AI in Swiss finance: a state of play in numbers
Everyone talks about artificial intelligence in banking. The figures settle the question of hype versus real adoption. And they are clear: the shift has already happened.
Between late November 2024 and mid-January 2025, FINMA surveyed around 400 supervised institutions. The result: roughly half already use AI or are building their first applications, and a further quarter plan to start within three years. On average, each institution concerned runs five applications in service and nine in development. On the banking side, the EY Banking Barometer 2026 points the same way: 78% are actively working on introducing AI, against about 53% a year earlier. A twenty-five-point jump in twelve months is not a gentle trend.
| Indicator | Figure | Source and scope |
| Financial institutions using or developing AI | ~50% | FINMA survey, ~400 supervised entities, Nov. 2024–Jan. 2025 |
| Among them, use of generative AI and chatbots | 91% | FINMA survey, April 2025 |
| Banks actively working on introducing AI | 78% (vs ~53% a year earlier) | EY Banking Barometer 2026 |
| AI applications per institution (average) | ~5 in service, ~9 in development | FINMA survey, 2025 |
| Global rank, private AI funding (cumulative since 2013) | 14th USD 4.73 bn | Stanford AI Index 2026 |
What these figures actually say
Two readings stand out. On one hand, usage is spreading fast across operations. On the other, Switzerland invests little: 14th worldwide, USD 4.73 billion of cumulative private funding since 2013, far behind the United Kingdom (USD 34.1 bn) or even Sweden. The country trains and employs some of the best AI talent in the world, yet funds innovation modestly. That gap between skills density and capital volume shapes which specialisations AI is reshaping: the large players push ahead while the start-up ecosystem stays under-capitalised.
Generative AI and "classic" AI: don't mix them up
The distinction matters for what follows. So-called classic AI sorts, scores and predicts: anomaly detection, risk models, credit scoring. Generative AI produces text, code and summaries the ChatGPT wave. The 91% recorded by FINMA relate to this second category, often grafted onto office tasks. Confusing the two leads to shaky decisions: you don't govern a fraud-scoring engine the way you frame a writing assistant.
Where AI already creates value in Swiss institutions
Use cases cluster where data is plentiful and tasks are repetitive. Here are the four fronts where the gains are tangible today, not in a forward-looking scenario.
| Function | Dominant technology | Concrete gain | Limit to watch |
| Portfolio management and advice | Predictive ML, robo-advisors | Continuous allocation and reporting, lower-cost advice | Model drift, over-optimisation |
| Compliance and AML | Adaptive scoring, NLP | Fewer false positives, targeted alerts | Explainability of decisions |
| Fraud detection | Anomaly detection | Near real-time response | Biased data, adversarial attacks |
| Client relationship and back office | Generative AI, chatbots | 24/7 answers, lighter admin load | Confidentiality, banking secrecy |
Portfolio management: advice augmented, not replaced
On the investment front, AI processes market-data volumes in real time that no team could track by hand. Robo-advisors bring basic advice to more clients at low cost. But the manager keeps control of the trade-offs and the relationship. The trap here is blind trust: a model overfitted to past data can be flat wrong at the first unfamiliar market regime. AI widens the field of view; it does not replace judgement.
Compliance and anti-money laundering: the most profitable ground
This is where the return on investment reads most clearly. Legacy AML setups rely on fixed rules that trigger mountains of false alerts. An adaptive engine learns to tell noise from signal. Take a typical case, made up with numbers for illustration: a Geneva private bank of 400 staff, with a compliance cell of nine analysts handling around 1,200 alerts a month, of which close to 85% are false positives. Loaded cost of an analyst: about CHF 135,000 a year.
By cutting the false-positive rate from 85% to 55% through scoring, the cell investigates around 430 genuinely relevant alerts instead of ploughing through more than a thousand. Three roles are redeployed onto higher-value investigation: roughly CHF 405,000 of annual capacity freed, against a licence and integration cost of the order of CHF 180,000 in year one. The maths turns positive from year one. Those who run audit and statutory review roles already see it: the time saved is reinvested in analysis, not in cutting headcount.
Client relationship: the UBS example, at full scale
The UBS case gives the scale. The bank logged 8 million prompts on its generative AI tools in the second quarter of 2025, a volume multiplied by four since late 2024. Its internal assistant "Red" is used by 52,000 employees, and more than 280 use cases are operational. The stated target: 15% efficiency gains. This is a long way from a confidential pilot.
Back-office automation: the least visible gains, the surest
Reconciliations, data entry, extraction from contracts, generation of regulatory reports: these are the least spectacular tasks, and yet the ones where AI pays back its entry ticket fastest. No demo effect, but hours recovered every week and fewer re-keying errors. Our advice: start there, not with the showcase use case. A smoother back office then funds the more ambitious projects.
What FINMA actually expects
Many institutions are waiting for "the" AI regulation. They miss the key point: FINMA has already set out its expectations, without new legislation. Its supervisory communication 08/2024, published on 18 December 2024, describes what it observes and what it requires. The guiding principle fits in one phrase: same business, same risks, same rules.
| Area | FINMA expectation | Operational implication |
| Governance | Clear responsibilities for each model | A designated owner per AI application |
| Risk management | Inventory and classification of uses | Register of applications, periodic review |
| Explainability | Understandable, justifiable outputs | Documentation, testing, human oversight |
| Data | Quality and protection (FADP) | Source control, minimisation |
| Outsourcing | Control over vendor dependency | Audit clauses, exit plan |
Explainability: the point that really sticks
One point deserves emphasis, because it is where FINMA raised the most concern: explainability. The authority found that in many cases staff neither understood nor could justify the outputs produced by their own models. Our position is blunt: a model you cannot explain has no business running in production at a financial institution. This is not a bureaucratic box to tick, it is a survival condition under inspection or in a dispute.
The Swiss framework against the European AI Act
Switzerland has chosen a different path from Brussels, true to form. No horizontal AI law: on 12 February 2025 the Federal Council instructed the FDJP to prepare, by the end of 2026, a draft for consultation. The aim: implement the Council of Europe Convention on AI, acting on transparency, data protection, non-discrimination and oversight. A sector-based approach, anchored in existing law.
| Dimension | Swiss approach | AI Act (European Union) |
| Philosophy | Sector-based, on existing law + CoE Convention | Horizontal, by risk tiers |
| Timeline | Draft to consultation by end 2026 | Phased application since 2025 |
| Reach for a Swiss company | National law, banking secrecy, FADP | Extraterritorial as soon as you operate on the EU market |
| Logic | Avoid over-regulation, preserve innovation | Graduated obligations, harmonised penalties |
Extraterritorial reach, the blind spot of Swiss banks
The classic mistake is to assume a Swiss bank escapes the AI Act because Switzerland sits outside the EU. Wrong. As soon as a system or its outputs touch the European market, the text applies. Groups active on both sides of the border must therefore work with two regulatory logics and anticipate their convergence rather than absorb it after the fact.
The risks the sector still underestimates
The gains are real, so are the blind spots. Three risks recur in FINMA's findings, and none is solved with a single tool.
- Bias and discrimination. A model trained on historical data reproduces the inequalities it contains in credit granting, in particular. Without auditing the datasets, you automate the problem instead of fixing it.
- Cybersecurity and confidentiality. AI systems churn through massive volumes of sensitive data. Every integration widens the attack surface. The typical breaking point: a consumer-grade tool wired to client data.
- Vendor dependency. FINMA repeats it in its risk monitoring: outsourcing critical functions to cloud third parties or model vendors creates a dependency you must put under contract and be able to unwind.
The healthy reflex is not to ban, but to frame before deploying: encryption, logging, no reuse of data for training, human oversight over high-impact decisions. The technology does not excuse the absence of guardrails.
Training and hiring: the blind spot of adoption
This is the subject people discuss least, and yet the most decisive one. You can buy a model. You cannot buy the internal capacity to use it properly. Our position at Fed Group, after supporting many hires across Swiss finance: the bottleneck is neither the budget nor the technology, it is human capital.
The profile that is truly missing: the business-to-technical translator
Contrary to a widespread idea, the first hire to make is not a data scientist. The profile most in short supply is the "translator": someone who knows the business compliance, management, risk and can talk to the technical side to frame a use case and challenge a result. Digital skills become a cross-cutting foundation, not an isolated speciality.
- Being able to frame a use case and measure its return, before buying any tool.
- Reading an AI output with a critical eye: where the data comes from, where the biases sit.
- Mastering the framework: FADP, banking secrecy, FINMA expectations.
- Steering the vendor relationship and auditing their models.
Train or hire: the right trade-off
These hybrid profiles are rare and expensive. To place their pay in the Swiss context, you have to reason in net terms: what these profiles actually take home depends heavily on the canton and on contributions. Betting on continuous training for the teams already in place is often faster and more retentive than the race for external hires.
Frequently asked questions
Does Switzerland have a specific AI law?
No horizontal law to date. Existing law applies FADP, code of obligations, banking secrecy. A sector-based draft is expected to go to consultation by the end of 2026.
Is a small financial firm covered by FINMA communication 08/2024?
Yes. It targets all supervised entities, in proportion to their usage. Even a single chatbot triggers minimum expectations of governance and traceability.
Does the European AI Act apply to a Swiss bank?
Only if it places AI systems or outputs on the EU market. For a strictly domestic activity, Swiss law prevails.
Do you need data scientists to get started?
Not first. The initial need is for people who can frame a use case and challenge a model. The business-to-technical translator matters more than pure technical depth.
Are generative AI and banking secrecy compatible?
Yes, provided the data stays within a controlled perimeter: controlled hosting, no reuse for training, logging. The risk comes from consumer tools wired to client data.
Read also
- How to become an accountant in Switzerland
- Study finance in Switzerland
- Fiduciary firms in Switzerland
- Financial controller compensation
- Accountant salaries in Switzerland
Useful resources and documents
- State Secretariat for International Finance (SIF) Artificial intelligence
- Swiss Financial Market Supervisory Authority (FINMA)
- Federal Office of Justice Regulation of artificial intelligence
Sources
- FINMA survey on the use of AI in Swiss financial institutions (24 April 2025).
- FINMA supervisory communication 08/2024, governance and management of AI-related risks (18 December 2024).
- EY Banking Barometer Switzerland 2026.
- Stanford AI Index Report 2026.
- Federal Council, press release of 12 February 2025 on AI regulation (admin.ch).
- UBS, second-quarter 2025 results (GenAI adoption figures).
