AI as the Coordination Layer for Anti-Money Laundering

Financial institutions have spent billions on anti-money laundering technology, compliance staff, and regulatory reporting. The investment is real, and so is the commitment. Yet the system as a whole continues to underperform — not because individual institutions are failing at their jobs, but because AML is fundamentally a coordination challenge. The gaps that matter most exist between institutions, between banks and regulators, and between the financial system and law enforcement.

Artificial intelligence offers the first credible path to closing them. The framework from the book _Reshuffle: Who Wins When AI Restacks the Knowledge Economy_helps us understand why. Unlike previous technologies, what makes AI different is the combination of three capabilities that have not previously existed together: the ability to create a unified representation of a problem or domain by reconciling and synthesizing data across multiple sources and modalities, the ability to reason across that representation and recommend decisions, and through agents, the ability to act on those decisions autonomously.

The coordination failures that need to be addressed exist at three distinct levels.

Level One: Inside the Institution

Start inside the bank itself. Customer risk data today is fragmented across transaction monitoring systems, KYC platforms, sanctions screening tools, and even marketing databases. Each system holds a piece of the picture. No single team sees all of it. A customer flagged in one system may appear clean in another. A pattern visible across three data sources may be invisible to any analyst looking at just one.

Reconciling these signals manually is slow, inconsistent, and error-prone. AI agents can traverse these siloed systems, reconcile conflicting data points, and surface a unified view of customer risk in real time. Beyond automating routine tasks, the deeper value of AI here is synthesis — connecting data that was never designed to connect and producing risk assessments more complete than any single analyst or system could generate alone.

This is where most industry attention is focused today, and rightly so. The elimination of manual reconciliation work alone represents a significant productivity gain for compliance teams. More importantly, a unified customer risk view changes the quality of decisions that analysts and investigators can make — decisions that were previously constrained by incomplete information become grounded in a full picture of customer behavior across all systems.

But this represents a local optimum. A better-informed bank is not, by itself, a better-functioning AML system. The gaps that most undermine the system exist not within institutions but between them — and closing those gaps requires the system’s actors to connect and participate in shared workflows in ways they currently do not.

Level Two: Between Institutions

Money launderers do not respect institutional boundaries. A typology that looks like isolated transactions at one bank may form a clear pattern when viewed across three. A customer who appears low-risk at one institution may have a history of suspicious activity at two others. The information exists in the system. It just cannot flow to where it is needed.

The US Congress recognized this problem when it created Section 314(b) of the USA PATRIOT Act, which provides a safe harbor for voluntary information sharing between financial institutions. The intent was sound. The execution has fallen short. As of 2024, just over 6,000 out of 324,000 eligible institutions were registered to participate. The reason is not legal uncertainty — FinCEN has clarified the scope of permissible sharing on multiple occasions. It is friction.

In practice, information sharing under 314(b) means a compliance analyst picking up the phone or drafting an e-mail or letter to a counterpart at another institution — a process that depends entirely on individual initiative, existing relationships, and the bandwidth to act on a suspicion that may or may not be reciprocated. There is no mechanism that makes this happen automatically. It only happens when someone chooses to make it happen, and chooses to invest the time and effort with no guarantee of a useful response.

AI changes this calculus entirely. Rather than humans reaching out to humans across institutions, agent-mediated information sharing could allow banks to request and exchange relevant data in a compliant, auditable, and scalable way. An AI agent that identifies a suspicious pattern could automatically query whether other registered 314(b) participants have seen related activity — receiving and synthesizing responses without any human having to initiate the outreach.

Critically, this does not require a central clearinghouse or a vendor acting as intermediary. Open-source agent communication protocols make it possible to build a decentralized, protocol-driven coordination layer that institutions can adopt without ceding control to a single commercial authority. Section 314(b)'s existing safe harbor provides the legal foundation. What has been missing is the operational infrastructure to make it real. AI finally provides that infrastructure.

Level Three: Between the Financial System and Law Enforcement

The third and most overlooked coordination failure runs between financial institutions, regulators, and law enforcement. It is also the one with the most direct bearing on whether AML actually fulfills its purpose.

Every SAR filed with FinCEN is, in theory, an input to law enforcement. Financial institutions invest significant resources in investigating suspicious activity, drafting narratives, and filing reports intended to help government agencies detect and prosecute financial crime. In practice, the institutions that file those SARs almost never learn whether their filings were useful. There is no feedback loop.

Without feedback, banks optimizing for SAR quality are flying blind. They cannot learn what kinds of narratives are most useful to investigators, what level of detail is appropriate, or which typologies are generating the most actionable intelligence. The industry’s de facto incentive becomes filing more SARs rather than better ones — a dynamic that burdens law enforcement with volume while delivering uncertain value.

AI offers a practical path out of this. Today, law enforcement analysts query SAR databases manually, looking for filings related to a specific entity or network. Imagine instead that AI agents perform this function — querying the database, mapping connections across filings, and presenting a synthesized intelligence picture to the investigator. Patterns that would take an analyst hours to assemble manually could be surfaced in minutes, with connections drawn across filings that no individual analyst would have the bandwidth to identify.

When an investigator finds that synthesized picture useful, a simple action through a review interface could capture that signal and trace it back to the specific SARs that contributed. Regulators could then relay that feedback to the filing institutions. This closes a loop that has been open for decades, and creates a mechanism for financial institutions to learn what useful looks like — and to get meaningfully better at it over time.

From Local to Global Optimum

Global trade did not scale just because ships got faster or ports were modernized. It scaled because containerization standardized how goods moved across every port and border in the world. A common infrastructure layer made coordination automatic rather than effortful, and transformed a collection of local efficiencies into a genuinely global system.

AML needs an equivalent infrastructure moment — not better tools inside individual institutions, but a common coordination layer across the system. The three levels described above — intra-institutional synthesis, inter-institutional information sharing, and regulatory feedback — are all, at root, coordination failures. AI makes it possible to address all three. But technology alone does not determine outcomes.

The rules, incentives, and feedback loops that govern the system will ultimately determine whether that moment arrives. The Bank Secrecy Act’s purpose is to enable financial institutions to provide highly useful information to the government — but the current system’s incentives reward volume over quality, and offer institutions no mechanism to learn from their own contributions.

Changing that requires action on both sides. Banks have underutilized Section 314(b) for years; AI finally makes meaningful inter-institutional information sharing practical rather than aspirational. Government must move too, embedding AI into how regulators process SARs and how law enforcement queries financial intelligence, rather than simply mandating its adoption elsewhere. That combination — banks coordinating across institutional boundaries, and government participating as an active node in the system it oversees — is what finally pushes AML past the local optimum each institution can achieve alone, toward a global one the system has never managed to reach