Every monitoring rule is a trade. Lower a threshold and the queue grows; raise it and something real may slip under it. Most banks make that trade on instinct and on the volume of the queue. The better input is outcome: of the alerts a rule raised, how many became suspicious activity reports, and why were the rest closed? And before a change goes live, which alerts that people escalated would it have dropped?

In short

  • Yield per rule, from the outcomes. InsightAML counts, for each rule, the alerts it raised and what became of them, this period and over the legacy system’s two years, and the alert-to-SAR rate of each.
  • Close reasons say where the noise comes from. On the demo bank’s legacy history, round-amount alerts became SARs 3.1% of the time (16 of 520); 453 were closed, most as own-account transfers, payroll or rent.
  • The proposal shows its cost. The tuning analyst proposes raising Round amounts min_count from 5 to 7, and says it would drop all 10 of this period’s round-amount alerts, including one that was escalated.
  • Only a lead tunes. A change needs a reason, becomes a new version of the rule, and keeps its history. No agent can apply it.
  • Every parameter is illustrative. The defaults are starting points: set them from your own risk assessment and tune them on your own outcomes.

Why tuning on volume goes wrong

A rule that raises too many alerts gets its threshold nudged until the queue is manageable. Nobody checks what the nudge removed, because the removed alerts are never raised and so never seen. The opposite failure is just as common: a rule nobody dares touch because no one can say what changing it would cost.

Industry
data

US banking agencies expect the models and systems that support BSA/AML compliance, including transaction monitoring, to be managed with sound model risk practices where they meet the definition of a model, scaled to the bank’s risk. In practice examiners ask how thresholds were set, when they were last reviewed, what the review looked at, and who approved the change.

Sources: Interagency Statement on Model Risk Management for Bank Systems Supporting Bank Secrecy Act/Anti-Money Laundering Compliance (April 2021); Supervisory Guidance on Model Risk Management, SR 11-7 / OCC Bulletin 2011-12; FFIEC BSA/AML Examination Manual. Check the guidance that applies in your jurisdiction.

The demo bank’s numbers

Northmere Bank is fictional; its transactions are generated and its legacy alert history (two years of a fictional previous monitoring system) is too. The figures below are what InsightAML’s tuning page counts from that data as of 20 November 2026. They describe the demo, not any real bank or benchmark.

RuleLegacy alertsLegacy SARsAlert-to-SARMost common close reason (legacy)
Structuring1603421.2%Single deposit explained by a documented vehicle sale (64)
Rapid movement / pass-through210188.6%Documented property or vehicle purchase (93)
High-risk geography1402920.7%Family support consistent with the declared profile (55)
Round amounts520163.1%Transfers between own accounts (173)
Velocity against profile613376.0%Annual bonus or one-off documented payment (193)
Funnel account601830.0%Tenant rent collection (20)
Dormant reactivation901213.3%Customer returned from abroad; explained to the RM (41)
Cash-intensive pattern230219.1%Seasonal takings (70)
Trade pricing anomaly451124.4%Price includes freight and duty (17)
Watchlist-adjacent counterparty25624.0%Link not confirmed: different entity on review (10)
The rule tuning page as AML lead: each of the ten rules with its parameters, alerts this period, alert-to-SAR this period and in the legacy system, the most common close reasons, and a Tune button, under a note that parameters are illustrative defaults
Rule tuning. Each rule’s parameters, this period’s alerts and decisions, the legacy alert-to-SAR rate and the close reasons, with a Tune button for a lead.

This period’s column is small on purpose: in the demo, 86 alerts were raised on the first run, and only a handful of cases have been decided. A rate of 50% from two decided alerts is not a yield; the page shows the count of decided alerts beside every rate so nobody mistakes one for the other.

Round amounts: where 453 closures came from

33%31%23%10%3%Legacy round-amount alerts, 520 in allTransfers between own accountsPayroll fundingRent or loan payments in round amountsEscalated, no SARSAR filed

Outcomes of the fictional legacy system’s round-amount alerts over two years.

The pattern is familiar: businesses move round sums between their own accounts and fund payroll in round transfers, and people pay rent and loans in round amounts. In this period, one of the ten round-amount cases, Ironwood CPA Partners, was closed for exactly that reason: six transfers totalling $232,000.00, all between its own accounts or to its payroll processor.

The proposal, and what it would cost

The tuning analyst read all ten rules (“10 rules read; 1 rule below 5% alert-to-SAR over the legacy history”) and prepared one proposal for a lead. It reruns the rule over this period’s data with the change, so the effect is counted, not estimated:

Raise Round amounts min_count from 5 to 7

Legacy system: 16 SARs from 520 alerts (3.1%). Most common close reason: Transfers between own accounts (173). Most common close reason: Payroll funding (162). This period: 10 alerts now; 0 with the change (10 fewer). It would drop 1 alert that was escalated: ALR-2026-000047 (Lakepoint Strategic Consulting LLC).

That last line is the one that matters. Lakepoint Strategic Consulting LLC is a layering case: four large credits paid out again within three days, wires with a high-risk jurisdiction, credits 15.2 times the declared monthly figure, and six round transfers totalling $566,000.00. It was escalated and its SAR is in draft. The round-amount alert was one of four on the case, so raising the count to 7 would not have hidden the case. It would have removed one of the facts the narrative cites. A lead can weigh that; a lead who never sees it can’t.

apply with a reason

reject

Outcomes
this period and legacy

Tuning analyst
reads yield and close reasons

Reruns the rule
with the change

Proposal
alerts dropped, escalated alerts dropped

AML lead

New rule version
history kept

Rule unchanged

The agent prepares the evidence. The lead applies or rejects it.

The agent activity page, with the tuning analyst's step on the monitoring rules at the top: 10 rules read; 1 rule below 5% alert-to-SAR over the legacy history, prepared, at L1
The tuning analyst’s run. At the top of the activity list: ten rules read, one below 5% alert-to-SAR on the legacy history, a proposal prepared at L1.

Who may change a rule

Tuning is a lead’s call. InsightAML refuses a change from an agent, refuses a change without a reason (“Say why the rule is tuned; the reason stays with its version”), refuses unknown or non-positive parameters, and saves every change as a new version with who, when, the parameters and the note. Every alert records the rule version that raised it, so a later reader can tell which threshold produced which alert.

A tuning reviewOwners in the tag
1
Read yield beside the decided countAML lead

A rate from two decisions is an anecdote; the legacy history gives the larger sample.

2
Read the close reasonsAML lead

Are the closures one explainable population (payroll, own accounts) or many?

3
Ask for the rerunTuning analyst

The proposal counts this period’s alerts with and without the change, and names any escalated alert it would drop.

4
Decide, with a reasonAML lead

Apply as a new version, or reject. The reason is part of the version.

5
Watch the next periodAML lead

Compare the new version’s alerts and outcomes with the old one’s.

TODAY The queue is too long, so the threshold went up

Nobody recorded which alerts the change removed, or who agreed to it. The next examiner asks how the threshold was set.

WITH INSIGHTAML The yield was low, the close reasons were explainable, and the cost was named

The proposal says what it drops, including the escalated alert. A lead applied or rejected it with a reason, and the version history keeps both.

Tune on what alerts became, and look at what a change would have missed.

The demo’s defaults are illustrative. The method is the point: outcomes in, a counted trade-off out, a lead’s reason on the version.

Keep reading