The modern insurance adjuster spends a disproportionate amount of their day acting as a digital librarian. They hunt through fragmented PDFs, cross-reference policy clauses against vet invoices, and manually assemble evidence folders before they can even begin the actual work of professional judgment. For years, this administrative friction has been the accepted cost of doing business in the insurance sector, a bottleneck where human expertise is wasted on data retrieval. This week, the conversation in the European insurtech space has shifted toward a different reality, where the preparation phase of a claim is handled entirely by an AI agent before the human expert even opens the file.
The Architecture of a Bottom-Up AI Rollout
Univé, a major Dutch cooperative insurance provider, has moved beyond the experimental phase of generative AI by integrating ChatGPT Enterprise across its entire organizational structure. The scale of the deployment is defined by a fundamental shift in who builds the tools. Rather than relying on a centralized IT department to develop and push a few monolithic applications, Univé empowered its staff to build their own solutions. This approach resulted in the creation of approximately 1,500 custom GPTs, each tailored to solve a specific, granular problem encountered by employees in the field.
This democratization of AI development has permeated every functional area of the company. The deployment is not limited to customer-facing roles but extends deep into the back office, including claims handling, underwriting, finance, human resources, legal, and IT. Even the executive leadership team utilizes these tools to streamline management processes. By treating AI as a general-purpose capability rather than a specific software product, Univé has effectively turned its workforce into a fleet of citizen developers who optimize their own workflows in real-time.
To support this scale, the company integrated OpenAI's enterprise-grade security platform into a rigorous internal governance framework. This framework ensures that the rapid proliferation of custom GPTs does not compromise data integrity or regulatory compliance. The goal was to create an environment where employees could experiment with high velocity without risking the security of the millions of members Univé serves across insurance, mortgages, and financial services. The result is a systemic reduction in task latency, where processes that previously took hours are now completed in minutes.
From Prompting to Agentic Workflows
The true transformation at Univé is not the presence of a chatbot, but the transition from a prompting model to an agentic workflow. In a traditional prompting model, the human is the primary driver: the user enters a command, and the AI provides a response. This still requires the human to know exactly what to ask and where the necessary data resides. An agentic workflow reverses this dynamic. In this model, the AI proactively prepares the work environment, gathering evidence and synthesizing context before the human ever engages with the task.
This shift is made possible by a technical implementation known as connector permission inheritance. Instead of creating a separate, complex set of security rules for the AI, the system inherits the existing permissions of the user. If an employee does not have access to a specific folder or database in the company's legacy systems, the AI agent cannot access it either. This ensures that the AI operates strictly within the boundaries of the user's existing authorization, eliminating the risk of unauthorized data exposure while allowing the agent to seamlessly pull information from across the enterprise.
This governance is reinforced by a multi-layered security stack including enterprise authentication, privacy impact assessments, and continuous monitoring. By aligning the AI's access rights with the company's existing identity and access management (IAM) protocols, Univé has turned security from a bottleneck into an accelerator. Employees no longer spend time requesting special permissions for AI tools; the tools simply know what they are allowed to see based on who is using them. This allows the AI to act as a highly skilled preparation assistant that handles the tedious assembly of information, leaving the human to perform the high-value act of decision-making.
Automating the Insurance Value Chain
In the realm of pet insurance claims, this agentic approach manifests as a complete redesign of the claims pipeline. Previously, a human adjuster had to gather claim files from multiple channels, scrutinize line items on veterinary invoices, and manually check them against the specific terms of the policy. Now, a workspace agent handles the entire assembly phase. The agent aggregates the files, verifies the invoice details, identifies missing documentation, and flags any anomalies that deviate from standard claim patterns.
Crucially, the agent does not decide whether to pay the claim. Instead, it produces a traceable recommendation report that outlines the evidence found and the policy clauses applied. The human expert enters the process at the final stage, reviewing the agent's prepared evidence and making the ultimate determination. This division of labor ensures that the speed of AI is balanced by the accountability of a licensed professional.
A similar transformation has occurred in the underwriting process. Underwriters typically face a queue of applications that must be triaged by risk level. The AI agent now pre-screens this queue, aggregating data from approved internal sources to analyze the risk profile of each applicant. It identifies missing evidence and prioritizes cases that require immediate human intervention. By removing the bottleneck of initial data gathering, the agent ensures that the underwriter's first action is an analytical one rather than a clerical one.
This operational model treats the AI as a partner in the workflow rather than an external tool. The objective is to eliminate the blank-screen problem, where a worker starts a task from scratch. Instead, the worker begins their day with a curated set of cases, each accompanied by a comprehensive dossier of evidence and a suggested path forward. The human remains the final authority, holding the legal and professional responsibility for the outcome, but the cognitive load of the preparation is entirely offloaded to the agentic system.
Univé's experience suggests that the success of enterprise AI depends less on the sophistication of the model and more on the flexibility of the operating model. By combining the autonomy of employee-led GPT creation with the rigidity of permission-based governance, the company has moved from using AI as a search engine to using it as an operational engine. The transition from hour-based work to minute-based work is the direct result of moving the AI from the end of the process to the very beginning.
