AI in insurances: A case study with BVAEB

ONTEC AI developed a system for the insurance institution that makes querying the internal data pool highly efficient.

BVAEB Speeds Up Insurance Case Processing with AI-Powered Search

BVAEB processes insurance cases based on extensive internal rulebooks that staff need to consult constantly in daily work. Together with ONTEC AI, BVAEB implemented an AI solution that guides employees directly to the relevant document passages and speeds up research, without replacing professional judgment.

At a Glance

• Customer: BVAEB Insurance (>1.3 million insured members)
• Problem: Time-consuming search through complex internal directives (tables, technical terminology, multiple versions)
• Solution: RAG-based AI search with source-document references, table processing, and terminology adaptation (not a generic chatbot)
• Approach/Go-live: Proof of concept (~1,700 documents), rigorous quality assurance (up to 200 test questions per cycle), rollout in early 2026 plus enablement program
• Benefits: Faster research, high accuracy, faster onboarding of new employees

About the Customer

BVAEB is Austria’s insurance institution for public sector employees, railway workers, and mining employees. It provides health, accident, and pension insurance for more than 1.3 million insured members and dependents across Austria. From birth through old age, BVAEB supports its members’ health through preventive care and health promotion programs. The organization is structured across several specialist departments, each with distinct areas of responsibility.

Starting Point and Challenge

In day-to-day case processing, including reimbursements and benefit decisions, employees regularly need to consult extensive internal rulebooks and guidelines known as “Dienstverfügungen” (service directives). These documents present several challenges:

• High document volume, making manual research time-consuming.
• Complex, detailed, and sensitive content, including medical terminology and organization-specific knowledge, that demands high precision.
• Tables that are hard for machines to parse as a central structural element, risking loss of semantic meaning (assignments, conditions, exceptions) during naive extraction.
• BVAEB-specific terminology that standard AI models don’t recognize without context, complicating search and interpretation.
• Frequent updates and multiple document versions, requiring the correct, current version to be reliably found and cited.

The core challenge was therefore:

• to quickly locate relevant passages
• interpret them correctly
• provide traceable and verifiable references to source documents
• and reduce time spent and error rates. This also included supporting the onboarding of new employees.

Objective

BVAEB wanted an assistance solution that would guide employees significantly faster to the right sections of its guidelines, without replacing professional accountability. Precision, answer quality, and traceability of information were the top priorities.

The Solution: “KIWI”, AI-Powered Search with a RAG Approach and Domain-Specific Extensions

Together with ONTEC AI, BVAEB built “KIWI” (German for “Artificial Intelligence for Knowledge Management”).

KIWI delivers answers in seconds with full traceability, but it doesn’t make decisions.

Technically, KIWI is built on a Retrieval Augmented Generation (RAG) approach and was deliberately designed not as a generic chatbot but as a clear, structured workflow:

Ask a question -> Get an answer -> Get guided directly to the relevant source passages

What Makes KIWI Different

KIWI isn’t a standardized, off-the-shelf RAG solution.

KIWI includes several capabilities co-developed specifically for BVAEB, combining ONTEC AI’s modular platform with targeted extensions built for this use case.

• Table processing as the key challenge: A large share of the relevant rulebook logic is embedded in tables. ONTEC AI developed specialized processing that transforms tables so that conditions, assignments, and exceptions remain semantically accurate and searchable.

• Domain language and organization-specific terminology: BVAEB-specific terminology was deliberately incorporated so that answers are professionally relevant and search performs reliably.

A hybrid search approach was implemented, combining semantic search (via vectors and LLMs that capture meaning and context) with classic lexical search (keyword matching), ensuring both contextual similarity and exact-term matches are reliably covered. This is especially valuable when purely semantic methods fail to capture domain-specific terms or phrasing. The combination makes results more robust.

• Safeguards against AI hallucination: Alongside prompt engineering (mandatory citations, answer formatting, handling uncertainty), KIWI benefits from the model-agnostic hallucination prevention built into the ONTEC AI platform.

Every generated answer is automatically checked against the source text by calculating the semantic match between the answer and the original document. If this falls below a defined threshold, the answer is discarded. The system only responds when the information is actually supported by the document. This ensures KIWI delivers only relevant, valid, and useful information, making it more reliable than the underlying language model alone.

• Integration with the existing document repository: KIWI accesses documents directly within BVAEB’s existing repository. The AI comes to the documents, not the other way around.
• Automated quality assurance and iterative optimization: Multiple test and feedback cycles were run with power users under strict acceptance criteria, including expected documents and passages.

An automated regression test suite was also built. For each test question, both the expected source document and the expected facts of the answer are stored. Every time the system changes, whether in search, document processing, or prompts, this suite runs automatically to ensure quality standards are maintained.

• Governance: responsibility stays with people: KIWI is designed as an assistant for finding information and orientation, not as an autonomous decision-making system. Feedback loops and accountability structures were mapped out to ensure quality assurance is organizationally embedded.

Project Journey and Rollout

Project Team

The project team combined ONTEC AI’s team with BVAEB’s team, including the digitalization department, project management, and power users and end users from specialist departments for testing and professional validation.

Development

Development was iterative and closely coordinated with BVAEB. It began with a proof of concept (POC) in which approximately 1,700 documents were reviewed and analyzed.

A key part of the implementation was prompt engineering: instructing the model on how to handle document information, such as answer format, mandatory citations, and handling uncertainty. This was complemented by the model-agnostic hallucination prevention integrated into the ONTEC AI platform, which automatically validates every generated answer against source documents and only allows verified information through.

The resulting solution was then validated and refined through repeated test and feedback cycles with power users.

A rigorous quality process was applied, with up to 200 test questions per cycle defined and verified, each including the expected source passages. These test questions were also converted into an automated regression test suite that ensures answer quality is maintained with every system change.

Beyond search and answer quality, the user interface and practical day-to-day usability were also progressively improved.

Rollout and Supporting Measures

The rollout took place in early 2026 and was accompanied by extensive training: content on an internal learning platform (covering how the system works, plus examples of good and bad questions), an internal handbook setting expectations (“what it is and what it isn’t”), and a course with a quiz. Feedback was established as a core part of ongoing improvement.

Results and Business Impact

KIWI helps employees process insurance and reimbursement cases more efficiently by making it faster to find and cite relevant sections of the rulebooks. Key results include:

• Faster document search: The hybrid search solution reduced the time spent searching for documents, freeing employees to focus on their core tasks.
• Higher accuracy: Combining LLMs with traditional search methods ensures more accurate results and minimizes AI hallucinations.
• Faster onboarding for new employees: The system shortened onboarding time, since new hires can independently and easily find the most relevant documents.
• Higher user satisfaction: Iterative testing and feedback optimized the solution, with power users reporting high satisfaction with the tool.

Key Takeaways for Other Organizations

Organizations building AI-powered knowledge management can draw several practical lessons from BVAEB’s project:

• Focus the use case clearly. BVAEB deliberately avoided a “general-purpose chatbot” in favor of an assistance tool that quickly guides employees to the right rulebook passages, supporting traceability and accountability.
• Document reality determines success. When critical knowledge is largely embedded in hard-to-parse tables, a standard RAG approach often falls short. Targeted, domain-specific processing is needed to preserve semantics like conditions, assignments, and exceptions.
• Take domain language seriously. Internal terms and abbreviations must be systematically accounted for, or match quality and user acceptance will suffer.
• Quality doesn’t happen by accident. Rigorous, repeated test cycles with power users, clear expectations including expected documents, and a structured feedback process are essential levers for reliably embedding the system into daily work.
• Enablement is its own success factor. Training, examples of good and bad questions, handbooks, and a feedback culture accelerate adoption and reduce misuse, especially in organizations with high compliance and quality requirements.