Why AI must accelerate AML investigations and why humans must stay at the core

New capabilities and new expectations are reshaping AML. Banks and regulators agree that static rules engines and siloed data no longer meet the demands of modern financial crime, and the impact shows up across every program. False positives dominate workloads. Customer friction rises. Teams grow year after year simply to meet regulatory expectations. Every institution feels the pressure.

The good news is that AI is now entering the operational stack, and this changes what AML teams can expect from their systems. These tools surface stronger signals earlier in the process and cut the noise that consumes so much investigative time. And this is where the much-needed modernization of AML processes begins. Collaborative federated machine learning networks and the future potential of agentic AI copilots will give banks more accurate detection and a way to direct human expertise toward activity that carries real consequences.

But as new AI capabilities mature, the question arises: what is the right balance between automation and human judgment in financial crime investigations?

AI is now essential to move beyond static rules and high false positives

The limitations of traditional AML systems are well understood. Static thresholds, basic peer comparisons, siloed data and limited adaptability all contribute to an alert that produces more noise than usable insight. The result is a program that absorbs resources without improving detection quality in any impactful way.

Recent research shows why AI is becoming central to addressing this. A 2024 study on fraud detection found that models incorporating human analyst feedback achieved higher accuracy, stronger recall and significantly fewer false positives. The key factor was the human input itself. Labels, rationales and contextual notes provided signals that automated methods could not infer from transaction data alone.

Academic reviews of deep learning in AML point to the same issue. Fully automated decisions remain unrealistic when banks must justify Suspicious Activity Reports to regulators. AI can triage and prioritize, and build narrative, but high-risk systems must include humans with the expertise to challenge automated suggestions and the context investigators bring to each case.

Humans remain essential: The research backs this

Across governance, ethics and law, researchers reach the same conclusion. Human oversight is not a safeguard that can be added at the end of the process. It is the mechanism that keeps high-stakes decisions accountable and defensible.

The research makes this unavoidable:

Human control is required for accountability and for decisions that reflect context and moral reasoning. Davidovic’s work highlights that systems influencing rights or regulatory outcomes must remain under meaningful human authority.

Poorly designed oversight turns people into passive approvers. Green warns that without the time, context and authority to intervene, oversight collapses into a rubber-stamp process that adds no real judgment.

High-risk systems must include humans with the expertise to challenge automated suggestions. Legal analyses of the EU AI Act also highlight that oversight is not symbolic. It requires individuals who understand the logic behind the system and can question it when needed.

Core elements of investigation cannot be automated. Public-sector studies, including Enarsson’s work, show that discretion, contextual interpretation and situational understanding often fall outside what models can encode, yet these factors are central to investigative outcomes.

Regulators echo this position. They expect explainability, auditability and a clear record of how decisions were made, especially when those decisions carry legal or regulatory weight.

And even AI developers acknowledge the limits. OpenAI’s own system cards caution that advanced models may produce incorrect information and require human review in consequential settings.

Human oversight does not slow AML programs down. It anchors them. AI can surface patterns and reduce noise, but the decisions that matter still rely on human intelligence, context and accountability.

The future: Agentic AI that supports investigators (not replaces them)

As AI reduces false positives and sharpens the signals that matter, the next step is to give investigators stronger support throughout the workflow. This is where agentic AI comes in. These systems are designed to work alongside investigators, not in place of them.

Agentic AI handles the operational load, so investigators can focus on judgment. It retrieves evidence, assembles context, summarizes case histories and produces narrative drafts that investigators can refine. It keeps track of new intelligence and brings forward the information that shapes better decisions.

Here’s what agentic AI allows investigators to:

🟣Work cases faster with context already assembled. Less time gathering data means more time assessing risk.

🟣Spend time analyzing behavior, not searching for supporting details. The system surfaces what matters instead of forcing investigators to chase it.

🟣Reach defensible conclusions with clearer evidence trails. Structured summaries and consistent documentation support auditability.

🟣Focus on alerts that genuinely require expertise. Routine tasks are handled by the system so investigators can prioritize complex cases.

The industry is already moving in this direction. Research in AML and fraud shows that combining automated pattern recognition with human-validated reinforcement improves both accuracy and operational efficiency. OpenAI’s own system cards reinforce the importance of human confirmation in consequential settings, which aligns directly with the requirements of regulated environments.

Agentic AI follows a straightforward principle. Automate the tasks that can be automated safely, and strengthen the human attention that determines investigative outcomes.

Consilient’s view: Humans, Federated Learning and agentic AI

Consilient’s approach brings together three elements that strengthen AML programs without compromising human judgment.

1. Collaborative federated AI

Learns from patterns across institutions without sharing customer data. Banks gain richer intelligence and stronger signals while maintaining full data protection.

2. Human-in-the-loop adjudication

Keeps investigators in control of decisions. Their rationale improves model performance, builds trust in outcomes and supports regulatory compliance.

3. Agentic AI copilots

Automate investigative tasks and assemble context so analysts can spend their time on true risk rather than manual processing.

These components address the two core challenges in AML:

• Scale and accuracy: AI surfaces behavior that static rules miss and reduces false positives, improving both coverage and efficiency.
• Nuance and accountability: Human judgment remains at the center of decisions that require explanation, proportionality and regulatory defensibility.

In this model, investigators do not step back. They step up. AI handles the volume. Humans make the decisions that matter.

The best AML systems are hybrid by design

The debate is not between automation and humans. AML programs need both. The evidence across research, regulation and industry practice points to the same outcome. The strongest systems use AI to deliver scale and signal, and rely on human judgment for context, accountability and the decisions that carry regulatory weight.

Federated learning strengthens this by giving institutions broader insight without sharing customer data. Agentic AI adds workflow support that helps investigators move faster and focus on real risk.

This is the direction of modern AML. AI reduces noise. Humans interpret what matters. And together, they create a program that is more accurate, more defensible and better able to identify the activity regulators expect institutions to find.

Consilient’s federated learning approach is built for this model. It identifies anomalous patterns across institutions without sharing customer data, keeps investigators in control of escalation and reporting decisions, and improves with every investigator outcome fed back into the system. The result is stronger signals, fewer false positives and a clearer rationale behind every decision. If you’d like to learn more, we’re always here to talk.