We use cookies to understand how this site is used. Privacy policy

    Skip to main content
    Kriv AI

    Finance / AML Compliance

    Perpetual KYC and Continuous AML Monitoring AI

    Periodic KYC review runs on a fixed schedule set by a customer's risk rating. Perpetual KYC and continuous AML monitoring AI re-score that risk as events happen, between reviews, which raises a different governance question: who validates the scoring and triage models making those calls, and how often.

    Perpetual KYC and continuous monitoring AI replace the periodic customer-file refresh with continual, risk-scored surveillance: transaction behavior, ownership changes, sanctions hits, and adverse media are re-evaluated on an ongoing basis rather than at fixed intervals, so a customer's risk profile updates as new information appears instead of waiting for the next scheduled review.

    perpetual kyc continuous

    What Perpetual KYC and Continuous Monitoring AI Actually Means

    Perpetual KYC and continuous monitoring AI replace the periodic customer-file refresh with continual, risk-scored surveillance: transaction behavior, ownership changes, sanctions hits, and adverse media are re-evaluated on an ongoing basis rather than at fixed intervals, so a customer's risk profile updates as new information appears instead of waiting for the next scheduled review.

    Traditional KYC operates on a review cycle set by the bank's own risk-based approach, low-risk customers on a longer cycle, high-risk customers more frequently, under FinCEN's customer due diligence rule, which requires covered financial institutions to establish risk-based procedures for ongoing due diligence, including monitoring to identify and report suspicious activity and maintaining and updating customer information on a risk basis.

    Perpetual or continuous KYC does not replace that regulatory baseline. It changes the monitoring cadence underneath it, using models that re-score a customer's risk on every relevant event, a large or unusual wire, a beneficial-ownership change, a new sanctions-list hit, negative media, instead of only at the scheduled review date.

    The AI component typically sits in three places: entity resolution and screening, matching a customer against sanctions and adverse-media sources on an ongoing basis rather than in a nightly batch; behavioral scoring, transaction monitoring that adapts its baseline per customer instead of applying one static rule threshold to everyone; and alert prioritization, ranking the resulting alerts by estimated severity so investigators work the highest-risk items first.

    None of that removes the compliance officer from the loop. FinCEN's rule and FATF's risk-based approach guidance both still require a documented, defensible rationale for why a customer's risk rating changed, whether a rule, a model, or a human analyst made the call.

    aml alert triage

    Why AML Alert-Triage AI Keeps Triggering False Positives

    Most false-positive floods trace back to the same root cause: an AI scoring or triage layer gets added on top of legacy rule thresholds without recalibrating either one, so the model inherits every over-broad rule and adds its own miscalibration on top of it.

    Rules-based transaction monitoring already over-alerts by design. Thresholds get set conservatively so a true positive is not missed, which examiners generally tolerate, and that produces a high volume of alerts that turn out not to be suspicious once an analyst looks at them.

    When AI is layered in to prioritize or auto-close a share of those alerts, teams often train the scoring model on historical analyst dispositions rather than on confirmed suspicious-activity outcomes. The model then gets efficient at reproducing whatever disposition pattern the analysts already had, closing the alerts investigators would have closed anyway, without necessarily reducing the false positives that were the problem in the first place.

    Feature drift compounds it. A model tuned against last year's transaction patterns for a given customer segment does not automatically hold once payment rails, customer mix, or typologies shift. If the retraining or recalibration cadence lags that drift, alert precision degrades quietly, sometimes for months, unless someone is monitoring for drift between formal validation cycles rather than only at them.

    The fix is rarely more AI. It is re-scoping what the AI gets measured against: calibrating alert thresholds and model confidence to case-disposition and suspicious-activity-report outcomes rather than analyst throughput, and building a feedback loop where an investigator's decision on an individual alert actually retrains or recalibrates the model instead of just closing the ticket.

    materiality based model

    Materiality-Based Model Validation for Continuous Monitoring and Scoring Models

    Materiality-based model validation applies the Federal Reserve's model risk management discipline, originally set out in SR 11-7 and revised in April 2026 by SR 26-2, conceptual soundness review, independent validation, and ongoing monitoring, in proportion to what a given AML model actually decides, rather than as one checklist applied identically to every model in the inventory.

    AML scoring, screening, and alert-triage models are quantitative by design: they take transaction and customer data as input and produce a quantitative score or flag used to inform a decision, which keeps them squarely inside SR 26-2's narrower 2026 definition of a model, unlike a generative or agentic tool that produces text rather than a measurable prediction. The validation discipline below still runs under that guidance's current name, SR 26-2, even where the reasoning traces back to the original SR 11-7 framework.

    SR 11-7 (and its SR 26-2 update) never mandated identical validation depth for every model. It directs validation effort toward the models whose failure would matter most, judged by the materiality of the decision the model drives, which is exactly the tiering question a continuous monitoring program needs to answer for its own scoring and triage models.

    Applied here, a model that only re-ranks the order analysts work an alert queue carries a different validation weight than a model that auto-closes alerts without human review, or one that recommends whether to file a suspicious activity report. The second and third categories warrant the deeper conceptual-soundness review, independent challenge, and outcomes analysis SR 11-7 describes. The first can sit on a lighter, more frequent cycle focused mainly on drift and stability.

    Continuous monitoring also creates a validation obligation periodic KYC never had. A model that re-scores risk on every transaction is making a steady stream of small decisions between scheduled validation dates, so ongoing performance monitoring, tracking alert volume, precision, and disposition rates against a documented baseline, has to run between formal validations, not just at them.

    Materiality tiering also determines who signs off. A queue-ranking model can reasonably be owned and monitored inside the AML operations team; a model whose output can lead to an account closure or a filed report needs the independent validation function SR 11-7 assigns to a group separate from the people who built or use the model.

    continuous monitoring progra

    What a Continuous Monitoring Program Needs, Architecturally

    Most stalled continuous-monitoring efforts are architecture problems before they are model problems: the scoring layer is only as continuous as the data and workflow underneath it.

    1. 1. A unified customer and transaction data layer

      Transaction history, KYC profile attributes, beneficial ownership data, and screening results need to sit somewhere the scoring model can query close to real time. Siloed data across separate onboarding, screening, and transaction-monitoring systems is the most common reason 'continuous' monitoring ends up running on a batch delay instead.

    2. 2. Event-triggered re-scoring, not only scheduled re-scoring

      The model needs to re-evaluate a customer's risk score when a defined event fires, an unusual transaction, an ownership change, a new sanctions or adverse-media hit, in addition to whatever fixed review cycle the bank's risk-based approach already sets.

    3. 3. An alert triage and prioritization layer

      A ranking step that orders the resulting alerts by estimated severity, so investigator time goes to the highest-risk items first instead of working a flat queue in arrival order.

    4. 4. Explainability and an audit trail per decision

      Every score and every alert disposition needs a recorded rationale an examiner can review afterward: which inputs drove the score, which rule or model version produced it, and who reviewed or overrode it.

    5. 5. A feedback loop from investigator decisions back into the model

      Case dispositions and confirmed suspicious-activity outcomes need to feed back into retraining or recalibration on a defined cadence, or the model drifts toward reproducing whatever bias existed in its original training alerts.

    governance controls examiner

    Governance Controls Examiners Expect Around These Models

    A continuous monitoring program usually runs on several distinct models at once, screening, behavioral scoring, and triage, and each of the controls below applies separately to each one.

    ControlWhat It CoversWhy It Matters Here
    Model inventory and ownershipEvery scoring, screening, and triage model logged with a named owner and risk tierAn incomplete inventory of the models actually in production is one of the most common examination findings
    Independent validationReview by a function separate from the model's developers and users, scaled to the materiality tierApplies SR 11-7's core validation elements to systems a bank may not have originally classified as 'models'
    Ongoing performance monitoringAlert volume, precision, and disposition-rate tracking against a documented baseline, between formal validationsContinuous scoring makes decisions every day between validation dates; drift can go unnoticed without this
    Change managementDocumented approval and a re-validation trigger whenever thresholds, features, or model versions changeA threshold tweak made to cut alert volume is itself a model change and needs the same scrutiny as a retrain

    kriv ai approaches

    How Kriv AI Approaches a Perpetual KYC or Continuous Monitoring Engagement

    The starting point is always an inventory: what is actually scoring, screening, or triaging customers today, who owns each piece, and which of those pieces were ever formally validated as a model in the SR 11-7 sense.

    From there the practice maps each model or scoring component to a materiality tier based on what it decides, queue ranking, auto-disposition, or a filing recommendation, and scopes validation depth and cadence to that tier rather than applying one template to everything.

    For monitoring and triage models specifically, that includes reviewing how alert thresholds and model confidence were calibrated, against confirmed outcomes or against analyst throughput, since the calibration target is usually the real driver of a false-positive problem, not the underlying algorithm.

    The practice also reviews the data architecture supporting continuous re-scoring (whether customer, transaction, and screening data actually reach the model close to real time), the audit trail each decision leaves for an examiner, and whether a feedback loop exists between investigator case outcomes and model retraining.

    The output is a governance structure the compliance and model risk teams can run themselves going forward: an inventory, a tiering methodology, a validation and ongoing-monitoring schedule aligned to SR 11-7 and to NIST's AI Risk Management Framework, and documentation an examiner can follow without a live translator in the room. No engagement here promises a specific reduction in false positives or examination outcome; the work is the governance structure, not a number.

    Straight answers

    Frequently asked questions about Perpetual KYC and Continuous AML Monitoring AI

    What is perpetual KYC and how is it different from periodic KYC review?

    Periodic KYC reviews a customer's file on a fixed schedule set by their risk rating. Perpetual KYC keeps that same regulatory obligation in place but re-scores the customer's risk continuously, on transaction behavior, ownership changes, sanctions hits, and adverse media, so a change in risk can surface between scheduled reviews instead of waiting for the next one.

    Why does AI-driven AML alert triage still produce a lot of false positives?

    Most of the time the AI was layered onto legacy rule thresholds that were never recalibrated, and was trained on historical analyst dispositions rather than confirmed suspicious-activity outcomes. It gets efficient at reproducing the existing disposition pattern rather than reducing the false-positive rate itself, and feature drift makes it worse over time if nobody is monitoring for it between validation cycles.

    What is materiality-based model validation for AML monitoring models?

    It means scoping validation depth, independent challenge, and how often a model gets re-validated to the materiality of what that model decides. A model that only re-ranks an alert queue gets a lighter, more frequent review; a model that can auto-close alerts or influence a filing decision gets the fuller SR 11-7 validation treatment, including independent review by a group separate from its developers and users.

    Does perpetual KYC replace the scheduled customer due diligence review cycle?

    No. FinCEN's customer due diligence rule still requires a risk-based review cycle and documented, updated customer information. Continuous monitoring runs underneath that obligation, catching risk-relevant events between reviews, it does not remove the requirement for periodic review itself.

    Which regulations and frameworks actually govern continuous AML monitoring models?

    The core requirement to monitor and update customer information on a risk basis comes from FinCEN's customer due diligence rule. Model governance for the scoring, screening, and triage models involved draws on the Federal Reserve's SR 11-7 guidance (revised in April 2026 by SR 26-2), and NIST's AI Risk Management Framework is a common structure for organizing inventory, measurement, and monitoring practices around those models.

    How does SR 11-7 apply to AML models that were never formally classified as 'models'?

    SR 11-7 defines a model broadly as any method that processes inputs into quantitative outputs used to inform a decision, which covers most transaction-scoring, screening-match, and alert-triage logic even when it was built as an operational tool rather than labeled a model. The practical implication is that an inventory built only from a bank's formal model list will miss several systems examiners would still expect to see governed.

    Talk to the team that would do the work

    Bring your requirements to a working session with the person who'll actually deliver.

    Book a Discovery Call