Webinar Summary: Admiral Handles Insurance Calls with AI Agents

Admiral handles millions of customer inquiries annually across countries including the UK, Italy, France, and Spain, and is building an AI layer to support this. The goal is to ensure customers can resolve issues immediately regardless of channel or situation, or be seamlessly connected to agents who understand the context.

The key metric is a 90% First Contact Resolution (FCR) rate. While 24/7 service, concise interactions, and high NPS are also pursued, the primary evaluation focuses on whether the customer's actual problem was resolved rather than how many calls the AI blocked. Intentional transfers to agents are not counted as failures.

Given that it is a regulated industry, Admiral established a principle that AI agents will not be deployed into operations until they perform at or above the level of existing human agents. Cases involving vulnerable customers and customers in arrears are routed to high-priority agent queues for the time being, and edge cases that previously relied on the experience of seasoned agents are explicitly validated on a use-case basis.

For organizational adoption, emphasis was placed on engagement and education rather than mandatory directives.

• Skeptical team members were allowed to experience the agents directly, and company-wide demos and training sessions were conducted.
• Consensus was first reached on changes to profit and loss, customer experience, and employee experience, rather than on technology.
• Governance, KPIs, and data standards were established from the outset to avoid reliance on intuition.

Admiral deploys by deeply experimenting with specific topics within one organization and then rapidly spreading the learnings to other organizations. In the first voice use case, the rate of resolution without transfer to an agent improved significantly, while customers recognized that the AI agent was faster.

The first case introduced in the webinar was Admiral Money's early loan repayment quote. Previously, this process took about 5 minutes, involving IVR selection, identity verification, waiting for an agent, backend calculation, and receiving a PDF via email. In the built workflow, identity verification is executed first as a separate sub-agent, and a sentry sub-agent determines whether the AI can handle the call. Vulnerable customers or customers in arrears are automatically routed to human agents.