Between 70% and 85% of AI projects in financial services fail to reach production or deliver measurable business value, according to research from Gartner, FinTellect, and the RAND Corporation. FINMA’s April 2025 survey of Swiss financial institutions found that approximately 50% of institutions use AI or have applications in development, yet governance frameworks concentrate on data protection risks rather than model risks such as bias, lack of explainability, and lack of robustness. FINMA’s April 2026 comprehensive guidelines went further: board-level AI governance will now be assessed as part of regular supervisory review. The implication is direct. Swiss banks are deploying AI at scale and building the programme conditions under which the global 85% failure rate applies: ungoverned models, undefined ownership, unmeasured outcomes. The difference in Switzerland is that FINMA is now actively looking for what is missing.
Volume and value are not the same thing
The pattern of misprioritisation is consistent across every market in this series, and Switzerland is no exception. AI investment concentrates first on high-volume, lower-consequence decisions such as document processing, basic transaction categorisation, and routine credit decisions below certain thresholds, because the efficiency case is straightforward and the sponsor is easy to identify. These are reasonable automation candidates. They are not where the most material return sits, and they are not the decisions that FINMA’s current supervisory focus is examining.
The decisions with the highest value at stake in Swiss retail and commercial banking combine high volume with severe financial consequence per error, and each carries a specific regulatory dimension that accelerates the case for AI. SIC5 real-time fraud scoring sits unambiguously at the top of the portfolio for the reasons established in the first article in this series: the ten-second processing window makes batch-scoring architecturally inadequate, irrevocable settlement means detected fraud is largely irrecoverable, and FINMA Guidance 02/2026 has specifically found that 42% of Swiss banks have no digital fraud policy, which means the majority of institutions do not have the governance infrastructure even to begin a properly governed AI fraud programme.
AML transaction monitoring is the second priority, driven by Switzerland’s strengthened AML obligations under FATF standards, expanded beneficial ownership requirements, and the consistent global finding that rule-based AML alert systems produce false positive rates of 90 to 95%. In Switzerland’s correspondent banking and commercial banking context, where transaction complexity is high and regulatory scrutiny of AML programme quality has intensified following enforcement actions at major Swiss institutions in recent years, the analyst burden created by rule-based systems is both operationally costly and regulatorily exposed.
Commercial credit fraud, including invoice fraud, CEO fraud, and synthetic identity applications, is the third priority, reflecting the doubling of CEO fraud cases between 2023 and 2024 documented by the Swiss Bankers Association and the specific vulnerability of commercial banking to AI-generated document forgery. AI models trained on behavioural and transactional signals can detect the anomaly in how a counterparty is behaving, not just whether a document passes format validation. That is precisely the gap that AI-generated fraud exploits.
FINMA’s supervisory signal has resolved the sequencing debate in a specific way: institutions that cannot demonstrate governance-grade AI fraud controls are facing examination findings before they face the fraud losses those controls would have prevented. The compliance requirement and the commercial opportunity are pointing at the same decisions.
The Three Reasons AI Programmes Fail and What They Look Like in Switzerland
The failure patterns break into three categories in roughly the same proportions across every market assessed. They are not technology failures. They are programme design, organisational, and measurement failures.
Governance failure accounts for around 42% of failed initiatives. Technology is selected before the problem is defined. The proof of concept proves the model works. Nobody is assigned to act on it. No production funding pathway exists at the start. FINMA’s Guidance 02/2026 survey found that three of the nineteen institutions surveyed lacked any steering committee with responsibility for digital fraud risk, a specific manifestation of governance failure in the Swiss context, where the absence of a named owner with budget authority is not only a programme failure condition but a FINMA supervisory finding. FINMA’s April 2025 AI survey found the same pattern in AI governance broadly: institutions deploying AI without the dedicated oversight mechanisms FINMA has explicitly described as expected.
Sequencing failure accounts for around 31%. Data was not ready. Nobody admitted it until the proof of value was already running. In Switzerland, the data readiness question has a specific texture for SIC5 fraud detection: labelled outcome data for irrevocable instant payment fraud is qualitatively different from historical card fraud data. Fraudulent SIC5 transactions that were authorised because the customer was deceived often appear as legitimate transactions in existing systems until a dispute is raised, and dispute records are not always linked back to the original transaction in ways that enable correct model training. Institutions that begin building AI SIC5 fraud detection without first auditing the quality and completeness of their labelled outcome data will discover the gap after the programme has been committed.
Measurement failure accounts for the remaining 27%. The model achieved its technical targets but nobody defined what a good business outcome looked like in dollar terms, or assigned accountability for achieving it. FINMA’s comprehensive AI guidelines require that institutions maintain continuous monitoring of AI system performance, precisely the ongoing measurement infrastructure whose absence drives this failure mode. A model in production without a defined economic baseline, without ongoing performance monitoring, and without accountability assigned to a named owner cannot satisfy FINMA’s monitoring expectations, regardless of how technically capable the model is.
| Failure mode | Share of failed initiatives | Swiss-specific dimension |
|---|---|---|
| Governance failure | ~42% | FINMA Guidance 02/2026 found 3 of 19 surveyed institutions had no digital fraud steering committee. FINMA AI guidelines require board-level oversight now assessed in supervisory reviews. |
| Sequencing failure | ~31% | SIC5 fraud outcome data often not correctly labelled in existing systems: customer-authorised fraud under deception appears as legitimate until dispute raised. |
| Measurement failure | ~27% | FINMA comprehensive AI guidelines require continuous monitoring. No baseline means no FINMA-defensible performance evidence. |
Source: Swiss Banking Practice analysis, consistent with findings across North America, EMEA, APAC, ANZ, LATAM, and South Africa markets.
The structural implication is direct. The programme conditions that produce the 85% global failure rate are present in Swiss banking. FINMA’s own survey evidence confirms it. The governance, sequencing, and measurement disciplines that would make AI programmes succeed are also the disciplines that produce FINMA-compliant AI governance. In Switzerland, there is no gap between building AI that works and building AI that satisfies the regulator. They require the same programme design.
What this means for sequencing
The correct first decision for most Swiss retail and commercial banks is SIC5 real-time fraud scoring, which scores highest across all prioritisation criteria simultaneously. It has the highest volume of any growing fraud category, the highest consequence per error on an irrevocable rail, sufficient labelled outcome data available if properly structured, and the most immediate and specific regulatory urgency from FINMA Guidance 02/2026. No other decision in the Swiss portfolio combines all four criteria as strongly.
AML transaction monitoring is the second priority, where the false positive burden case is the same as in every other market and the FATF compliance dimension adds regulatory urgency specific to Switzerland’s position as a major international financial centre. Commercial credit fraud prevention is the third priority, where the data requirements are met by existing commercial credit systems and the doubling of CEO fraud cases provides the value case.
The value of making this sequencing explicit, through a structured cross-functional conversation that surfaces data readiness gaps, names the sponsor who will own production outcomes, and establishes the economic baseline, is not the output. It is the failure prevention that the process creates. These are precisely the conditions whose absence drives the majority of Swiss AI programmes to fail in the ways FINMA has now publicly documented.
Part 2 of 3.
Sources
FINMA. Guidance 02/2026: Digital Fraud at Banks. April 2026. FINMA. Comprehensive AI Governance Guidelines. April 2026. FINMA. Survey on AI Use in the Swiss Financial Sector. April 2025. FINMA. Guidance 08/2024: Governance and Risk Management when using Artificial Intelligence. December 2024. Swiss Bankers Association. Results of the Preliminary Study on Collaborative Fraud Prevention. March 2025. Gartner. AI Project Failure Rates. Multiple editions, 2023–2025. FinTellect AI. Why 80% of AI Projects in Finance Fail. 2024. Swiss Banking Practice. Decision Portfolio and Failure Mode Analysis. Internal analysis, consistent across North America, EMEA, APAC, ANZ, LATAM, and South Africa markets.
