Unipol accelerates innovation in the insurance sector thanks to Generative AI

Cluster Reply has created the Foundation Layer – Unipol GenAI Stack, a common technological and organisational foundation that makes the adoption of GenAI coherent and governed, avoiding duplications and providing shared models, data, and guardrails for chatbots, automations, and applications.

The three-tier architecture combines an AI Data Layer with Azure AI Search and Vector DB (over 60,000 documents and 200,000+ metadata) for vector search and contextual generation, a Core AI Layer that orchestrates generative models and agents for use cases such as chatbots, coding assistants, and RPA, and an AI Service Layer that exposes AI capabilities to business applications. The platform offers centralised access to LLMs, unified management of embeddings, semantic indexing, and knowledge base, shared orchestration services, automation (RPA, Process Mining), and application integration, with guardrails for quality, ethics, and compliance; it is already integrated with ServiceNow, Essig, Liquido, UEBA, and FileNet.

Benefits:

- Accelerated time-to-value: release of new applications in 1–3 months thanks to the reuse of common assets and components, with costs under control.
- Modular scalability: extendable to new domains and technologies without impacts on the architecture.
- Governance and compliance: robust control model, fully aligned with the AI Act.
- Data valorisation: vector indexing and semantic access for faster and more accurate access to the knowledge base.

Cluster Reply has developed a Virtual Information Assistant based on GenAI that provides agents and operators with immediate and reliable access to business knowledge, reducing ticket inflow and response times.

Born to manage large volumes of information in the network of over 2,000 agencies, the chatbot integrates a RAG (vector indexing) with a network of specialised agents - document understanding, product disambiguation, checks on business systems, dialogue management - and a chatGPT-like interface with conversational memory and guardrails for ethical and controlled use.
Integrated with the Foundation Layer of the Group, it handles requests in natural language, automatically retrieves relevant content from the knowledge base (over 50,000 UEBA documents and 200,000+ metadata from circulars, manuals, and regulations) and synthesises precise and policy-aligned responses, reducing manual search and openings to the Help Desk.

Benefits:

- Reduction of the load on the back office: significant drop in tickets thanks to conversational automation.
- Improved time-to-resolution: requests fulfilled in a few seconds thanks to semantic retrieval.
- Unified access to knowledge: a single chat interface to consult the entire information asset.
- Scalability and integration: extendable asset to new areas (claims, settlement, administrative) and integrable with other GenAI agents of the Group.

Sprint Reply has supported the Unipol Group in developing a network of AI-enabled agents that convert traditionally manual tasks into conversational interactions: they understand context, validate data in real-time, automate operations on core systems, and, when necessary, direct complex cases to the back office via pre-filled tickets complete with all the information.

The solution combines Azure OpenAI (GPT) for natural language and reasoning, LangChain and LangGraph for workflow orchestration, Python for application logic and integrations, and UiPath for scalable execution (AI + RPA/BPM), ensuring quality, control, and compliance.

Benefits:

- Reduction of tickets & time-to-resolution: lower inflow and faster responses thanks to conversational automation and pre-filled tickets.
- Operational efficiency: streamlining the back office and reducing repetitive tasks with direct execution on core systems.
- Quality, control & compliance: real-time validations, expert involvement when necessary, traceability, and alignment with policies.
- Scalability & integration: extendable architecture to new processes and integrable with AI/RPA stacks and business systems.

Sprint Reply has developed AI Guardrail, a solution based on the asset LoopMind, to ensure monitoring, control, quality and compliance with ethical and regulatory standards. It combines real-time controls with LLM-as-a-judge, human-in-the-loop validations, complete traceability of input/output/evaluations, and a multi-model (proprietary and open-source LLMs) to balance performance and costs.

The solution is a containerized web app with microservices, integrable into the corporate ecosystem via standard connectors or custom developments, supporting various authentication mechanisms, heterogeneous storage, and deployment on-premises or cloud; it offers KPIs and visibility on workflows to uniformly govern all GenAI use cases.

Benefits:

- Control, security & compliance: real-time supervision, centralised policies, and targeted human intervention reduce risks, protect data, and ensure compliance (including the AI Act).
- Quality & reliability of outputs: automated checks and operational reviews ensure accurate, consistent, and ethical content.
- Efficiency & cost optimisation: unified governance, workflow automation, and multi-model strategy improve productivity and spending.
- Traceability & continuous improvement: comprehensive logging enables audits, performance analysis, and progressive tuning.
- Scalability & integration: microservices architecture and standard connectors quickly extend coverage to additional processes and applications.