For financial institutions, choosing an AI model is no longer only about which one performs best. According to Napier AI, AI strategy is increasingly shaped by where a model was trained, who trained it, what data it used and where it runs. These factors don’t necessarily conflict with performance, but they are a different kind of measure.
This shift was the focus of a recent panel at the Napier AI /AML Assembly. Practitioners from technology, financial services and payments took part, and the session ran under the Chatham House Rule.
Running software in local or dedicated environments is nothing new. Accessing certain AI models in a fully disconnected way has been harder, because some providers still prohibit it. That limits the options for institutions with strict data residency requirements. Panellists agreed that as sovereign cloud offerings for AI mature, control over data will become a baseline expectation rather than a premium extra.
The case for control goes beyond compliance. When a multinational seeks to expand a credit facility, several banks may start from the same information and could even put the same prompt into the same general-purpose model. What sets them apart is their institutional and tacit knowledge. Training AI on that knowledge is a real differentiator. Napier AI notes that the same applies to financial crime compliance, where typologies, risk appetite and investigator judgement are hard-won assets.
On regulation, panellists said its purpose is to protect and set boundaries, not to make AI better. A lighter touch may speed deployment, but it carries risk. Firms that skip explainability could later have to reverse engineer autonomous systems before they can adjust them, which is slow, costly and sometimes impossible. Napier AI says it treats traceability and auditability as a core discipline, so that every alert, disposition and escalation can be explained to a regulator.
There is a paradox, though. Sovereignty pushes institutions to keep data close, yet research suggests the industry could catch more crime by sharing. Appetite for sharing is strong in payments and growing in banking, particularly for consortium analytics. Its value is clearest in authorised push payment fraud, where signals about receiving accounts help institutions shut down mule accounts faster.
Synthetic data offers another route. Napier AI took part in the Anti-Money Laundering and Synthetic Data Project alongside the FCA, the Alan Turing Institute and Plenitude Consulting. The project produced a fully synthetic dataset that the FCA is making available through its Digital Sandbox. Napier AI has since applied frequency-based AI algorithms to large synthetic datasets in the FCA Supercharged Sandbox.
The way to reconcile sovereignty and sharing is to separate raw data from the intelligence derived from it. Collaboration can happen through signals, patterns and typologies while customer data stays under an institution’s control. Napier AI’s cross-customer learning follows this principle. The panel suggested firms ask five questions of any vendor: where the model was trained and on what data, where it runs, whether its outputs are explainable, what leaves their environment, and how the approach fits each jurisdiction’s data-sharing rules.
For more information, read the full story here.
Read the daily FinTech news
Copyright © 2026 FinTech Global
Copyright © 2026 RegTech Analyst
