Länsförsäkringar’s customer-facing AI chatbot illustrates both the potential and the complexity of deploying generative AI in a regulated insurance environment.

The chatbot solution was designed using a multi-agent architecture, incorporating guardrails, domain-specific “area agents”, validation layers, and a tone-control agent to ensure brand consistency and accuracy. This layered approach was critical to managing risks such as hallucinations, inappropriate outputs, and security threats, especially given that customer-facing AI cannot rely on users to verify responses.

After 18 months in production, the chatbot handled approximately four out of five customer interactions, demonstrating strong efficiency gains and enabling 24/7 service. An unintended but valuable outcome was that employees also began using the chatbot to support their own queries, effectively turning it into an internal knowledge tool.

Important lessons around governance and data quality were learnt during the implementation process. AI systems surfaced outdated or hidden content (e.g. contact details on obscure pages), reinforcing the need for strong content management and controlled data access. AI success was found to depend on less on the technology itself and more on data readiness, infrastructure, and organisational change.

Key takeaways for other mutual insurers include prioritising robust guardrails and validation layers for customer-facing AI, being ready to face data and content quality issues in the early stages of implementation, designing for speed and accuracy simultaneously, not just correctness, and recognising that change management (culture, trust, processes) drives most of the value.