Why Do AML and KYC Compliance Programs Keep Failing, and How Do AI Agents Fix Them? Key Findings From Skan AI

Anti-money laundering (AML) and Know Your Customer (KYC) alert volumes rise with transaction volumes, which strains reviewer capacity, and banks must still meet the 30-day deadline for filing a suspicious activity report (SAR) as backlogs build. Banks use robotic process automation (RPA) bots, which replay fixed click paths, and rules engines, which fire alerts when a threshold trips, to add throughput, but analysts still need to interpret the case evidence.

The missing piece is context: what an alert means, where its evidence sits, and who holds the authority to decide.

Why do AML and KYC compliance programs keep failing?

AML and KYC compliance programs keep failing because the investigation work between detection and review still runs on manual evidence gathering. The failure has three structural causes:

• Alert noise: False positive rates can reach or exceed 90% of alerts (BCG, 2025), so analysts spend much of their review time clearing activity that was never suspicious.
• Headcount ceilings: Banks commonly assign up to 10 to 15 percent of full-time equivalents to KYC and AML alone (McKinsey, 2025). The TD Bank consent order shows what happens when staffing falls behind: “in 2022, the AML headcount decreased to less than 2020 totals, while the number of transaction monitoring alerts of potentially suspicious transactions continued to rise” (FinCEN, 2024).
• Documentation gaps: The FFIEC manual expects investigators to document their conclusions and the specific reason for filing or not filing (FFIEC). Examiners may follow a single alert through the entire process.

Rules engines and RPA address thresholds and keystrokes, while the heavier workload sits in investigation. Recalibrating thresholds often produces another spike in false alerts, and compliance teams spend most of their time collecting data and in “stare and compare” sessions instead of investigating (McKinsey, 2017).

RPA automates that collection with scripts, but the scripts fail when an application’s interface changes without notice: a robot can stop processing for days because it cannot find a moved or removed button (PwC, 2017).

What can AI agents for AML and KYC compliance actually take on?

AI agents for AML and KYC compliance can take on the assembly work of banking compliance investigations while analysts keep the decisions on SAR filing, escalation, and customer exits. Within that boundary, agents prepare the case and can handle:

• Alert triage and prioritization
• Evidence gathering across core banking, case management, screening and KYC systems
• Sanctions, PEP and adverse media review
• Event-driven KYC refresh
• SAR narrative and disposition drafting for analyst approval

How are AI agents different from RPA, rules engines, and copilots?

A rules engine fires when a threshold trips. An RPA bot replays a fixed click path. A copilot answers when an analyst asks it something. An agent works toward a goal, such as “disposition this alert with evidence.” It chooses which systems to query, adapts when a case varies, records its rationale, and stops at a defined escalation threshold.

Use four tests to identify rebadged RPA in a vendor demo:

• Give it an alert whose evidence sits in a system it was not configured for.
• Rename a field on one screen it reads.
• Present an exception that falls outside its documented path.
• Ask for the rationale record behind one alert it closed.

If any of these needs a script change, you are looking at RPA under a new name.

Where do AI agents fit across the AML and KYC lifecycle?

The fit differs by workflow, because each stage has different evidence sources and decision rights.

Sanctions, PEP, and adverse media alert review

An agent can review sanctions, politically exposed person (PEP), and adverse media (negative news coverage of a customer) alerts by comparing date of birth, nationality, and ownership data against the list entry, clearing name collisions (different people who share a name) that the data rules out, and writing the reason for each clearance.

The Wolfsberg Group states that a screening hit does not by itself indicate sanctions risk (Wolfsberg Group). In a controlled Federal Reserve staff experiment, language models reduced sanctions-screening false positives by 92% relative to the best fuzzy-matching baseline (a name-matching method that tolerates spelling differences) (Federal Reserve, 2025).

Transaction monitoring triage and investigations

For each alert, the agent pulls account history, the KYC profile, counterparties, and prior alerts, then summarizes the activity against the customer’s expected behavior. At one retail bank, generative AI (GenAI) agents produced investigative summaries for compliance officers, and case resolution time improved (BCG, 2025).

KYC onboarding, UBO mapping, and perpetual refresh

The Financial Crimes Enforcement Network (FinCEN) states that there is no categorical requirement to update customer information on a continuous or periodic schedule; updating is risk-based and follows from normal monitoring (FinCEN).

Agents can:

• Trace ownership chains, the links between a company and the ultimate beneficial owners (UBOs) who own or control it, against the 25% ownership and control prongs (eCFR, 2026).
• Flag discrepancies against registry data.
• Open a refresh when the bank learns of facts that reasonably call previously obtained beneficial ownership information into question, the trigger FinCEN’s 2026 exceptive relief sets for re-collecting that information (FinCEN, 2026).

In the EU, the draft ongoing-monitoring guidelines of the Anti-Money Laundering Authority (AMLA) expect both periodic and event-driven reviews (AMLA, 2026).

SAR narratives and disposition drafting

The agent assembles an evidence packet with a link to each source record and drafts a narrative that covers who, what, when, where, why, and how, the structure set out in the FFIEC manual’s Appendix L (FFIEC). The analyst edits the draft, decides and files.

Which decisions stay with human analysts?

Final filing, escalation, and exit decisions stay with people. The FFIEC expects the SAR decision-maker, whether an individual or a committee, to hold authority over the final filing decision (FFIEC). Everest Group places SAR filing, autonomous alert closure, onboarding denial, and sanctions blocking among the more sensitive workflows that need stronger controls (Everest Group, 2026). A workable decision-rights map looks like this:

Your risk appetite decides whether low-risk clearance belongs in the execute column, and sampled QA should come with it.

What governance, model risk, and audit controls will regulators expect?

Regulators apply the frameworks they already use to agents: an effective decision process, retained evidence, and effective oversight.

• U.S. model risk: SR 26-2 superseded SR 11-7 and the 2021 BSA/AML model risk statement on April 17, 2026. It excludes generative and agentic AI from its scope and leaves governance of those tools to the bank’s own risk management practices (Federal Reserve, 2026). The OCC describes bank use of these tools as limited to specific cases “with guardrails and human-in-the-loop accountability” (OCC, 2026).
• FinCEN: A SAR is due within 30 calendar days of initial detection, with a 60-day cap where no suspect is identified, and banks must keep supporting documentation for five years and provide it on request (eCFR, 2026). FinCEN’s October 2025 FAQ says there is no requirement to document a decision not to file (FinCEN, 2025). The FFIEC manual still advises documenting reasons, so record both outcomes.
• FATF: FATF lists explainability, human oversight, privacy and cybersecurity as safeguards for new AML/CFT tools (FATF, 2021).
• EU AI Act: High-risk systems need automatic event logging and human oversight, including the ability to override outputs (EU AI Act Art. 12, Art. 14). Stand-alone high-risk obligations apply from 2 December 2027 (EUR-Lex, 2026).
• AMLR and AMLA: The AMLR applies from 10 July 2027 (EUR-Lex, 2024), and the Anti-Money Laundering Authority begins direct supervision of up to 40 cross-border institutions, selected in 2027, from 2028 (AMLA, 2026).

Log every agent action with its inputs, sources and rationale, and keep the record replayable for examiners. Restrict PII access by role, and retain agent records alongside SAR supporting documentation, which carries a five-year rule. Require every drafted statement to link to a source record, and route unsupported claims to the analyst.

How do AI agents connect to case management, screening, and core banking?

Agents read from case management, screening, core banking, KYC, and registry systems, through application programming interfaces (APIs, the standard connections that let one system request data from another) where they exist and through application screens where they do not. They write drafts back to case management.

The harder problem is procedural: standard operating procedures (SOPs, the written steps for a task) can omit the tacit ways analysts handle exceptions, and event logs, the timestamped records systems keep of what happened, miss steps that cross applications without a shared data layer (a common store that all the applications write to). A Context Graph of Work, a map of how work actually moves across applications drawn from real process activity, is designed to close that gap.

Skan AI is an agentic AI platform for regulated enterprises that builds a Context Graph of Work from real processes, so banks can ground agent decisions and execution in how work runs.

Before setting an agent’s scope, a bank’s team needs to see how work actually moves across its applications, and the bank’s own observed process activity shows that movement. Skan AI Virtual Assistant (Skan VA) observes work across mainframes (large central computers), virtual desktop infrastructure (VDI, desktops hosted on a central server), legacy desktop, and web applications under a privacy by design approach, meaning privacy protections are built in from the start, and Skan AI combines end-to-end process depth with capture that needs zero integrations, APIs, or logs to see the work.

Skan AI PRISM applies pixel-level redaction, which masks sensitive content in screen images, on-premises, as part of its privacy by design approach, before data leaves the customer environment, and the Skan VA is designed not to capture email content or keystrokes. Raw screen data stays inside the customer perimeter. Skan AI holds SOC 2 Type 2 attestation and ISO 27001 and ISO 42001 certifications, and is compliant with the Health Insurance Portability and Accountability Act (HIPAA), the General Data Protection Regulation (GDPR), and the California Consumer Privacy Act (CCPA).

The products run in sequence on one Context Graph of Work:

• Skan AI Blueprint identifies candidate areas through a lighter opportunity scan.
• Skan Intelligence observes those identified areas in depth.
• Skan AI Agents execute under Agentic Operating Procedures (AOPs).

AOPs encode the workflow steps, decision rules, and escalation thresholds an agent must follow. Within the agentic AI platform, the software the agents run on, Agent Studio is the configuration tool that sets each agent’s scope, permissions, and escalation paths, and a governance layer records its actions, decisions, and rationales for case-level review. A neurosymbolic architecture, which pairs learned models with explicit rules, keeps an explainable reasoning path from observed activity to each recommendation.

Enterprises choose where the platform runs: on-premises (on the customer’s own servers) inside the customer firewall, in a private cloud (cloud infrastructure dedicated to one organization), or Skan AI hosted. Skan AI also supports readiness for the EU AI Act, the European Union’s regulation on artificial intelligence.

How much impact can AI agents deliver, and how do you build the case?

Build the case on your own measured baseline. ComplyAdvantage’s 2026 survey found 89% of institutions take up to 30 minutes to resolve a single transaction-monitoring alert (ComplyAdvantage, 2026). The model below applies illustrative inputs:

The monthly alert volume, false-positive share and before-and-after handling times are illustrative assumptions, not measured results. The analyst-hour totals are calculations from those assumptions.

On build versus buy, the interagency third-party guidance is clear that using a vendor does not reduce the bank’s own responsibility for compliance (Federal Reserve, 2023). Phase the rollout so decision rights expand only as evidence accumulates (Deloitte, 2026):

1. Measure the current baseline per alert type.
2. Run evidence gathering and triage in draft-only mode.
3. Allow low-risk clearance with sampled QA.
4. Add SAR drafting, with filing kept human-only.

Where can you get the full AML and KYC framework?

The decision guide explains how far to extend agent authority as evidence builds, and the AML and KYC agentic compliance whitepaper sets out the capacity model and decision-rights approach in full. To learn more, see the analysis of why AML compliance programs keep failing.

If you’re ready to see these recommendations applied to your own alert workflows, you can request a Skan AI demo and see how complex context enables reliable, compliant agent execution.