Alpes Bank’s journey to creating value at scale with generative AI

Alpes Bank

Alpes Bank ¹ is one of Switzerland’s leading universal banks, renowned for providing high-quality banking services. It has been in business for over a century and employs 4000 people, serving one million customers across Switzerland. While peers shifted focus to digital channels, Alpes Bank doubled down on personal, in-branch relationships as its core differentiator. Though it continues to invest in digital services, the bank remains resolutely branch-centric. Moreover, this stance reflects its customer base, which is, on average, 51 years old and upper-middle class. They value in-person service and are willing to pay a premium for it.

Although customers interact with the bank across multiple channels, the bank’s own study of its channel economics has found that the traditional brick-and-mortar business is still more profitable than its digital channels (see Appendix 1).

This traditional business model, however, has not stopped Alpes Bank’s retail banking division from selectively adopting some artificial intelligence (AI) tools. For example, its customer analytics team has worked with a model that predicts “next best offers.” Similarly, the risk and compliance division has sourced a machine learning solution for predicting credit default. So far, the bank has followed a buy-first approach, sourcing AI applications from external vendors and focusing on integrating such solutions into its existing information technology (IT) infrastructure.

Tamara Maurer’s mandate

Alpes Bank took notice of GenAI when competitors began offering GenAI-driven investment advice at a fraction of its traditional cost. If it worked, this could prove an existential threat to Alpes Bank’s old high-touch model. “If GenAI delivers financial services for cents,” one chairman asked, “how long further can we justify our in-person business approach?”

As a defensive measure, the board launched a standalone venture led by Tamara Maurer, Alpes Bank’s Head of Digital Transformation, to demonstrate how GenAI might be deployed without compromising service levels.

Tamara was empowered to move quickly, bypassing some traditional processes. However, given Alpes Bank’s traditional culture, which prioritizes risk avoidance, she needed to demonstrate clear business value to secure ongoing support. The CEO summarized the criteria for her first GenAI pilot in an email to Tamara: “By quarter-end, I want a no-regret GenAI pilot, something we can roll out quickly, that shows clear business value, and does not break anything.”

Finding a GenAI use case

Tamara invited a handful of data scientists from the business division’s existing data and AI teams. She also brought in colleagues from the IT department to ensure technical alignment. She scheduled a one-day ideation offsite. Although the session unfolded with less structure than initially planned, the whiteboard was quickly filled with ideas. The group ultimately converged on three promising options (see Appendix 2):

• Alpes Bank email Assistant: Customer service agents rely on a single structured document, the 483-page “Business Case List,” to answer many of the customer e-mails they handle. However, accessing this file via the bank’s existing wiki requires exact search terms, making the process error-prone. A retrieval-augmented generation (RAG) ² assistant could instantly draft responses but customer service agents would probably need to verify and sometimes edit the email draft. This partial automation of email writing could reduce the average handling time, ³ which is a closely monitored board-level metric for the contact center.

Further, the participants agreed on guesstimates to evaluate the potential impact and implementation effort of these three options (see Appendix 2). As the workshop concluded, Tamara made a call: “In 12 weeks, I either walk into the executive board meeting with a GenAI pilot, or with explanations.” She pointed at the three shortlisted options and said to her team, “Pick one. Which use case do we commit to?”

No one argued for the internal co-pilot; it would not ship in time. No one wanted the reputational risk associated with a public chatbot. The email assistant was the only option that could potentially fit both the deadline and the bank’s risk tolerance. Tamara circled it in red.

Envisioning a GenAI solution

Tamara’s initial enthusiasm did not last long. By the time she reconvened the workshop participants, the mood had shifted. Some on the team wondered if Alpes Bank could build the GenAI email assistant independently. They admitted they had never worked with GenAI before.

Two days of internet research followed. Vendor websites promised “instant GenAI for customer service.” Still, none survived deeper scrutiny: licenses were vague about Swiss data residency, a strict legal requirement for Swiss banks, and multilingual support rarely extended beyond English. The conclusion was stark: there was no product the bank could simply license; building was inevitable, and so was outside help.

The break came almost by accident. Tamara slipped out mid-week to the “Swiss AI Forum,” more for fresh air than to find a specific GenAI solution. However, she found herself in a coffee queue conversation with Lucien Gétaz, the technical lead at a startup called AILabs. As an AI integration partner, AILabs had just completed a RAG pilot for a Swiss insurer. Intrigued, she invited Lucien for a remote demo.

On Friday afternoon, Lucien fired an anonymized customer email into Alpes Bank’s sandbox environment, ⁵ and the GenAI returned a flawlessly formatted reply in German, French, Italian, English, and to everyone’s surprise, also a decent attempt in Romansh. ⁶ The silence that followed was louder than any applause could have been. Tamara broke it with a single sentence: “We need that.”

The following week, Lucien and his team of AI engineers squeezed into Alpes Bank’s office space for an exploratory workshop.

At the start of the workshop, Tamara wrote the guiding mantra on the whiteboard: “Decision required today - how do we ship a prototype in 12 weeks?” Together, the AILabs and Alpes Bank teams developed a basic framework for leveraging RAG to automatically generate responses to customer e-mails (see Appendix A3 for a description of the envisioned solution). This framework assumed that Alpes Bank would use a commercially available large language model (LLM) in the cloud and that AILabs would perform the required development work to integrate these services with Alpes Bank’s existing IT infrastructure.

Tamara knew that other paths existed (as outlined in Appendix 4, where Tamara chose API access to an LLM via Microsoft Azure), ones that might avoid the long-term dependencies of relying on AILabs and anchoring the solution to a specific cloud and LLM stack. But the 3-month deadline left no room to pursue them. That evening, she formalized the decision in a brief internal memo: “GenAI Pilot v1.0: Conviction Over Perfection.”

Sprint one (Week 1–3): Proving it can work

Tamara initially built an internal core team, recruiting additional data scientists, an experienced business analyst, and a software engineer. Additionally, she reached out to Antoine Bellini, the Head of Customer Service, and asked him to convene a small group of customer service agents. These agents would serve as a sounding board, sharing insights from their daily interactions with customers.

To give structure to the collaboration, Tamara mapped key responsibilities on a whiteboard: Lucien and his AI engineers would lead technical development, focusing on building the RAG pipeline and designing the required prompts. Alpes Bank, in turn, would define the business requirements, establish secure access to internal systems, provide the data AI engineers needed, set up cloud infrastructure, adapt the existing email application, and advise on how real-world customer e-mails should be phrased.

Tamara insisted on using Alpes Bank’s established four-gate project framework (scope, demo, testing, and review). At the same time, AILabs proposed agile sprints organized on a Kanban board, ⁷ a simple, visual method for tracking work progress across columns such as “To Do,” “In Progress,” and “Done.” To accommodate both sides, Tamara designed a framework that combined Kanban for daily work management and Alpes Bank’s formal milestones for executive oversight.

The first major hurdle was data. Alpes Bank’s data scientists extracted 10 years of customer e-mails from the customer relationship management system and moved them to a secure sandbox. Alpes Bank’s team developed simple routines to mask personal information ⁸ in e-mails using tools provided by a cloud vendor that the bank’s legal and compliance team had already approved for another project. However, deeper governance processes were temporarily put on hold. This allowed AILabs to start testing RAG workflows immediately on a set of test e-mails. Responses returned within 10 seconds; some were impressively accurate, others were still off the mark.

Sprint two (Week 4–6): From working to first useful results

Technically, AILabs was moving fast. The AI engineers refined the assistant’s retrieval logic by adjusting how customer e-mails were interpreted and by testing different filtering steps to cull irrelevant cases. They measured improvements by checking how consistently and quickly the GenAI system responded to messages. Updates were regularly shared via Slack. ⁹

A sense of success shaped the tone of the second sprint review. AILabs presented a stable prototype capable of generating replies for standard e-mails. Tamara described it as a promising baseline: the email assistant worked in controlled conditions. From the AI engineering side, the message was clear; this was a good result for this point in the project.

And yet, around the table, a different perspective quietly emerged. The customer service agents who tested the tool during the meeting had a mixed reaction to its capabilities. A few appreciated its speed, while others found the suggestions too generic. Still, nobody framed this as a failure. By contrast, Tamara noted a growing sense of cautious optimism. The prototype had cleared the project’s demo stage, and the team could now shift from proving feasibility to sharpening its usefulness.

At this stage, no one questioned the project’s overall direction, but a subtle tension emerged over what “done” really meant and whether the solution was “good enough.” AILabs measured success by technical readiness, for example, ensuring that the system responded to customer e-mails within 10 seconds and reliably retrieved information. In contrast, Alpes Bank’s team focused on business impact, such as whether the assistant could reduce the time needed to respond to customer e-mails. Meeting one set of criteria did not automatically guarantee meeting the other.

Sprint three (Week 7–10): Reopening what was closed

AILabs flagged a problem that had not been visible earlier. The assistant struggled to interpret the diverse ways customers phrased their requests. Fixing this problem would mean redoing work that Tamara had already scoped, tested, and declared “done.” Some iteration was expected, especially in an exploratory pilot. However, for the bank’s team, this change cut deeper. The RAG pipeline was one of the few elements that had felt complete. It had been reviewed and signed off in meetings, documented in slide decks, and treated as a closed chapter in the project plan. Reopening it felt like breaking an implicit contract, calling the rest of the project into question.

AILabs added the task to the Kanban board. The Alpes Bank’s team updated the schedule to accommodate the delay. But the mood had shifted. It was not just the RAG pipeline. Several other tasks were quietly reshuffled. From AILabs’ perspective, this was normal. Experiments revealed new insights, and subsequent course corrections were made. However, from the bank’s side, the frequent changes made the project feel less stable.

Meetings also began to lose their balance. They shifted from shared checkpoints to technical briefings. Lucien and his AI engineers showcased graphs and detailed test results. To bring the project back into alignment, Tamara requested clearer summaries that translate technical insights into business language.

By the end of the sprint, the GenAI email assistant could generate near-complete replies for many standard customer inquiries. From AILabs’ perspective, the system was approaching deployable quality. One AI engineer summed it up: “It is not just faster. It is more consistent.” Then Tamara asked: “So, is it done?” The question landed with quiet weight. Technically, the assistant was working. But for the bank, “done” meant something more: documentation, signoffs, user testing, and a compliance review.

Sprint four (Week 11–12): Progress, not perfection

AILabs was convinced that the GenAI email assistant was nearly ready for deployment. But deeper issues began to surface. During a review, AILabs shared a new finding: the assistant struggled with a specific subset of customer e-mails. Messages from the Canton of Grisons, ¹⁰ often written in standard German with occasional Romansh words, disrupted the retrieval logic. AILabs’ AI engineers flagged this as a fairness risk: if the assistant performed well with standard German but failed when the input deviated slightly, it could lead to biased response quality. Their recommendation: hold off on the broader rollout until the issue is fixed.

Tamara saw it differently. “If we wait for perfection,” she said, “we will not finish in time to gather the data we need to prove the solution creates business impact. This is augmentation, not automation. The customer service agent remains in control. We will roll out and address the fairness issue in the next sprint. I am not trading momentum for an ideal state we may never reach.”

In the days that followed, a compromise emerged between AILabs’ AI engineers and Tamara. The rollout would extend to more customer service agents but still be framed as “experimental.” A work item for a patch was added to the backlog for a future sprint. The pilot phase formally ended. During the limited rollout that followed, the GenAI email assistant reduced the average handling time for simple inquiries by 21%. Early feedback from customer service agents was also optimistic.

The board meeting

As the initial excitement faded, the email assistant’s limitations became harder to ignore. In a note to Tamara, the CEO praised the pilot’s measurable impact but set a clear condition for its future rollout: without GenAI-specific governance, the bank would not use the email assistant beyond the experimental pilot phase.

The CEO put the pilot’s future on the agenda for the next executive board meeting. The board meeting would come down to two questions: (1) Is the risk under control, and (2) who is accountable for it day to day?

On the morning of the meeting, Tamara walked into the boardroom with a short slide deck. She plugged her laptop into the screen, knowing the meeting’s outcome would determine whether the pilot would be scaled or switched off.

“One week ago,” she began, “a customer service agent opened an email that said: ‘I lost my card’. Our GenAI email assistant produced a full reply in seconds. It was good enough that the agent could stop writing and start reviewing.”

She let that land, then clicked to the next slide. “Later that same afternoon, another customer wrote in. Same issue, lost card, but in German, with a couple of Romansh words mixed in. The assistant still generated a reply. But this time, it was slightly off.”

The room went quiet. “That is the point,” Tamara said. “We are used to systems that behave exactly the same way every time. GenAI does not. It infers. It fills gaps. It is powerful, but it may hallucinate a bit.”

The CEO looked up. “Hallucinates how?”

“With our normal systems,” Tamara replied, “the boundaries are crisp. If something goes wrong, we can trace it to a rule, a workflow, or a line of code. With the GenAI assistant, a tiny change in phrasing can change the outcome, and we cannot always explain why it chose one answer over another. Drafting an email sounds simple, but under the hood, the system is interpreting intent, retrieving internal content, and composing text, all at once.”

The CEO interjected: “Which is exactly why I said we need GenAI-specific governance.”

Tamara nodded. “Agreed. The pilot proved we can create business value with GenAI. Now the question is whether we can control the risk to keep it running. I have taken what we learned during the experimental rollout and turned it into a governance process with specific roles (i.e., Head of AI, Tech Owner, Business Owner, Risk and Compliance, and End-user). This would help us distribute accountability across both technical, business, and compliance functions, ensuring that governance is not one-sided” (see Appendix 5 for details).

The CEO leaned forward and tapped the table. “This is comprehensive. But translate it for us. What does it mean in practice?”

Tamara paused. “It means three things,” she said, choosing her words carefully. “Competent. Compliant. Calculated.”

She held up one finger. “Compliant: we meet regulatory expectations. For example, data handling, auditability, and documented controls.”

A second finger went up. “Calculated: we track business value, where the assistant saves time, and where it might fall short. That also means no convenience use cases. We must focus on use cases where we can measure the effect GenAI has on business value.”

A third finger went up. “Competent: we build the ability to identify and manage hallucinations. GenAI will always hallucinate a little. Hence, the governance question is not, ‘How do we eliminate hallucination?’ The real question is: Where are we willing to carry that risk, and are we able to manage it?”

She looked around the room. “That is why governance cannot sit only with Tech and Compliance. With GenAI, the people using the system are the first line of defense. Our customer service agents must act as co-governance agents: reviewing every draft, rating the quality of the output, and flagging failures. Without that feedback loop, the assistant becomes unsafe. So, the decision today is simple: do we shut this down because it is not perfect, or do we keep it running by building the competence to govern it?”

A new mode of governance

This governance model marked a shift for Alpes Bank. Business units had long been consumers of technology, not stewards. But GenAI changed the rules. With its non-deterministic outputs and context-dependent risks, determining what is wrong could no longer be left to technical staff and compliance alone. Business units, and especially end users, now had to step up as co-governance agents capable of exercising judgment in gray areas.

As board members gathered their materials, quiet conversations filled the room, some energized, others cautious. The CEO approached Tamara, visibly satisfied: “This is exactly what I needed to see. It gives us a way to move forward without breaking anything. But the real work starts now. The business units are not used to owning these responsibilities. We will need to convince, train, and support them.”