I wrote this article with the assistance of AI tools (Claude), which also helped build the figures. The experience, sources and conclusions are mine, and I checked every claim and citation.
Executive summary
Problem. Swiss banks and insurers already use AI: in FINMA's 2025 survey of around 400 institutions, about half used it or were building it. Switzerland has no AI-specific law, but FINMA expects the existing rules on governance and risk to cover AI. Most talk about AI safety is about where data goes and who can call a model. Much less is about recording what the AI decided, and why.
Approach. I compare four habits from the month-end close (reconcile, keep the audit trail, separate preparer and reviewer, decide what is material) with what FINMA's guidance asks for, and with what each looked like in my own AI project.
Findings. Each habit matched a failure I saw: a test report that covered less than it seemed, a model call cut off without warning, a reviewer that is a second opinion rather than an independent check, and a final answer that dropped its caveats.
Limits. One project of my own and three recorded runs of one question. It shows how these failures happen, not how often.
Takeaway. Governing AI needs no new principle. It is the discipline finance already has, applied to a new kind of preparer.
At the end of every month, finance teams close the books. Before anyone signs, the numbers are reconciled, the trail is checked, someone other than the preparer reviews them, and the team decides which differences matter. None of this is new. It is what makes a number trustworthy enough to sign.
My background is in making financial numbers provable: IFRS reporting, consolidation, SOX controls. For the last two years I have been building AI systems. In my last article I described what happened when I let AI agents build one of them: every stage was reported "done", and the headline feature had never worked. Looking back, most lessons from that project are ones that finance learned long ago.
Verdaca is my working name for that system: it runs a small panel of AI models on a question and writes a receipt, a JSON record of the evidence, the draft, the critique, the outcome and the cost of every model call.
Regulators are asking the same questions
Switzerland has no AI-specific law, and FINMA says so directly: "To date, there is no AI-specific legislation in Switzerland." The existing rules on governance and risk management apply to AI, scaled to how material each application is (FINMA, 2024). A broad AI act is not planned. In February 2025 the Federal Council decided to ratify the Council of Europe's AI Convention. Any new rules should be "as sector-specific as possible", and a draft bill is due for consultation by the end of 2026 (Swiss Federal Council, 2025).
FINMA's Guidance 08/2024 reads like a close checklist: an inventory of AI applications, data quality, tests and ongoing monitoring, documentation, explainability and independent review.
AI is already inside Swiss finance. In FINMA's survey of around 400 financial institutions, about half used AI or had applications in development, and another quarter planned to within three years. Among those using AI, 91% also used generative AI (FINMA, 2025).
The United States has no federal AI law either, and a December 2025 executive order set up a federal task force to challenge state AI laws (Executive Order No. 14365, 2025). For banks, US supervisors replaced their model-risk guidance, the rules for how banks check the models they rely on, in April 2026. The new version says generative and agentic AI models "are not within the scope of this guidance" (Board of Governors of the Federal Reserve System et al., 2026). Leaving generative AI out of a framework does not remove the risk, or the duty to manage it (Godlove & Buchanan, 2026). In the EU, the AI Act's main obligations for stand-alone high-risk systems now apply from 2 December 2027, after the July 2026 amendment (European Union, 2026).
So the AI-specific rules are still being written. The duty to govern risk already applies. If AI touches a decision, you need to show your work.
Two layers of AI governance
"AI safety and governance" covers two different jobs, and it helps to keep them apart.
Layer 1: where your data goes and who can use the model. This is what cloud platforms offer. Anthropic's Claude models are available directly from Anthropic and through Amazon Bedrock, Google Cloud and Microsoft Foundry. Each route differs in who runs the model, where data is processed and which controls come with it. Here is a short summary of each provider's public documentation, as of October 2026.
Layer 2: what the AI decided, and why. The controls above tell you who can call a model and where the data goes. Some go further: Amazon Bedrock, for example, can check whether an answer is grounded in its source documents (Amazon Web Services, n.d.-b). But these are filters applied at the moment of the call. They are not a decision record of what evidence an answer relied on, what challenged it, who signed off, and whether its caveats survived. That second layer is what this article is about, and it is the layer I work on in Verdaca. The two layers solve different problems. A fully locked-down model can still give a confident answer that nobody can reconstruct.
The rest of this article is about layer 2. Here are four habits from the close, and what each looked like in my own AI project.
1. Reconcile before you report
In finance, a number is not finished until it ties to something independent: a bank statement, a sub-ledger, a confirmation. For an AI answer, that can mean recomputing its numbers from the source, checking that every citation exists, or asking the same question more than once and comparing the answers (Wang & Wang, 2025).
My project once reported 411 tests passing and none failing. That was not true. The test folder was not tracked, so 36 failures sat where the automated checks never looked. The green signal was real; it just covered less than it seemed to. Later, when the system made real model calls, I recomputed every cost on its records from token counts and list prices. They matched. That second check is the one I trust.
Ask: what independent source does this AI result tie to, and who checked it?
2. Keep the audit trail, not the summary
Auditors don't accept "the system calculated it". They want the trail.
Every run of my system writes that receipt, and for each model call it also keeps the tokens, the cost and the reason the call stopped. That trail caught something no summary would have shown. In the first run, one model call was cut off at its output limit, and the critique was built on a truncated argument. FINMA asks for the same kind of record for material applications: documentation of "assumptions, limitations, testing and controls as well as fallback solutions" (FINMA, 2024). When the model itself can't be inspected, "opacity does not eliminate accountability; it increases uncertainty" (Godlove & Buchanan, 2026). An auditor would record a scope limitation and qualify the opinion: sign, but state clearly what could not be checked. The AI equivalent is to say so in the answer: mark it as limited, or escalate it to a person, rather than present it as complete.
My trail is still incomplete. No human sign-off is recorded yet, and the receipts are hashed, not signed.
Ask: if this decision is challenged in a year, what record will you show?
3. Separate the preparer from the reviewer
The person who prepares a reconciliation does not approve it. That is segregation of duties.
In my system, a reviewer model writes its own counterargument before it sees the draft, then critiques the draft. That is meant to avoid a model grading its own work, a setup that research shows is biased (Yang et al., 2026). But it is not full independence, because the reviewer judges the draft against its own argument. FINMA saw a similar gap in the institutions it supervises: it "did not observe a clear distinction between the development of AI applications and the independent review in all cases" (FINMA, 2024).
In banks, the "second line" is a separate risk team that challenges the work of the people who build and use the models. A model inside my own system cannot be that.
My reviewer is a second opinion, not a second line.
Human review has its own weak spot. In one study, showing researchers which evidence supported each AI claim lowered their trust, yet they kept relying on the AI's edits anyway (Martin-Boyle et al., 2026). A reviewer on paper is not always a reviewer in practice.
Ask: who reviews the AI, and what makes them independent of whoever built it?
4. Decide what is material, and protect it
Not every difference matters. Materiality is how finance decides which ones do, and FINMA applies the same idea to AI: controls should match how material the application is (FINMA, 2024). FINMA lists factors such as the financial impact, the number of clients affected and the consequences of errors (FINMA, 2024). A low-impact productivity tool does not need the same validation as AI that influences credit or autonomous transactions (Godlove & Buchanan, 2026).
Often the most material part of an AI answer is its qualifications. In one of my runs, the reviewer wrote that other parts of a report "may add mitigations". The final answer dropped that hedge and stated the conclusion flatly. This is a known pattern: in one study, AI summaries of scientific papers were nearly five times as likely as human-written ones to generalize beyond the source (Peters & Chin-Yee, 2025). In finance we know it as the qualified audit opinion that becomes "the auditors signed off" by the time it reaches the board.
Ask: which caveats in this output are material, and does the final answer keep them?
What I take from this
None of these four habits needs a new principle. Some need new tooling, such as a record of every AI run. The discipline is the one finance already has: reconcile, keep the trail, separate duties, and decide what matters.
Governing AI is mostly that discipline applied to a new kind of preparer.
The honest limitation: my evidence is one project of my own and three recorded runs of one question. It shows how these failures happen, not how often. FINMA says it will refine its expectations as its experience grows (FINMA, 2024). When it does, I will compare them with these four habits.
References
Amazon Web Services. (n.d.-a). Data protection. In Amazon Bedrock user guide. https://docs.aws.amazon.com/bedrock/latest/userguide/data-protection.html
Amazon Web Services. (n.d.-b). Detect and filter harmful content by using Amazon Bedrock Guardrails. In Amazon Bedrock user guide. https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html
Anthropic. (n.d.). API and data retention. Claude Platform Docs. https://platform.claude.com/docs/en/manage-claude/api-and-data-retention
Board of Governors of the Federal Reserve System, Federal Deposit Insurance Corporation, & Office of the Comptroller of the Currency. (2026, April 17). Revised guidance on model risk management (SR 26-2). Board of Governors of the Federal Reserve System. https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm
European Union. (2026). Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI). Official Journal of the European Union, L 2026/1744. https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
Executive Order No. 14365. (2025, December 11). Ensuring a national policy framework for artificial intelligence. Daily Compilation of Presidential Documents, DCPD-202501186. https://www.govinfo.gov/app/details/DCPD-202501186/
FINMA. (2024, December 18). Guidance 08/2024: Governance and risk management when using artificial intelligence. Swiss Financial Market Supervisory Authority. https://www.finma.ch/en/news/2024/12/20241218-mm-finma-am-08-24/
FINMA. (2025, April 24). FINMA survey: Artificial intelligence gaining traction at Swiss financial institutions. Swiss Financial Market Supervisory Authority. https://www.finma.ch/en/news/2025/04/20250424-mm-umfrage-ki/
Godlove, T., & Buchanan, J. (2026). From model risk management to AI assurance: Integrating AI risk management, independent validation, and AI auditing for generative and agentic AI. Transactions on Engineering and Computing Sciences, 14(5), 19–62. https://doi.org/10.14738/tecs.1405.12205
Google Cloud. (n.d.-a). Claude at scale on Google Cloud: Frontier AI, built for enterprise production [Blog post]. Google Cloud Blog. https://cloud.google.com/blog/products/ai-machine-learning/claude-at-scale-on-google-cloud-frontier-ai-built-for-enterprise-production
Google Cloud. (n.d.-b). Gemini Enterprise Agent Platform and zero data retention. Google Cloud Documentation. https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention
Martin-Boyle, A., Leckey, C. A. C., Brown, M. C., & Kaur, H. (2026). PaperTrail: A claim-evidence interface for grounding provenance in LLM-based scholarly Q&A. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26). ACM. https://doi.org/10.1145/3772318.3791101
Microsoft. (2026). Compare hosting options for Claude models in Microsoft Foundry. Microsoft Learn. https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/claude-models-hosting-comparison
Microsoft Azure. (2025). Claude in Microsoft Foundry is now generally available [Blog post]. Azure Blog. https://azure.microsoft.com/en-us/blog/claude-in-microsoft-foundry-is-now-generally-available/
Peters, U., & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science, 12(4), 241776. https://doi.org/10.1098/rsos.241776
Popov, A. (2026, October 4). I let AI agents build my governance engine. Here's what they didn't tell me. Verdaca. https://verdaca.com/blog/i-let-ai-agents-build-my-governance-engine
Swiss Federal Council. (2025, February 12). AI regulation: Federal Council to ratify Council of Europe Convention [Media release]. https://www.admin.ch/gov/en/start/documentation/media-releases.msg-id-104110.html
Wang, J. J., & Wang, V. X. (2025). Assessing consistency and reproducibility in the outputs of large language models: Evidence across diverse finance and accounting tasks [Working paper]. arXiv. https://doi.org/10.48550/arXiv.2503.16974
Yang, J., Hu, Z., Qiu, C., Deng, Z., Jiao, X., & Zhou, T. (2026). Quantifying and mitigating self-preference bias of LLM judges (arXiv:2604.22891). arXiv. https://doi.org/10.48550/arXiv.2604.22891