Anyone looking into investing for the first time faces significant hurdles: complex technical terms, a wide range of products, and a reluctance to ask seemingly simple questions such as “What is an ETF?”, “What does ‘reinvesting’ mean?” or “How does a savings plan work?”. Younger people in particular are looking for accessible, digital ways to learn at their own pace. To bridge this gap, VisualVest, a digital online broker and subsidiary of Union Investment, has launched a project to develop an LLM-based chatbot.
As a financial institution, VisualVest attaches the utmost importance to compliance with regulatory requirements. The bot must not only convey sound financial knowledge and detailed product information in an understandable way, but must also ensure that every response is legally sound and technically accurate. Particularly when using Large Language Models (LLMs), guaranteeing this precision and reliability consistently and over the long term presents a significant technical hurdle.
For the project, VisualVest sought support to combine generative AI with regulatory requirements. This led to the creation of VestAI, an AI chatbot that has been providing information on the VisualVest homepage since January 2026.
An expert analyses each question and decides which specialist expert is responsible.
Depending on the topic, the query is forwarded to the appropriate experts, such as the product portfolio expert for questions about VisualVest products or the default expert for all queries outside the defined area of expertise
The specialist experts use Retrieval Augmented Generation (RAG) to access fact-checked, curated knowledge content, such as website content, FAQs and product information.
VisualVest’s requirement was clear: the AI chatbot’s responses must always remain within a regulatory framework. At the same time, however, it should utilise the advantages of generative AI – natural conversation, flexible responses and understandable language. How can this freedom be reconciled with a strictly defined legal framework? Working as a cross-functional team, iteratec therefore developed a specialised multi-expert architecture – a system comprising several LLMs, each with specific expertise in a particular subject area. This architecture is robust when dealing with enquiries outside defined areas of responsibility. Here’s how it works:
This provides users with reliable financial information in natural language. The process is fast. The system is continuously optimised based on user and legal feedback.
Large Language Models offer the advantage of being able to interact in natural language. However, it is precisely this flexibility that poses a problem: with natural language, there are an infinite number of possible questions. So how do you ensure that the system remains compliant with regulations even when faced with unforeseen phrasing? In this project, a multi-layered quality approach was developed:
Expert Testing: Legal experts reviewed question-answer combinations in advance to assess potential regulatory compliance.
Red Teaming: Targeted attempts to elicit undesirable responses from the system reveal blind spots.
LLM-as-Judge: A key component is this method, in which an LLM evaluates the outputs of other LLMs against defined quality metrics such as factual accuracy, consistency and traceability.
Human-in-the-Loop: Random checks by subject matter experts continue to be carried out during ongoing operations.
A key success factor was the direct involvement of subject-matter experts in the development of the prompts, particularly from the legal department, who worked hand in hand with the technical AI specialists.
The result: shorter iteration cycles and a deep mutual understanding between subject-matter and technical perspectives. The integration of subject matter experts as co-developers also signals a paradigm shift in product development: the focus is less on pre-defined requirements management and more on genuine collaboration on an equal footing.
VestAI demonstrates that state-of-the-art AI technology and strict regulation are not mutually exclusive. A systematic approach, intelligent architecture and close stakeholder involvement have made the solution possible. The project provides a strong impetus for the responsible use of AI in regulated sectors – from finance and healthcare to the public sector.
