The decision-time evidence gap

Swiss financial institutions use AI. FINMA's own survey of around 400 supervised institutions found that about half already use it or are developing first applications, and that smaller institutions mostly rely on applications developed outside. At the same time, very little of this AI reaches a regulated decision in production: in a 2025 study of Swiss financial institutions, 11% of AI initiatives had scaled, and 94% of private banks reported no revenue effect from AI in 2025.

The usual explanation is "explainability": the model is a black box, so compliance will not let it near a decision. That diagnosis points the wrong way. Nobody in a Swiss bank is asked to explain neurons. What FINMA, an auditor or a client asks, sometimes years later, is far more concrete: what did the AI propose, which rule ran, what did the person see, who decided, and why?

That question has a name. We call it the decision-time evidence gap.

What the supervisor actually asks

In December 2024 FINMA published what it looks at when a supervised institution uses AI (Guidance 08/2024). It is not a law and it does not prescribe a tool. It lists seven things an institution should be able to show: an inventory with a risk class, clear responsibilities, data quality, testing and ongoing monitoring, documentation, explainability, and independent review.

The sharpest sentence is about explainability. FINMA observed that results "häufig nicht nachvollzogen, erklärt oder reproduziert und somit nicht kritisch beurteilt werden können", that they often cannot be traced, explained or reproduced, and therefore cannot be critically assessed. It added that when decisions had to be justified to clients, employees, the supervisor or the audit firm, it looked at explainability in depth.

Under the monitoring heading there is a quieter sentence that matters just as much: FINMA expects institutions to analyse the cases "in denen die Ausgabe von den Anwendern ignoriert oder verändert wurde", the cases where a person ignored or changed the AI's output. In other words: the override is part of the record, and the record must exist.

This is not new for Swiss finance. For anti-money-laundering records, the law has required for years that institutions create them "so that a competent third party can form a reliable judgement" of transactions and business relationships ("Belege so erstellen, dass fachkundige Dritte sich ein zuverlässiges Urteil ... bilden können", AMLA Art. 7). For those decisions, the record that answers "why was this decided" has always been the institution's duty, and FINMA's 2024 guidance now asks the same question of AI. AI makes the decision faster and the record harder to assemble.

Why the model log is not the answer

The first instinct is to keep the model's logs and, if asked, run the model again. Two facts stand in the way.

First, generative models do not reliably reproduce their own output. In a published test, a large open model was asked the same question one thousand times at temperature zero, the setting that is supposed to make it deterministic. It produced eighty different completions. The cause is mundane: server load and batching change the arithmetic. Running the model again tomorrow does not tell you what it said today.

Second, even a perfect replay would miss what the decision depended on: which version of the sanctions list was in force at that hour, which policy pack applied, what the officer saw on the screen, and the reason the officer gave for approving, rejecting or overriding. None of that lives in a model log. It lives, in most institutions, in e-mails, screenshots, ticket comments and memory.

So the supervisor's test, "traced, explained or reproduced", cannot be passed by the model. It can only be passed by a record that is captured outside the model, at the moment of decision, and that stays readable afterwards.

What such a record contains

A decision record that closes the gap holds five things, each kept apart from the others:

1. The contributions, attributed. What the source documents said, what the AI proposed, what the screening tool returned, what the person wrote. Each one labelled with its origin and its moment, never merged into one text.
2. The checks that ran before anyone read a score. Which policy pack, which version, which result. A score cannot buy a pass.
3. The refusal, when support was missing. A case does not close quietly; if evidence is absent, the record says so and the case escalates.
4. The human decision, with the person's own rationale. One named, authorised person signs. An AI recommendation is not a decision.
5. The moment, sealed. The order of events, the versions in force and the timestamp, kept so that the future can add to the record but cannot rewrite it.

This is the shape FINMA's seven questions take when they are asked about one decision instead of a whole programme. It is also what a Swiss institution has always promised its supervisor and its auditor, and never its competitors or the public: confidentiality outward, a reliable judgement inward.

What this means for a bank, an asset manager, a vendor

For a bank or an asset manager, the practical question is not "do we use AI" but "for the AI-assisted decisions we already take, can we show these five things from one record today, and what does it cost us to assemble them when asked?" Where the answer is "from e-mail and memory, in days", the pilot that is waiting for compliance will keep waiting. Where the answer is "yes, in minutes", there is nothing to buy.

For a vendor selling AI into Swiss institutions, the same question arrives from the buyer's compliance team, usually late in the sale: how will your output be reconstructed and proven to an independent auditor later? A vendor that can hand over its decisions in a form the bank's record can carry has removed the last objection; a vendor that cannot has handed its sale to the compliance calendar.

For an auditor, the gap is simply the difference between a file that can be read and a file that has to be rebuilt.

Where TAIA stands

TAIA builds the record. It does not replace an institution's AI, screening, case-management or core systems, and it does not sell a model: whatever AI an institution uses, the record of the decision is captured outside that model and stays readable to the people entitled to read it. In TAIA's own hosted environment, the records and their backups are stored in Zürich. The contract for that record, which fields a decision record must carry to be readable later, is published openly as the TAIA Calibre specification. The first step with an institution is not software. It is a two-week measurement on the institution's own closed files: what reconstructing the basis of one AI-assisted decision costs in a year, and the single point at which recording it at decision time would remove most of that cost. If the burden turns out not to be material, that is a valid result, and we say so in writing.

TAIA is pre-market. Its platform runs on synthetic data; it has no production bank data and no paying client yet; it is not approved by FINMA, and it is not an audit firm. Your auditor reads the record. We engineer how it is produced.

---

Sources. FINMA, Aufsichtsmitteilung 08/2024, "Governance und Risikomanagement beim Einsatz von Künstlicher Intelligenz", 18 December 2024. FINMA, press release on its AI survey of supervised institutions, 24 April 2025. Swiss Fintech Innovations / OST / ELCA, "AI from PoC to Production", 29 October 2025. KPMG Switzerland, Private Banking Study 2026. Federal Act on Combating Money Laundering (AMLA, SR 955.0), Art. 7, as in force from 1 October 2026. Thinking Machines Lab, "Defeating Nondeterminism in LLM Inference", 10 September 2025. Stanford RegLab / HAI, "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools", 23 May 2024 (legal research tools incorrect in 17% to 34% of queries even with retrieval).