Din Sundhedsfaglige A-kasse (DSA) is Denmark’s unemployment insurance fund dedicated exclusively to health professionals. Originally founded for nurses in the 1940s, DSA has expanded to support seven health professional groups, delivering specialised guidance on unemployment benefits, job searching, and career development. 

With 75,000 enquiries per year across multiple channels, DSA needed a scalable way to provide fast, accurate member support without overloading its service team. Puzzel’s Virtual Agent (VA) became central to that transformation. 

The Challenge

By 2021, DSA saw a clear pattern:

• Enquiries were increasing year-over-year.

• Members expected round‑the‑clock support.

• Manual responses and outdated FAQ content made it difficult to keep guidance accurate.

• Their first‑generation chatbot relied heavily on predefined answers, making maintenance slow and leading to inconsistent replies. 

As Taifour Zavari, AI Consultant at DSA, described at Elevate 2026: 

“We were in what I call a scaling hell, constantly categorising new questions, updating manual answers, and trying to keep the bot relevant in a world that changes fast.” 

DSA needed a smarter, more flexible solution that could: 

• Understand diverse member questions 
• Provide accurate responses using current DSA information 
• Scale to support all member types, not just recent graduates 
• Reduce manual maintenance 
• Offer more proactive, human‑like guidance 

The Solution: From Rule-Based to Orchestrated AI

Phase 1 (2021): A Structured but Limited Virtual Agent

DSA launched its first Puzzel Virtual Agent chatbot using manually categorised emails and predefined answers. While innovative for its time, it struggled with: 

• Recognising subtopics 

• Maintaining updated responses 

• Handling varied questions beyond a narrow graduate-focused scope 

Phase 2 (2024): Introducing GenAI + Website Search

In 2024, DSA evaluated how to scale support for all members. They adopted a generative AI model for their Virtual Agent, powered by Puzzel’s search and content retrieval. 

This version of the chatbot: 

• Searched the DSA website for relevant articles 

• Used an LLM to construct natural, concise answers 

• Reduced manual content maintenance 

However, it still relied on a single tool - the search engine - which created constraints: 

• The chatbot misinterpreted nuanced questions 

• Small talk and contextual understanding were poor 

• It repeated answers 

• It couldn’t perform specialised actions (e.g., reviewing a CV or LinkedIn profile) 

As Taifour summarised, “When you only have a hammer, everything starts to look like a nail.” 

Phase 3 (2025–2026): A Multi‑Agent AI System with Orchestration

Recognising the need for deeper intelligence and flexibility, DSA re‑engineered its entire Virtual Agent chatbot ecosystem into a multi‑agent, orchestrated AI model. 

Eight specialised Virtual Agents

Right now DSA has eight different types of Virtual Agents that can handle each case individually. There are no human agents, if the VA chatbot cannot handle the questions it will go directly to the phone. 

Every incoming question is analysed by a central AI orchestrator which routes it to the correct specialist agent - whether unemployment benefits, job search advice, LinkedIn guidance, or career development. 

Each Virtual Agent is trained on dedicated knowledge and instructions, enabling: 

• Higher accuracy 

• More personalised guidance 

• Human-like conversational flow 

Instead of ending each answer abruptly, the chatbot proactively suggests what the user may need next, mimicking a skilled service advisor. 

For example, if a member asks, “When can I fill out my unemployment benefits card?”, the chatbot now: 

• Calculates the exact date (using calendar logic, not just website text) 

• Provides a tailored explanation 

• Suggests the next likely question 

• Keeps full conversation context 

Proactive Intent Detection

One of DSA’s most forward-thinking features is a proactive intent detection system. When a member logs into the chat, the bot extracts key profile data and pulls their full message history with the DSA service team. It then analyses this inbox to generate two suggested questions that are likely to be top of mind for that member before they have typed a single word. This experimental capability, now live in production, helps members get to the right answer faster and enables the chatbot to provide more personalised support from the very first interaction. 

Major Knowledge Base Clean‑up 

The website contained 534 articles, including duplicates, marketing content, and outdated COVID pages. After analysis of real chatbot dialogues, DSA: 

• Removed 300+ irrelevant pages 

• Consolidated content 

• Added missing articles 

• Built a Q&A‑structured knowledge base designed specifically for AI 

This resulted in a 63% reduction in content load making the content in their knowledge base more accurate and more efficient with just 200 high‑quality articles, with more refinements coming. 

The impact

Significantly improved accuracy 

The orchestrated chatbot keeps full context, avoids repeated answers, and offers precise guidance, even when information isn’t explicitly stated on the website. 

Higher member satisfaction through proactive, personalised support 

The “next action predictor” creates a more natural, human experience that guides members through their journey rather than answering in isolation. 

Support for all member groups 

The system auto‑detects member status when logged in, tailoring responses for:

• New graduates 

• Unemployed members 

• Specific professional groups (e.g., radiographers, nurses) 

• Reduced manual maintenance 

By restructuring the knowledge base and removing duplicates, DSA’s team now maintains a much cleaner, more reliable content set. 

Early results: growing member adoption

Since go-live in March 2026, chat traffic has increased by 40% year-over-year for April, a strong early signal. DSA is now tracking whether this growth translates into a measurable reduction in phone and email volumes, which will confirm that the chatbot is successfully deflecting enquiries from high-cost channels. 

Better insights for continuous improvement 
DSA reads real user dialogues to identify where the bot needs fine‑tuning, a qualitative approach that ensures improvements target real issues. Each week, the team samples 50 conversations and reviews them individually to assess quality. Cases that underperform feed directly back into configuration improvements, creating a human-led feedback loop that Taifour prefers over automated sentiment scoring: “Sentiment analysis gives me a number, but I don’t rely on that number. Having someone read the conversations tells me where we actually need to improve.” 

“The real work isn’t just in prompts, it’s in the knowledge we feed the bot. Cleaning, structuring, and consolidating our content made the biggest difference.” 

— Taifour Zavari, Your Health Professional A‑kasse (DSA) 

Looking ahead

DSA plans to continue expanding its specialist agents and exploring new member use cases such as: 

• Automated CV reviews 

• Job application support 

• Personalised career insights 

• Intelligent appointment booking 

With Puzzel’s AI capabilities, DSA is building a truly member‑centric support experience that evolves in real time with user needs.  The next measure of success will be whether phone and email volumes have declined in parallel with this chat growth, data DSA is actively monitoring as April closes.