
AI for Financial Services in Switzerland: the 2026 Guide

By
Ennio Limbach Brun del Re
AI for financial services is no longer a future topic in 2026. Almost half of the independent wealth managers in Switzerland already use AI in individual processes, yet only a small share runs applications in production. This guide shows what AI concretely delivers for a small financial firm today: which use cases pay off first, what getting started costs, what regulation expects, and how you move from a first pilot to reliable operations in four phases.
Where does the Swiss financial industry stand in 2026?
Between April and June 2026, Lucerne University of Applied Sciences and Arts (HSLU) and the Zurich-based WealthTech company Finup surveyed more than 40 independent wealth managers in a structured study (AI Use Cases and Crypto Assets in External Asset Management, July 2026). The result describes the starting point more precisely than any forecast:
- 49 percent of the wealth managers surveyed already use AI in selected processes; another 39 percent are exploring or planning.
- Of the 65 reported AI applications, 66 percent are experiments with public tools, 23 percent are pilots, and only 11 percent are scaled implementations.
- 85 percent name data protection as the biggest hurdle, and only 27 percent consider their own data ready for AI.
- The most common use cases: investment research, KYC and AML checks, and client communication.
The study also reveals what drives the respondents: 93 percent name expected efficiency gains as the most important driver, while positioning as an innovator plays a minor role at 32 percent. One wealth manager interviewed put it plainly: administrative tasks have grown steadily over the years, mainly because of regulatory requirements, and the hope is that AI gives firms back some of that time for the end client.
The study also shows a clear divide. In language- and content-heavy areas such as investment research, proposal writing, and client reporting, AI use is already visible. In workflow- and data-intensive processes such as client onboarding or reconciliation with custodian banks, it remains low despite the perceived potential. If you start today, begin where language and documents sit at the center, then expand toward workflows.
My reading of these numbers: the bottleneck is not the technology. What is missing is prepared data, clear ownership, and a measurable goal for the first pilot. That is exactly where this guide starts.
The same development is simply further along at the large end of the market. McKinsey estimates the annual value that generative AI can add to the banking sector alone at USD 200 to 340 billion (The economic potential of generative AI, June 2023). What insurers, neobanks, and large wealth managers have built in year-long programmes since then is now available to a small financial firm in weeks. The reason: the building blocks are standardized, the models are affordable, and hosting in Switzerland is a solved problem.
What has changed for small financial firms?
Three years ago, AI in finance meant hiring your own team, building your own model, running your own data center. That is over. Three things have changed.
First, the cost. You pay for modern language models per use, not as a capital project. A pilot today costs what a feasibility study used to cost.
Second, the hosting. Time and again, the first thing raised is the question: where does my data live? The answer is straightforward today. Protected Swiss cloud environments, for example Microsoft Azure in the Zurich North region, keep client data in the country and cleanly separated from public models.
Third, the building blocks. Connectors to CRM, document management, and accounting systems exist off the shelf. An AI agent does not have to replace your existing stack. It becomes an intelligent orchestrator on top of the systems you already use.
Together, this changes the entry logic. An AI initiative used to start with a budget decision and an annual plan. Today it starts with a single process, a manageable investment, and a result within a few weeks. The question is no longer "can we afford AI?" but "which process delivers the fastest proof?"
Which use cases pay off first?
AI does not have to start with a transformation programme. The biggest returns often come from automating everyday, high-volume tasks. Five fields have proven themselves in practice:
- Client reporting and dashboards. Generate performance, risk, and compliance reports directly from core systems, with templating, versioning, and audit logs. The output is reproducible and auditable.
- Document processing. Incoming contracts, tax forms, and KYC files are read, classified, and filed. Downstream work such as compliance review starts with prepared data.
- Compliance screening. KYC, AML, sanctions and PEP screening, periodic re-screening, discrepancy detection. The machine flags anomalies; the human decides.
- Client interaction. Meeting preparation, drafts for follow-up emails, answers to standard queries with human approval.
- Back office and accounting. Process invoices, match payments, reconcile accounts, flag anomalies.
As a result, the role of employees shifts away from processing toward reviewing results and exceptions. That is exactly where the value clients pay for is created.
What does this mean for independent wealth managers?
The study shows why the pain is real: the effort for onboarding, KYC, and reporting has grown for years while teams have stayed the same size. A concrete example from my own work: T&T International, a wealth manager with a very small team, handles high volumes of compliance work and reporting with an operations application. It reads bank reports, prepares the necessary documents including compliance checks, and shows clients a consolidated view of what they hold where: multiple banks, currencies, and products. The time required has dropped by 80 percent.
In the world of large institutions, my work with additiv shows how the same idea scales: hybrid advice, where AI supports the advisor before and after meetings, or discretionary portfolio management with self-service onboarding and automated processing, launched at one of Europe's largest neobanks. What matters for a small firm is the logic, not the size: capture data properly once, prepare documents automatically, free up people for advice.
More on the landing page AI for wealth managers. Explored in depth in the pillar article AI for independent wealth managers: 7 use cases that already work today.
What does this mean for insurance brokers?
For brokers, the leverage sits on two sides. In the existing book: a quote comparison agent compares all of a client's offers within seconds and prepares an individualized email. The client sees before the consultation what they save or gain in additional coverage. In acquisition: good new business starts with the occasion, not the call. AI detects buying signals such as site expansions or hiring programmes and prepares the outreach.
Here too, the view of the large players is instructive. An advisor assistant, introduced at one of the world's largest insurers, suggests next best actions based on client history and trigger factors and sends occasion-based emails automatically. For a broker with a team of three to thirty people, a fraction of that is enough: find the right occasions and reach out personally at the right time.
More depth on the landing page AI for insurance brokers.
What does this mean for the back office?
Wherever numbers and documents flow, automation pays off fastest: invoice processing, reconciliations, tax preparation, periodic reports. These processes have clear inputs and outputs and a manageable risk profile. That makes them ideal as a first pilot.
A typical pattern from practice: invoices arrive by email and are entered, posted, and approved manually. An agent reads the attachments, matches them against vendors and contracts, proposes the posting, and flags deviations. The human reviews the exceptions, not every receipt. The same pattern works for reconciliations with custodian banks, for preparing tax documentation, and for periodic quality checks in master data.
How does an AI agent work in practice?
The term sounds abstract; the implementation is not. An AI agent is a software component with four capabilities:
- Perceive: it reads input from documents, APIs, or system events.
- Plan: it decides which steps are needed, for example calling a model or querying a database.
- Act: it invokes tools, produces results, and triggers workflows.
- Remember: it keeps track of context and past steps so decisions stay traceable.
What matters for daily work: the agent uses the same tools you use today. Via APIs and connectors it talks to your CRM, document management, and accounting. Where no interface exists, screen automation takes over. Core processes, data, and systems stay in place; the intelligence arrives as a layer on top.
In a regulated industry, the agent runs in assisted mode. It creates drafts, flags anomalies, and prepares decisions. Approval stays with a human. Every action is logged, with timestamp and reasoning. That produces exactly the trail that auditors and supervisors expect.
In daily work, the collaboration looks like this: the agent produces the prepared dossiers in the morning, your team reviews and refines, and only after an approval does anything go to clients or into the archive. At the start you review strictly; as the hit rate grows, you loosen the review density where the agent works reliably. Trust here is not created by a contract clause but by measured quality over weeks.
What does regulation require?
I deliberately keep this short and cautious, because the details depend on your licence and business model.
- FINMA: Supervisory notice 08/2024 of December 2024 sets out expectations for governance and risk management when using AI (FINMA, supervisory notice 08/2024). The core: you must know where AI is used, who is responsible, and how errors surface. That requires clarity, not a corporate structure. What the seven measures of the notice concretely demand and which minimal version is enough for a small team is explained in FINMA guidance 08/2024 explained.
- Data protection (revFADP): Processing personal data is governed by purpose limitation, data security, and transparency (revised Federal Act on Data Protection, Fedlex). For AI this means in practice: keep data in the country, control access by role, never feed client data into public models. The ten most important duties as a working list are in the revFADP checklist for financial firms; the location question is covered in depth by AI hosting in Switzerland.
- FinSA: Documentation and advisory duties toward clients remain in force (FinSA, Fedlex). AI can prepare and standardize meeting records; responsibility for the advice stays with the human.
- ISA: Insurance intermediaries face their own information and documentation requirements (ISA, Fedlex). Again: AI may support; the obligations remain.
My experience: if you build auditability, data residency, and human approval in from the start, you avoid expensive rebuilds later. Compliance here is not a brake but a design principle.
What does it cost to get started?
Hardly anyone quotes numbers. I quote them as orders of magnitude, because the exact figure depends on the scope of the pilot: a compliant setup sits in the mid five-figure franc range; a lean pilot comes in below that, plus manageable monthly costs for operations, models, and support. Larger or hybrid environments sit above that.
The math usually works out quickly. If automation frees 60 to 100 hours of manual work per month, the setup typically pays for itself within 6 to 12 months. The details are in the cost article What does AI cost for financial services firms?, and the readiness check assesses your starting point in a few minutes.
Which mistakes should you avoid?
From the study and from my own projects, four patterns emerge that slow pilots down:
- Data not ready. Only 27 percent of the wealth managers surveyed consider their data ready for AI. If you underestimate data preparation, you pay for it twice later.
- Nobody owns it. A pilot without an owner loses momentum after the first demo.
- No measurable goal. "Let's try AI" is not a goal. "We halve the time per client report" is one.
- An oversized first attempt. Three workflows at once mean triple the coordination and no clean result.
Projects that reach production do the opposite: one process, one owner, one target metric, one fixed review slot per week.
How do I proceed? Four phases
Phase 1: Assessment (2 to 4 weeks). Map processes, quantify bottlenecks, check data readiness. The output is a pilot candidate with high volume, clear inputs and outputs, and moderate risk.
Phase 2: Pilot (4 to 8 weeks). Automate one workflow, run it in shadow mode, and compare human results with the agent's output. Measure error rate, turnaround time, and user feedback.
Phase 3: Rollout (2 to 3 months). Bring the team on board, define responsibilities and approvals, switch governance and logging to production. Then connect adjacent workflows.
Phase 4: Operations (ongoing). Monitor quality, maintain models and rules, keep an eye on regulation. Operations is not the end of the project but the normal state.
Two things I consider decisive across all phases. First, a fixed rhythm: fifteen minutes a week on the metrics, a monthly decision on what to expand. Second, honest communication within the team: the agent takes over processing so people gain time for clients and the demanding cases. Teams that experience it this way bring the next use cases themselves.
From my practice: The pilots that fail rarely have a technology problem. They lack a measurable goal or an owner. The pilots that reach production start with a single, clearly bounded process and one person who looks at the numbers every Friday.
Why I see it this way
I do not write this from the outside. At additiv I am part of how insurers and neobanks build advisory applications: guided advice for complex topics, advisor assistants that prepare client meetings, self-service onboarding in portfolio management. With Humanyx I translate the same logic for small teams: Leadstreams for sales, Taskstreams for operations. And with clients such as T&T International I see that the math works even for a team of a few people, provided the entry point is chosen well.
This perspective shapes the recommendations in this guide: start small, think about regulation from day one, measure impact, then scale. It also shapes the limits I set: AI in my projects replaces neither advice nor accountability. It returns time that was tied up in processing, and it makes decisions more traceable. Anything beyond that is promise, not practice.
Conclusion: your next step
The industry is moving, the technology is ready, and the costs are manageable. What is usually missing is simply a well-chosen first process. This week, pick one workflow that regularly steals hours from your team, for example reporting, document intake, or compliance reconciliation. Check your starting point with the readiness check, and if you want an independent assessment: the page AI consulting is the direct route to a conversation.
FAQ
Is AI worthwhile for a team of five?
Yes, often especially so. Small teams feel repetitive work directly, and the entry cost of a pilot is in the range of a part-time hire for a few months. What matters is a clearly bounded first process.
Do I need my own AI model?
No. Standard models running in a protected Swiss cloud environment and connected to your systems via connectors have proven themselves. A custom model only makes sense for very specific requirements.
What happens to my client data?
In a compliant setup, client data stays in Switzerland, for example in the Azure Zurich North region. It does not flow into the training of public models. Access is role-based and logged.
How quickly does a pilot pay for itself?
As a rule of thumb from my own projects: 6 to 12 months if the pilot automates a process with high manual effort. Without a measurable goal it takes longer, or never happens.
Do I have to report my use of AI to FINMA?
That depends on your licence and supervisory regime. FINMA sets out expectations for governance and risk management, not a general duty to report every use of AI. Clarify your specific case with a specialist; see the note in the regulation section.

About the author
Ennio Limbach Brun del Re
Ennio Limbach Brun del Re is the founder of Humanyx, Leadstreams and Taskstreams. Through his time at additiv, he brings years of WealthTech experience and supports financial services firms in adopting AI systems.
LinkedIn

