The Safest Payment Data Is The Data That Never Happened
Central banks want AI in supervision. They do not want a model hallucinating its way through confidential payment data and producing a regulatory footnote that reads like a chatbot apology with legal consequences.
That is the useful tension behind BIS Project AISE.
On April 15, the BIS Innovation Hub published Project AISE, short for AI Supervisory Enablement, as a technical proof of concept for AI-enabled supervisory analytics in retail payments. The project is grounded in retail-payments supervision use cases and uses synthetic data throughout, with the explicit aim of producing practical lessons before authorities move toward real-data experimentation (BIS Innovation Hub, April 15, 2026).
That structure is the story. AISE is not asking whether a large model can summarize a report. Everyone in finance has seen that demo. It is asking what kind of architecture lets a supervisor trust, reproduce, challenge, and document an AI-assisted conclusion.
In regulation, the bottleneck is not the prompt. It is the evidence trail.
Supervisory AI is not useful because it sounds fluent. It is useful only when it can show its receipts.
Payments are the right test bed because they combine scale, network structure, operational risk, fraud patterns, third-party dependencies, and confidentiality. They are also politically radioactive. Real payment data can expose consumers, merchants, banks, non-bank payment firms, and behavioral patterns that no supervisor wants leaking into a vendor demo environment.
Synthetic data solves only part of that problem. It lets authorities test methods without exposing live records. It does not guarantee that the methods will survive contact with live supervision. That is why AISE is framed as a staging layer, not a production deployment.
The Model Is The Least Interesting Piece
The BIS describes AISE as a comparative test of alternative analytical architectures under identical conditions. The listed components include machine-readable reporting, knowledge graphs, deterministic analytics, and constrained model-based synthesis. The point is to learn which foundations add value and which complexity can be avoided (BIS Innovation Hub, April 15, 2026).
That is a more mature question than “which model wins?”
A supervisor’s problem is not one document. It is a messy chain of reporting, entity relationships, alerts, internal notes, third-party dependencies, previous examinations, and risk judgments. If an AI tool cannot preserve the link between output and source, the supervisor gets a faster way to produce unauditable prose. That is worse than a spreadsheet. At least the spreadsheet is honest about being ugly.
AISE’s design language points in the opposite direction. BIS says governance and security are built in, and it names concrete risks: unsupported outputs, prompt injection, retrieval manipulation, weak reproducibility, and unclear accountability. The proof of concept tests whether AI-assisted capability can operate inside a constrained architecture that anchors outputs in evidence, preserves provenance, and supports structured human review (BIS Innovation Hub, April 15, 2026).
Those are not decorative controls. They are the difference between an assistant and an unlicensed shadow analyst.
The broader BIS work on AI in policy reinforces the same point. Its G20 report says BIS Innovation Hub projects using AI span suptech, climate-risk analysis, cyber security, and monetary-policy technology. In the AISE entry, BIS describes a virtual assistant for on-site supervision that integrates generative AI into supervisory workflows to automate data extraction, improve pattern detection, standardize on-site supervision, and streamline report generation through modular reusable tools (BIS G20 report, October 2025).
But the same report warns that explainability is a core constraint. It says sophisticated machine-learning models can improve predictions, but complex interactions among variables make it hard to identify what drives output. For generative AI, it highlights the risk of confident factual errors and says human supervision remains necessary for tasks that require logical reasoning (BIS G20 report, October 2025).
That is exactly why AISE matters. It treats AI supervision as institutional plumbing. The model sits inside a controlled workflow. It does not replace the workflow.
Synthetic Data Is A Governance Choice, Not A Toy Dataset
Synthetic data often gets dismissed as fake data for fake proofs of concept. That is too simple.
In supervision, the alternative is frequently no experimentation at all. Real supervisory and payments data are difficult to share, hard to anonymize fully, and dangerous to move across vendors or jurisdictions. Synthetic data gives authorities a way to test analytical patterns, tooling interfaces, governance controls, and review processes before requesting access to sensitive production data.
Other regulators are moving in the same direction. The UK’s Financial Conduct Authority published an April 15 research note on a synthetic-data anti-money-laundering project with the Alan Turing Institute, Plenitude, and Napier AI. The FCA said detailed financial data is often constrained by legal and privacy limits, and the project created a fully synthetic dataset based on real UK retail-banking data and realistic money-laundering scenarios so firms and regulators can test detection methods without sharing raw customer records (FCA, April 15, 2026).
BIS had already been testing adjacent payment-supervision ideas. Project Hertha, published in June 2025, explored transaction analytics for identifying financial-crime patterns in real-time retail payments. It used no real customer data and instead developed a representative synthetic retail-payments ecosystem for a single jurisdiction (BIS Project Hertha, June 5, 2025).
AISE is different because it pushes beyond detection into supervisory workflow. The question is not only whether an algorithm can spot a pattern. It is whether a supervisory institution can make the pattern usable: sourced, repeatable, explainable enough, and reviewable by humans with statutory responsibility.
That is a narrower ambition than the AI sales deck. It is also more likely to survive procurement, audit, and court.
The Implication For Financial Regulators
Project AISE suggests a practical sequence for AI adoption inside financial authorities.
First, build synthetic supervisory environments. They will not answer every production question, but they let teams test architecture and controls before sensitive data enters the loop.
Second, make reporting machine-readable. AI cannot rescue supervisory material if the underlying inputs are trapped in inconsistent documents, stale templates, and attachment archaeology. A model can summarize garbage. That does not make the garbage regulated.
Third, use knowledge graphs and deterministic analytics where they are enough. If a rules-based or graph-based method gives a traceable answer, do not route it through a generative model just to make the output look modern. Save model-based synthesis for the parts where language and judgment are genuinely useful.
Fourth, design for review. The output should make a supervisor faster, not make accountability disappear. Every claim needs evidence. Every evidence link needs provenance. Every model-assisted synthesis needs a human decision point.
That architecture will disappoint anyone expecting supervisory AI to look like consumer chat. Good. Supervision is an evidentiary business.
The more important market read is for banks and regtech vendors. Authorities are starting to reveal what they will expect from AI systems used in regulated finance: local deployability, data controls, retrieval integrity, evidence anchoring, reproducibility, and workflow accountability. A vendor that cannot explain those elements will struggle to move past demos, even if its model benchmark slide looks strong.
AISE is still a proof of concept. It does not prove that central banks can safely run AI over live payment data tomorrow. It proves something more useful: supervisors are learning to ask the right implementation questions before the vendor arrives with a chatbot and a compliance brochure.
That is how serious AI adoption begins in finance. With fake payments, real controls, and reproducible workflows.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.