Munich Re

The organisation

Munich Re is one of the world's largest reinsurers, providing risk cover to primary insurers, corporations and public sector clients across more than 160 countries. With a history spanning well over a century, the Munich-based group combines deep actuarial and underwriting expertise with an active investment in digital and AI-driven solutions, from automated underwriting platforms to data-led approaches for managing climate and catastrophe risk.

The challenge

In large-scale reinsurance, claims and renewals decisions draw on a vast body of dedicated knowledge: policy wordings negotiated years earlier, treaty terms across multiple cedants, regulatory guidance from several jurisdictions, and decades of case files. That knowledge sits across document stores, SharePoint libraries, core systems and individual experts' heads, and the typical research cycle takes weeks before a claim can even be considered for settlement, extending reserve exposure and risk on the balance sheet. Munich Re wanted to change that economics with AI, but a regulated insurance tenant cannot absorb ungoverned experimentation, six prototype teams each needed to build on the same secure, compliant foundation rather than inventing their own.

The solution

The delivery team was engaged to provide a Governance and Security Architect focused on Power Platform and Copilot Studio, working alongside embedded developers across six parallel prototyping teams, each targeting a different knowledge source and use case in claims and renewals research. The work ran on two tracks in parallel. From the architecture seat, the team owned tenant-level governance and security for the wider rollout: environment strategy, data loss prevention, connector policy and knowledge source classification, the guardrails every prototype would inherit. From the delivery seat, Copilot Studio developers joined the six prototyping teams to build hands-on, working across Power Platform and Copilot Studio to bring each concept to life. The use cases were drawn from the parts of the claims and renewals cycle where research time turns directly into reserve exposure: locating the governing policy wording and endorsement history for a notified loss, reconciling treaty terms and exclusions across multiple cedants, assembling precedent from decades of settled and declined case files, checking jurisdiction-specific regulatory guidance before a coverage position is taken, pulling renewal history and loss experience for a treaty coming up for negotiation, and summarising submission documentation ahead of underwriting review. In each case the agent does not decide anything. It finds, cites and summarises, and the underwriter, claims handler or actuary keeps the decision. That distinction is what made the governance model work in a regulated tenant. Every answer carries a citation back to an approved source document, so a coverage position can be traced to the wording it came from. Agents are grounded only on classified, permitted knowledge sources rather than on the open tenant, personal and cedant-confidential data is fenced by environment and connector policy, and conversation logging gives compliance an audit trail of what was asked and what was returned. Governance came first by design: every team was given a pre-approved authoring environment, a defined set of knowledge sources, and a publishing path through security and compliance review, so subject matter experts and developers could focus on agent logic and user experience because the safe boundary was already drawn. The outcome was not six disconnected experiments but one governed platform with six teams building safely and quickly on top of it.

How we ran it

1. 01

   Establish governance

   Set tenant-level environment strategy, data loss prevention, connector policy and knowledge source classification before any prototype went live.

2. 02

   Define guardrails

   Gave each prototype team a pre-approved authoring environment and an approved set of knowledge sources to ground agents on.

3. 03

   Build in parallel

   Embedded Copilot Studio developers into all six prototyping teams to deliver agent logic and user experience hands-on.

4. 04

   Review and publish

   Routed every agent through a security and compliance review path before publishing, keeping delivery fast without compromising control.

The result

Munich Re now has six AI knowledge prototypes running on a single, governed Copilot Studio foundation, with research that previously took weeks compressed towards minutes for the global subject matter experts using them. Relevant precedent and policy detail is surfaced rather than missed, and because every answer is cited back to an approved source, a coverage position can be evidenced rather than asserted. Shorter research cycles shorten the distance between notification and a reasoned reserve or settlement decision, which is where the balance sheet benefit sits. Every agent is confined to approved, classified knowledge sources, the rollout is secured and auditable at tenant level, and Munich Re holds a repeatable pattern for extending AI into further claims, underwriting and renewals workflows with confidence.