All insights
Strategy2026-02-058 min read

Agentic Workflows for AML Alert Triage: Structuring the Pre-Investigation

How to deploy agentic systems to investigate transaction monitoring alerts, enforcing strict boundaries between automated data gathering and human disposition.

Transaction monitoring systems in commercial banking are calibrated for high sensitivity, resulting in alert queues where false positive rates routinely exceed ninety percent. Analysts spend the majority of their shift not making risk decisions, but manually retrieving account histories, counterparty profiles, and corporate registry documents across fragmented internal systems.

Agentic systems are frequently proposed as a mechanism to clear these queues autonomously. This is a fundamental misapplication of the technology in a regulated environment. The objective of an agent in Anti-Money Laundering (AML) operations is not to replace the investigator, but to automate the deterministic retrieval and synthesis of contextual data so the investigator can execute a defensible disposition in minutes rather than hours.

Structuring the Pre-Investigation Sequence

An effective alert triage workflow decomposes the investigator's manual routine into discrete, observable tool calls. When a transaction anomaly triggers an alert, the workflow instantiates an agent with a specific, bounded mandate: build the case file.

The agent navigates the bank's internal APIs to pull the customer's Know Your Customer (KYC) profile, recent transaction behavior, and historical alert dispositions. It must follow a strict procedural graph rather than deciding its own investigation strategy.

  • Querying the core banking system for the last ninety days of transaction velocity.
  • Retrieving the ultimate beneficial ownership (UBO) structure from the client onboarding repository.
  • Executing a deterministic search against approved adverse media databases.
  • Formatting the collected evidence into a standardized pre-investigation narrative.

The Human Disposition Gate

The boundary between agent behavior and human authority must be hardcoded into the workflow. The agent compiles the evidence and highlights specific risk indicators based on predefined rules, but it lacks the system permissions to change the state of the alert to closed or dismissed.

A human analyst reviews the synthesized narrative alongside explicit links to the underlying source documents. The analyst remains the sole actor authorized to execute a "clear" or "escalate to Suspicious Activity Report (SAR)" command. This isolation ensures that liability for the final regulatory decision rests entirely with the bank's designated personnel, while significantly increasing their throughput.

Controlling Hallucination in Adverse Media Screening

Adverse media screening is highly susceptible to entity resolution errors and hallucinated connections. If an agent is granted unbounded access to search the public internet, it will inevitably conflate individuals with similar names or misinterpret historical news articles.

To mitigate this, the agent must be restricted to a tightly controlled retrieval pattern utilizing only bank-approved, structured intelligence feeds. The agent acts strictly as a summarization layer over factual records retrieved by deterministic database queries.

  • Forcing the agent to cite specific article identifiers for any adverse finding.
  • Mandating exact-match parameters for corporate entities and strict proximity matching for individual names.
  • Configuring the workflow to fail open—escalating to a human immediately—if the search results are ambiguous or contradictory.

Measuring Investigator Throughput and Unit Economics

The justification for deploying an agentic triage system is measured strictly in unit economics. The cost of running the agent—which includes inference compute, orchestration overhead, and API calls to external data providers—must be substantially lower than the fully loaded cost of an analyst performing the identical retrieval tasks.

Financial institutions must establish baseline metrics for manual alert resolution times prior to deployment. Success is defined by a measurable reduction in the average handling time per alert without a corresponding increase in quality assurance failure rates during secondary review.

Ensuring Auditability for Regulatory Exams

Bank examiners expect to see the exact lineage of information that led to an alert disposition. When an agent is introduced into the process, the compliance burden extends to proving the agent executed its retrieval mandate accurately and without omission.

The system must persist a complete record of the agent's execution graph alongside the final analyst decision. This ensures that when a historical alert is sampled during an audit, the bank can reconstruct the exact data the agent fetched, the precise configuration utilized during synthesis, and the human sign-off that concluded the investigation.

Next step

Find your best AI workflow opportunity

A 30-minute discovery call: we look at two or three of your current processes and tell you plainly which are worth automating and which are not.