How Alpian Integrates AI into Banking
At heyneo, we are quite keen on new technologies. So, we wanted to understand exactly how AI is being introduced into a Swiss bank. We met someone even more geeky than us: Damien Chambon, Head of Cloud, Data & AI at Alpian. With over fifteen years of experience, he has guided numerous AI projects from prototype to production, particularly in regulated environments.
We met with him to discuss and ask all our questions. How do you combine LLMs and banking data? Why should certain calculations never be entrusted to generative AI? Where do the models run? How is AI used in the app and by the teams? How does Alpian protect data and distribute roles between AI and employees?
We are sharing this fascinating first meeting to understand what really happens behind the scenes when a Swiss bank like Alpian starts integrating AI into its products and operations.
[Interview conducted on August 31, 2026]
How does AI impact the banking market and why must Alpian adopt this technology to avoid being outpaced by the competition?
AI is changing the economics of personalization: what was once reserved for traditional private banking (tailored advice, responsiveness, availability) is becoming scalable. For Alpian, whose promise has always been to democratize this experience, AI is not an option but rather an accelerator to realize our vision.
Moreover, competition is not limited solely to banks: customer expectations are shaped by the best tech products they use daily. A banking application is compared to ChatGPT, Gemini, or Claude, not just a competitor’s e-banking application.
However, one point must be clear: Alpian has never intended to replace advisors with AI. Our model is precisely based on human relationships; that’s what distinguishes us. AI provides support: it absorbs repetitive tasks, prepares the ground for advisors, and synthesizes information so they can dedicate their time to what is most valuable: relationships, judgment, and advice. At Alpian, we don’t automate relationships; we automate what prevents us from dedicating time to them.
And this is where the perspective is interesting: the competitive risk is not being replaced by AI, but being outpaced by those who have made their teams more efficient thanks to it.
What does Alpian aim to achieve with AI at an operational level?
In the market in general, AI at the operational level always focuses on the same use cases: customer support, document production, developer assistance, and internal information retrieval. Nothing truly specific to banking there, and that’s precisely the point: these gains are now accessible to any organization, regardless of its size.
What changes for a bank is the level of requirement. Each use case must be auditable, explainable, and compliant with the Swiss regulatory framework. This is the whole difference between creating an impressive demo and putting a system into production for real banking clients.
At Alpian, we pursue three objectives.
1. Customer Experience
Our AI assistant allows our clients to analyze their expenses in natural language, in four languages, with responses based on their transactions. An immediate, personalized response directly in our mobile application.
2. Operational Efficiency
AI absorbs routine tasks (case synthesis, knowledge base search, developer support) so our teams can focus their time where human input makes a difference.
This efficiency relies on an invisible but decisive foundation: data. Every event on our banking platform (a transaction, an account opening, a client interaction) is continuously captured and feeds our central data warehouse. A single, structured, and always up-to-date source.
Our teams can thus directly query this data via internal AI agents, in natural language, without depending on a technical team for each analysis. A team using AI today competes with much larger structures: the competitive advantage shifts from size to speed of execution.
3. Risk Management
This aspect includes anomaly detection, compliance screening, and the reduction of false positives that unnecessarily tie up analysts. It is the least visible aspect, but in a regulated bank, it is perhaps where AI creates the most value by making vigilance more precise, not less human.
AI does not replace any decision. It prepares, accelerates, filters, and leaves the final judgment to human discretion.
How does Alpian embrace technology and how does it arbitrate its technical choices?
We do not develop our own AI models. This would make no sense, neither economically nor technically, given the billions invested by major laboratories. We consume the best models on the market via Google’s managed services, our platform from day one, with our clients’ data hosted in Switzerland.
However, everything that constitutes Alpian’s value, we build: the orchestration of agents, the secure and private connection to our clients’ banking data, and of course, the controls that frame each response. The LLM (Large Language Model) is a component; the product is everything around it.
Our doctrine can be summarized in one sentence: buy what doesn’t differentiate us, build what does.
We favor open and open-source solutions, especially when they are widely adopted as industry standards. This is the case for infrastructure, monitoring, and interoperability protocols between AI agents. An open standard is an assurance of reversibility, and we do not want to be captive to any proprietary technology.
How do we arbitrate? Every technology goes through the same filter.
First, the regulatory framework: a Swiss bank chooses its suppliers with a formalism that a startup does not have. Each critical provider is subject to a formal risk assessment, and data localization is a non-negotiable constraint.
Next, team size: every technological brick we integrate is a brick we will have to maintain for several years. Poorly chosen, it’s guaranteed technical debt, and a modest team cannot afford to accumulate it.
Finally, speed: in the AI ecosystem, the landscape changes every six months. We therefore prioritize choices that preserve our ability to change our minds.
This results in a clear philosophy: proven, almost “boring” technology for the foundations, and innovation concentrated where the customer sees it. Audacity is not about trendy technologies. Audacity is knowing where to innovate and where to be “sober.”
Does AI work on the client’s device or server-side?
Today, for our assistant, inference is done entirely server-side, on Google Cloud, and our clients’ data is hosted in Switzerland. Our client’s phone never sees the model: the application communicates with our secure APIs, which orchestrate the AI from our infrastructure.
This choice is based on three reasons.
The first is technical: models capable of bank-grade reasoning are too large to run on a smartphone and will remain so for some time.
The second is regulatory, and it is the most structuring: in a bank, every response generated by AI must be traceable and auditable. We must be able to reconstruct what the model answered and on the basis of what data (context). This requirement imposes a centralized architecture: you cannot easily audit what runs on millions of heterogeneous devices.
The third is operational: server-side, a model improvement or a security fix instantly benefits all our clients, without depending on an application update.
Subsequently, I am convinced that the future will be hybrid. On-device AI is progressing rapidly. Google and Apple are integrating models directly into their systems, and it has real advantages: near-instant responsiveness and “privacy by design” since the data never leaves the device.
For targeted tasks such as intelligent form pre-filling, instant categorization, or offline assistance, this is a serious avenue that we are closely monitoring.
But let’s be clear (and this is the crux of the problem): complex reasoning on financial data will remain server-side, where we control security, model version, and auditability.
In the banking environment, the question is not just “where is inference fastest?”, but “where is it most governable?”.
My conviction: local for user comfort, server for everything that involves trust.
How does Alpian distinguish LLMs from other forms of AI used to analyze financial data?
LLMs excel where understanding and expression are needed: grasping a client’s intent when they ask a question in their own words, in their language, and then formulating a clear answer. This is a revolution for customer relations: the machine finally adapts to the human, not the other way around.
But these same models have a structural limit: they are probabilistic. They generate the most plausible answer, not necessarily the exact one. In conversation, this is a strength, but in financial calculation, it is disqualifying.
That’s why, in our systems, an LLM never produces a number. When a client queries their expenses, the language model understands the question and then delegates the calculation to deterministic tools that execute exact queries on their data. The amount displayed does not come from a model’s imagination: it comes from the database, to the cent. The LLM merely clothes it in words.
Financial analysis and risk management fall under another, older and more proven family of AI: predictive AI. Here we are talking about statistical models and classic machine learning.
Fraud detection is the perfect illustration: it involves identifying, in real-time, a transaction that deviates from a client’s usual behavior, with a measured error rate and adjustable thresholds.
It’s the same logic for compliance screening or identifying trends in financial data: these approaches have qualities that generative AI does not. They are measurable, reproducible, and explainable. They can be tested on historical data, their accuracy quantified, and why they trigger a particular alert documented. Exactly what a regulator (and a client) is entitled to expect when it comes to money or security.
In summary: generative artificial intelligence for interaction, predictive for numbers and vigilance, and humans in the loop for every decision that matters.
The three combine in a single product where the client sees only a fluid experience, but under the hood, each AI family strictly stays in its place.
How is AI integrated into the Alpian client application?
The first AI function visible to our clients is a conversational assistant integrated directly into the application: everyone can query their finances in natural language, in the four languages of our clients.
“How much did I spend on restaurants last month?”, “What is my biggest expense item this year?”: the answer arrives in a few seconds, based on their real transactions.
Technically, this is the direct application of the doctrine I mentioned before: the numbers come from the client’s data, the model converts them into words. And the whole thing runs on our Swiss infrastructure, with the controls of a bank, not in a generic chatbot connected to the Internet.
And we must be honest about one point: the model is the easy part. Most of our work focuses on what surrounds it: the safeguards that block questions outside its scope of expertise, the systematic evaluation of responses before each production release, continuous monitoring once online. This is far less spectacular than a demo, but it is exactly what separates a prototype from a banking product.
Around this assistant, AI also works in the background in the application: automatic transaction categorization transforms a cryptic bank statement into readable information and forms the basis of all the expense analysis that the client sees on their screen.
Our current developments are driven by one motto: move from reactive to proactive.
Today, the assistant responds when asked. Tomorrow, we want it to proactively flag what deserves attention: an unusual expense, a subscription increase, an approaching deadline, or an atypical month, for example.
The goal is not to multiply notifications—nobody wants more of those—but to surface the right information at the right time.
It’s important to note that there’s a red line to all this: the assistant illuminates the client’s finances; it does not provide investment advice. Everything related to advice remains the responsibility of our human experts. AI prepares the ground. The relationship of trust, however, is built with real people.
How is AI used internally by Alpian teams?
Every Alpian employee has an enterprise AI assistant, connected to our tools and internal knowledge base with the same security and confidentiality requirements as our banking systems.
Our conviction is simple: we cannot claim to offer clients the services of an AI-augmented bank if our own teams work as they did in 2015.
In customer service, AI handles the first level of support. We have built an assistant capable of answering a large portion of requests, including procedures, product functionality, and frequently asked questions, by drawing on our internal knowledge.
Today, it operates alongside our teams. The next step: making it directly accessible to the client in the application, for an immediate response, at any time, and in their language. And we want to go further: connect this assistant to our internal systems so that it doesn’t just answer, but can also diagnose.
For example, a client explains that they can no longer pay online. The assistant can check the status of their card, note that the “online payment” option is not activated, and explain how to enable it. Always the same architecture: the model understands, our systems verify.
But it is on the side of our technical teams that the transformation is most profound: AI is revolutionizing their way of designing software. Our engineers work with code assistants that no longer just complete a line: they understand an entire project, write complete functionalities, generate tests, propose code reviews, and write documentation.
The profession is shifting: engineers spend less time writing code and more time specifying, architecting, and verifying it. They become the orchestrator of their own work. A disruptive paradigm shift.
Finally, for management, our business departments directly query the bank’s data in natural language, via agents connected to our data repository.
A question that yesterday would have triggered a ticket to the data team now finds its answer in minutes, and the data team can dedicate itself to what truly requires its expertise.
About Damien Chambon
Damien Chambon is Head of Cloud, Data & AI at Alpian. He leads the teams responsible for cloud infrastructure, the data platform, and AI products. He regularly speaks at conferences on the adoption of artificial intelligence.
