.svg)
Sovereign AI and shared intelligence: can anti-money laundering have both?
Insights from a panel at the Napier AI /AML Assembly
Which AI model performs best is no longer the only question financial institutions are asking. Increasingly, the AI strategy is being shaped by where a model was trained, by whom, with what data, and where it runs. These factors are not necessarily in tension with performance, but they are different measures.
That was the starting point for a recent panel discussion bringing together practitioners from across technology, financial services and payments. The session was held under the Chatham House Rule, so the views below are summarised without attribution.
Why data sovereignty is moving up the agenda
Running software in a local or dedicated environment is not new. What has been harder is accessing certain AI models in a fully disconnected manner. Some model providers still do not allow this, which limits the options available to institutions with strict data residency or control requirements. The panel agreed that as sovereign cloud offerings for AI mature, the ability to control data will become a baseline expectation rather than a premium feature.
The more interesting argument was about why control matters beyond compliance. Consider a large multinational corporate looking to expand its credit facility. It will typically approach several banks at the same time, each starting from broadly the same information. In principle, every bank could put the same starting prompt into the same general-purpose model. What separates them is how each institution gets from that starting point to a decision: its institutional and tacit knowledge.
The ability to train AI on that knowledge is a genuine differentiator. It is also why institutions are protective of where their data goes and which models it touches. The same logic applies to financial crime compliance, where an institution's typologies, risk appetite, and investigator judgement are hard-won assets.
AI / AML regulation, speed and explainability
The panel also considered how regulation, and in some market's deregulation, will affect AI adoption. One view was that regulation is not designed to make AI better. Its purpose is to protect and to set boundaries, from limiting bias in models through to governing how AI is used in customer-facing processes.
A lighter regulatory touch is likely to accelerate deployment across organisations. But speed carries a risk. Institutions may reach a point where AI is running compliance checks across the business, only to find they must reverse engineer those systems to understand why decisions were made before they can adjust them. That challenge grows as AI systems gain more autonomy, with agents able to interpret their environment and adapt their behaviour with less human direction.
In other words, firms that skip explainability to move faster are storing up problems for later. Recording how an AI system reaches each decision is far easier to build in from the start than to piece together once the system is live, which can be slow, costly and sometimes impossible. At Napier AI, we treat traceability and auditability as a core discipline, not an add-on: every alert, disposition and escalation should be explainable and auditable to a regulator. We explored this in more depth in The accountability gap: AI, automation and AML compliance.
The data sharing paradox
If sovereignty pushes institutions to keep data close, research consistently points out the other way: the industry could detect more financial crime if it shared more. Does a narrow, sovereign approach to data contradict the case for collaboration?
One view from the panel was that the appetite for sharing is strong, particularly in the payments industry but increasingly in banking. Interest is growing in consortium analytics, where signals are shared across a group of institutions while appropriate levels of privacy and confidentiality are preserved. Its value is most visible in authorised push payment (APP) fraud. Rather than looking only at the account sending the money, institutions can see signals about what is happening with the receiving account, helping them identify and shut down mule accounts faster. Card networks are also keen to play a role in coordinating intelligence across the industry.
Synthetic data offers another route to collaboration without exposing customer data. Napier AI was part of the Anti-Money Laundering and Synthetic Data Project, a multi-stakeholder initiative including the Financial Conduct Authority (FCA), the Alan Turing Institute and Plenitude Consulting. The project began with real banking data that had already been anonymised at source, with no personal data included in the initial request. That data was enhanced with realistic synthetic money laundering scenarios to create a fully synthetic dataset, which the FCA is making available through its Digital Sandbox so that firms can test new detection approaches on a shared foundation. Napier AI has since applied novel frequency-based AI algorithms to large-scale synthetic financial datasets as part of the FCA Supercharged Sandbox.
The challenge is that the answer will differ by jurisdiction. Sharing data in a way that is both useful and compliant remains difficult, but the appetite is clearly there. For more on this shift, read From institutions to networks: The future of AML detection.
Reconciling data sovereignty and fincrime collaboration
Sovereignty and collaboration are less contradictory than they may seem at first. The key distinction is between raw data and intelligence derived from it. Sovereignty governs where data is held, where models run, and who controls them. Collaboration can happen at the level of signals, patterns, and typologies, without customer data ever leaving an institution's control.
This is the principle behind Napier AI's approach to cross-customer learning, which improves defaults, configurations and typology coverage for every customer without any customer transaction or alert data leaving its own tenant.
For institutions assessing AI for anti-money laundering (AML), the panel's discussion suggests five questions worth asking of any model or vendor:
- Where was the model trained, and what data?
- Where does it run, and can it operate in a disconnected environment if required?
- Can every output be explained and evidenced by a regulator?
- What leaves your environment: raw data, or only derived signals?
- How does the approach align with data-sharing rules in each jurisdiction where you operate?
Looking ahead
The future of AI in AML will not be decided by model performance alone. Institutions that combine control over their data and institutional knowledge with a willingness to share intelligence responsibly will be best placed to stay ahead of financial crime.
The Napier AI / AML Index 2026–27 is coming soon, with new data on how AI is shaping anti-money laundering across global markets. Sign up to be among the first to receive it.
