AML's Real Bottleneck Is the Narrative

A compliance team at a mid-size bank files 300 SARs a month. Each narrative takes roughly four hours to write. That's 1,200 hours of writing, every month, from a team of maybe 15 people.

The monitoring system that generated those alerts took milliseconds per transaction, and the rules engine that flagged them ran in batch overnight. The bottleneck was never detection.

Where the hours go

Transaction monitoring catches the signal. A rule fires, or a model scores a cluster of transactions above a threshold, and an alert lands in the queue. A Level 1 analyst triages it in 20 to 40 minutes. If it looks real, a Level 2 investigator pulls records, reviews account history, and decides whether to file.

Then comes the SAR.

FinCEN's SAR form has dozens of structured fields for account numbers, dates, and dollar amounts. But Part V is a free-text narrative, and that narrative is the actual product. It has to describe who did what, when, how the activity was detected, what makes it suspicious, and what the institution did about it. A FinCEN examiner who reads it should understand the case without logging into the bank's systems.

Writing that narrative from scratch means pulling data from the case management system, the core banking platform, the CDD files, and sometimes external databases. The analyst synthesizes all of it into prose that's specific enough to be useful and careful enough to survive regulatory examination.

4.8 million SARs were filed in fiscal year 2025. Each one has a narrative someone wrote by hand.

Why detection got all the investment

The AML technology market spent the last decade on the front of the pipeline. Better rules engines, ML-based scoring, network analysis, behavioral analytics. The pitch was always the same: reduce false positives, catch more real activity, automate the alert.

It worked, partially. Some large banks have migrated significant portions of their rule-based monitoring to machine learning, and vendors claim 30 to 70 percent reductions in false positive rates. But even at a 50% reduction, you've cut a 95% false positive rate to a 90% false positive rate. The alert queue is still enormous.

And nothing on the detection side touches the narrative. A smarter alert still lands on an analyst's desk, and that analyst still opens a blank text field and starts typing.

Meanwhile, 77% of financial institutions report difficulty hiring and retaining AML analysts. The people writing the narratives are the scarcest resource in the pipeline, and their most time-consuming task is the one that got the least automation.

What happens when you point an LLM at a SAR

The obvious approach is to feed transaction data into a language model and ask it to write the narrative. Hummingbird, which builds SAR workflow software, tried exactly this with general-purpose LLMs.

The model fabricated transaction details, inventing account numbers that didn't exist in the source data and generating plausible-sounding narratives about activity that never happened.

This is the wrong kind of wrong. A SAR is a legal filing with a federal agency. A human analyst who writes a sloppy narrative gets a coaching conversation. A system that invents facts in a federal filing creates regulatory liability that compliance officers won't accept.

General-purpose LLMs produce plausible text, but SAR narratives demand provably grounded text where every fact traces back to a specific record in the institution's systems.

Structured generation instead of freeform drafting

The systems that work in production don't ask the LLM to write a narrative from scratch. They decompose the SAR into sections, extract verified fields from the case file, and ask the LLM to generate prose for each section using only the provided facts.

The difference is the constraint surface. Instead of "write a SAR about this customer," the prompt is closer to "given these 14 transactions between these dates, involving these accounts, describe the pattern of activity." The LLM's job shifts from creation to articulation, turning structured data into readable prose without adding information.

LLMs handle variation in sentence structure, connect facts with appropriate causal language, and produce prose that reads like a human wrote it. The constraint is that every factual claim in the output must map to a field in the input, with no synthesis, no inference, and no "based on similar cases."

Research on co-investigator AI systems reports 70% narrative completeness and 61% time savings using this structured approach. The output is a draft that an analyst reviews and edits in roughly 45 minutes instead of four hours.

What the analyst still does

Even with structured generation, the analyst's role shifts rather than disappearing.

The hardest part of a SAR narrative is judgment: deciding which transactions are relevant, what pattern they form, and why it's suspicious rather than unusual. "Customer sent five wires to the same beneficiary in three days" is a fact. "Customer appears to be structuring wire transfers to avoid reporting thresholds" is an inference. The LLM handles the facts, but the inference that connects them to suspicious activity remains the analyst's call.

Review also catches what structured generation misses. Context that lives in relationship manager notes, prior SARs on the same customer, information from law enforcement requests. These inputs don't always make it into the case management system's structured fields, which means they don't make it into the LLM's prompt.

The production workflow looks less like "AI writes the SAR" and more like "AI drafts the SAR, analyst reviews and adds judgment." The four-hour task becomes a 45-minute task. The analyst writes four times as many SARs in the same shift, or spends the recovered hours on the investigations that actually need human attention.

FinCEN's effectiveness standard changes the math

In April 2026, FinCEN proposed a rule that would re-anchor AML programs around effectiveness rather than procedural compliance. Instead of proving you followed the right steps, you'd need to demonstrate your program actually works, that it detects and reports illicit activity at a rate that justifies the investment.

Under a procedural standard, automating SAR narratives is a risk. What if the examiner doesn't like that a machine wrote part of it? What if the narrative quality doesn't match a human analyst's output? The safe choice is to keep humans writing every word.

Under an effectiveness standard, the math flips. If structured generation lets your team file more complete, more consistent SARs with fewer errors and faster turnaround, that's a measurable improvement in program effectiveness. The regulatory question shifts from "who wrote this" to "does this work better."

The proposed rule hasn't been finalized. But the direction is clear enough that compliance teams are building toward it. The institutions that figure out human-plus-LLM workflows for SAR narratives now will have the operational data to demonstrate effectiveness when the rule lands.

The narrative is the product

Most of the AML technology conversation focuses on the alert. Better detection, fewer false positives, smarter scoring. Those matter. But 4.8 million SARs a year means 4.8 million narratives, each one a multi-hour writing task performed by an analyst who could be investigating the next case instead.

LLMs won't replace the analyst, but they'll give them back the hours they currently spend typing.