeliness (generally 30 days from initial detection, 60 without an identified subject, with continuing-activity reviews on their own cycle) means the pressure that produces weak narratives never lets up.
The three properties
Grounding. Every factual assertion in the narrative — every amount, date, counterparty, and pattern claim — must be traceable to something in the case file: transaction records, KYC data, alert output, prior SARs. This sounds obvious and fails constantly, because narratives are written by humans (or now, models) summarizing large files under deadline, and summarization drifts. A narrative that says “fourteen structured deposits between January and March” when the file supports twelve between February and April is not a typo problem. It is a filing signed under the institution’s name that does not match the institution’s own records — the exact discrepancy an examiner or defense attorney finds first.
Articulated suspicion. “The activity appeared consistent with structuring” is a label, not a narrative. A defensible narrative walks the reasoning: what the expected activity for this customer was, what was observed, why the deviation is not plausibly explained by the customer’s profile or stated purpose, and how the pattern maps to the typology. This is the part that establishes the institution exercised judgment rather than pattern-matched — and it is precisely the part that a template, or a model prompted to “write a SAR narrative,” fakes most convincingly. Fluent assertion reads like reasoning to a hurried reviewer. It does not read like reasoning to an examiner comparing the narrative against the alert history.
Completeness. The narrative must account for the material activity in the review — including what cuts against the suspicion. Omitting the three months of normal behavior between suspicious clusters, or the plausible business explanation the customer gave, doesn’t strengthen a filing; it creates the impression of a curated story when the full record surfaces later. Completeness is also the property that quietly dies at volume, because checking a draft against the full case file is the most expensive part of quality review.
Enter the drafting models
LLMs are already writing first-draft narratives at scale, and the appeal is real: the work is summarization under a format convention, the volume is high, and investigator time is the binding constraint. But notice what an ungoverned drafting model optimizes: it produces the most plausible narrative conditioned on the case summary it was given. Plausibility is anti-correlated with exactly the three properties above. The model will fill evidentiary gaps with typology boilerplate (grounding failure), assert suspicion in confident compliance prose (articulation failure), and smooth over the inconvenient parts of the record because they weren’t in the prompt (completeness failure). The result is a new and worse failure mode: narratives that are uniformly well-written and intermittently untrue. Thin narratives announced themselves. Fluent ungrounded ones pass review.
The fix is not banning drafting models — the volume pressure is real and the capability is genuinely useful. The fix is refusing to let the narrative be a single-pass generation. In a governed drafting workflow, the narrative is assembled by a reasoning graph, not emitted by a prompt: extraction nodes pull facts from the case file with references; a claim-binding gate blocks any sentence whose factual content lacks a source in the file; a reasoning node constructs the suspicion argument explicitly — expected activity, observed activity, ruled-out explanations — rather than asking the model to sound suspicious; a completeness check reconciles the draft against the alert set and transaction record, flagging what the narrative fails to account for; and the investigator reviews a draft whose every claim is annotated to its evidence. The narrative gets faster and more grounded, and the institution keeps an artifact behind each filing: the trace showing how the narrative was built, which is the answer to the question examiners have started asking — not just “is this narrative adequate” but “how is it produced?”
A SAR narrative is testimony: the institution’s written account of why it suspected what it suspected. Testimony is judged on grounding, reasoning, and completeness — and a drafting process that can’t demonstrate all three under examination shouldn’t be signing the institution’s name.