Financial Crime Has a Network. Defenders Need a Shared Mind.

Every bank sees a customer. The criminal sees a network.

One institution sees an account opening. Another sees a sequence of instant payments. A crypto platform sees conversion into a digital asset. A telecom operator sees a device reused across identities. Law enforcement sees a known associate. Each participant can behave correctly and still miss the scheme, because the signal exists in the relationship between their fragments.

This is the structural weakness at the heart of financial-crime defence: the adversary coordinates across the system while the defenders investigate inside institutional boundaries.

AI can make an individual bank faster at screening alerts. That is useful, but it does not solve the geometry of the problem. A better model looking through the same narrow window still has a narrow view.

The next step is not one omniscient database. It is a shared mind: institutions able to learn from one another, identify network patterns, and coordinate action without abandoning privacy, legal boundaries, or accountability.

Criminals do not respect the org chart

Most financial institutions are designed to manage risks they can observe directly. Their systems know their customers, accounts, transactions, alerts, and investigations. Their controls are organized around that perimeter.

Financial crime is organized around avoiding it.

A mule network can distribute activity across dozens of accounts and several providers. Each account may look ordinary in isolation. Amounts stay below familiar thresholds. Money moves quickly enough that no institution holds it for long. Identities, devices, beneficiaries, and corporate structures overlap just enough to coordinate the scheme, but not enough to trigger a simple rule at any one bank.

The FATF's work on data pooling and collaborative analytics describes the problem directly: criminals exploit information gaps while engaging with multiple domestic and international institutions, each with a partial view. FATF also identifies the possible upside of collaboration—better risk understanding, fewer false positives, and more effective reporting—while insisting that sharing remain consistent with data-protection and privacy law.

That combination matters. The answer to fragmentation cannot be indiscriminate surveillance. It has to be better coordination with explicit limits.

More intelligence inside each silo is still a silo

Banks are already applying machine learning to transaction monitoring, entity resolution, alert prioritization, and case preparation. These systems can find subtler patterns and help investigators focus their attention. As I argued in The First Banking AI Winners Will Be Invisible, financial crime is exactly the kind of operational domain where AI can create value before customers ever see it.

But local optimization has a ceiling.

Imagine five banks, each holding one piece of a laundering chain. If every bank improves its model by 20 percent, each still lacks the other four pieces. The models may become more confident without becoming more complete. Worse, a shared vendor or common training approach can teach every institution to miss the same new pattern.

The BIS Innovation Hub's Project Aurora tested a different architecture. Using synthetic data, machine learning, network analysis, and privacy-enhancing technologies, it compared siloed monitoring with national and cross-border collaborative approaches. Its central result was directional but important: collaborative analysis was more effective at detecting laundering networks than institutions analysing activity alone.

The unit of intelligence changed. Instead of asking whether one transaction looked suspicious, the system asked what the network was doing.

A shared mind is not a shared database

The phrase “shared mind” can sound like a demand to put every customer's data into one place. That would be both politically unrealistic and often legally wrong.

Collective intelligence does not require collective possession of all information.

Different architectures can reveal useful patterns while limiting exposure. Federated learning can bring a model to data held by participants rather than moving raw data into a central store. Secure multi-party computation can let several parties compute a result without revealing every underlying input. Homomorphic encryption can enable limited computation over encrypted data. Tokenization, hashing, access controls, and purpose limitation can narrow what is exchanged and why.

These tools are not magic. They introduce cost, complexity, and their own attack surfaces. They cannot repair an unclear legal basis or poor governance. But they change the design question from “Can we pool everything?” to “What is the minimum information required to detect this pattern safely?”

Project Hertha, run by the BIS Innovation Hub with the Bank of England, explored precisely that question in real-time retail payments. On a large synthetic dataset, payment-system analytics using a limited set of data points helped participating institutions find 12 percent more illicit accounts than they otherwise would have found; for previously unseen behaviours, the reported improvement was 26 percent. These were simulation results, not a production guarantee. Their value is in showing that a network-wide signal can supplement—not replace—the institution's own investigation.

Privacy and detection are design constraints, not opposites

Financial institutions often discuss information sharing as a trade-off: either protect privacy or catch more criminals. That framing produces paralysis.

Both objectives are legitimate. Both have to be engineered.

Singapore's COSMIC platform offers a useful governance pattern. Participating financial institutions can securely share information about customers exhibiting multiple specified red flags when stipulated thresholds are met. The important idea is not simply that institutions share. It is that sharing has a defined purpose, participation model, threshold, and legal framework.

The 2026 FATF global review of public-private partnerships makes the same point at broader scale. It identifies durable partnerships as mechanisms for exchanging typologies, risk trends, and, where permitted, operational information. Their effectiveness depends on legal foundations, governance, secure channels, and clarity between financial-crime and data-protection authorities.

This is also why a bank's supposed data problem is often really a permission problem. A shared mind must know not only what data exists, but who may use which signal, for which investigation, under which authority, and for how long.

Build the minimum viable shared mind

A financial-crime consortium does not need to begin with a global platform. It can begin with one typology, one legal perimeter, and one measurable outcome.

I would build it through six layers:

Shared purpose. Define the crime pattern and public-interest objective narrowly. “Fight financial crime” is not a usable mandate. “Detect coordinated mule activity across instant-payment participants” is.
Minimum signal. Identify the smallest set of fields, derived indicators, and network relationships needed. Do not collect information merely because it might become useful later.
Permission architecture. Document legal basis, participant eligibility, allowed uses, retention, access, audit, and redress. Every query and disclosure should have an accountable owner.
Collaborative analytics. Select the least intrusive architecture that can answer the question: secure exchange, federated analysis, privacy-enhancing computation, or carefully governed pooling.
Human investigation. Treat the model as a way to generate leads and assemble context, not as an autonomous accusation. The limits described in “Human in the Loop” Is Not a Governance Model still apply: reviewers need time, evidence, authority, and escalation paths.
Learning loop. Return confirmed outcomes to participants, test for displacement and bias, update typologies, and retire signals that create noise without stopping harm.

This framework deliberately separates seeing together from deciding together. Participants can contribute to network detection while retaining responsibility for customer action. In some cases, a payment-system operator or public authority may generate a supplementary signal; the bank still evaluates it against its own information and obligations.

Measure networks disrupted, not alerts produced

The easiest metric in financial-crime technology is the number of alerts generated. It is also one of the least useful.

A shared system should be judged by whether it finds activity that siloed controls missed, reduces duplicated investigation, improves the quality and speed of actionable intelligence, and helps stop money before it disappears. It should also measure false positives, privacy incidents, unequal impact, investigator workload, and the time between detection and intervention.

In the United States, FinCEN's Section 314(b) programme provides a safe harbour, subject to its requirements, for participating financial institutions sharing information to identify and report possible money laundering or terrorist activity. The mechanism is voluntary. That is a reminder that legal permission creates an opportunity; operating discipline determines whether institutions use it well.

The future of financial-crime defence will not be one super-model declaring who is suspicious. It will be a network of institutions able to combine limited signals, test competing explanations, and coordinate action under clear rules.

Criminals already understand the system as a whole. Defenders do not need to know everything about everyone. They need to stop seeing every fragment alone.