OUR CLIENT
Our client is a globally operating financial services and insurance group with a long-standing market presence and a comprehensive product portfolio spanning life, non-life, health, and investment solutions. Serving both individual and corporate clients worldwide, the organization operates across a broad financial ecosystem that includes asset management, pension planning, and banking-related services.
The company places strategic emphasis on long-term value creation, responsible business practices, and sustainable growth. As a regulated financial institution, it operates under stringent compliance requirements including the EU AI Act, demanding that any AI initiatives meet the highest standards of governance, security, and auditability.
THE CHALLENGE
As AI adoption accelerated across the financial services industry, the company recognized the need to move beyond siloed pilot projects and establish a coherent, enterprise-grade foundation for deploying AI agents at scale. The challenge was not simply to adopt large language models (LLMs), but to do so in a way that was architecturally sound, operationally sustainable, and compliant with internal and external regulatory obligations.
The company needed a strategic blueprint – a production-ready Agentic AI environment that could integrate with existing infrastructure, scale responsibly, and deliver measurable business outcomes across claims, underwriting, customer service, and analytics.
Key challenges included:
- Lack of a unified architecture capable of supporting AI agents in both cloud and on-premises environments, alongside existing Microsoft Azure, SAS, and Google Cloud Platform (GCP) investments.
- Absence of clear guidance on agentic design patterns, agent orchestration frameworks, and multi-agent collaboration strategies suited to complex insurance and financial workflows.
- Uncertainty around LLM selection, embedding models, and vector database infrastructure required to power document-intensive use cases such as policy retrieval, claims analysis, and underwriting support.
- Growing need to address AI governance, cost management, and compliance with the EU AI Act, which mandates formal AI risk management and lifecycle oversight.
- Limited internal visibility into trade-offs between cloud-native and on-premises AI model serving, including hardware, performance, and cost implications of CPU versus GPU inference.
Data architecture for AI agents
OUR SOLUTION AND APPROACH
Algomine delivered a comprehensive Agentic AI architecture recommendation covering all layers of the platform, from foundational AI model infrastructure through to application-level deployment, governance, and evaluation.
Our approach was structured, vendor-pragmatic, and deeply aligned with the client’s existing technology investments.
1. Reference Architecture Design
We designed a layered Agentic AI reference architecture built on Microsoft Azure as the primary cloud platform, with defined integration points for GCP and on-premises environments. The architecture encompasses six distinct layers: AI model and services, agent orchestration, data (structured and unstructured), application hosting, developer tooling, and governance. Each layer was assessed against the client’s hybrid infrastructure requirements, including Azure AI Foundry, Microsoft Fabric, Copilot Studio, Azure OpenAI, SAS, and Google Vertex AI. Special attention was paid to the architecture’s openness, ensuring the platform could support models from multiple providers, including open-source alternatives and locally hosted instances.
2. Agentic Design Patterns and Framework Selection
Algomine evaluated and recommended five core agentic design patterns suited to the client’s use cases: Reflection, Tool Use, ReAct (Reason + Act), Planning, and Multi-Agent Collaboration. For each pattern, we mapped specific insurance and financial services scenarios, enabling the client to select the appropriate architecture depending on task complexity, autonomy requirements, and data sensitivity. Framework recommendations were provided for both cloud-native deployments (Microsoft Agent Framework, LangChain/LangGraph, Google ADK) and on-premises use cases (LlamaIndex, CrewAI), with clear criteria for when to use code-first versus low-code approaches such as Copilot Studio.
3. LLM and Embedding Model Strategy
We delivered a comprehensive evaluation of available large language models, covering Azure OpenAI (GPT-4o, GPT-4.1, GPT-5), Google Gemini 2.5 Pro, and open-source models including Meta Llama, Mistral, and DeepSeek. Model selection guidance was grounded in benchmark data (MMLU, HumanEval, HellaSwag) and aligned with the client’s latency, quality, and cost requirements. For embedding infrastructure, we recommended text-embedding-3-large for English and multilingual workloads, and Qwen3-Embedding for Polish-language document processing, paired with Azure AI Search as the primary vector database and guidance on alternative solutions such as Weaviate and Qdrant for self-hosted deployments.
4. Data Architecture for Agentic AI
A dedicated data architecture was designed to support both structured and unstructured data needs. For document-intensive workflows, we specified a RAG (Retrieval-Augmented Generation) pipeline architecture using vector embeddings, hybrid search, and semantic chunking strategies. For structured data, we recommended the creation of AI-specific data marts within the client’s existing data warehouse, enabling LLMs to safely consume transactional and operational data without exposing complex SQL schemas directly. Integration connectivity was addressed through Microsoft Fabric’s Data Factory, self-hosted integration runtimes for on-premises sources, and real-time streaming via Azure Event Hubs.
5. On-Premises GPU/CPU Inference Guidance
For workloads requiring data sovereignty or latency-sensitive on-premises inference, Algomine provided detailed guidance on serving frameworks, including vLLM (recommended for production throughput), SGLang (recommended for complex multi-step agentic reasoning), and Ollama (suitable for prototyping). We also delivered a practical cost model for GPU infrastructure, quantization trade-offs (INT4, INT8, FP16), and memory calculations to help the client evaluate build-vs-buy decisions for on-premises model hosting.
6. Governance, Evaluation, and AI Act Compliance
Given the client’s regulated operating environment, Algomine embedded AI governance as a first-class concern throughout the architecture. We mapped Microsoft Purview as the central governance hub for data lineage, agent lifecycle oversight, prompt logging, and compliance monitoring across all deployed agents. A structured evaluation framework was defined for both performance (using RAGAS, DeepEval, and LLM-as-a-Judge approaches) and security (using Microsoft AI Red Teaming and PyRIT for adversarial testing). The governance model was explicitly aligned with EU AI Act requirements for AI risk management systems (AIMS), ensuring the client could demonstrate control, auditability, and traceability of AI-driven decisions.
THE RESULTS
The engagement delivered a production-ready strategic and technical foundation enabling the client to accelerate responsible AI adoption across its business units.
Key outcomes include:
| Area | Key Outcome |
|---|---|
| Architectural Clarity | A fully documented, layered Agentic AI reference architecture aligned to the client’s hybrid Azure and GCP environment, reducing time-to-design for future AI initiatives. |
| Faster Agent Deployment | Clear framework selection criteria and design pattern playbooks enabling engineering teams to build and deploy AI agents without starting from first principles on each project. |
| LLM Cost Optimization | A model selection and consumption strategy (API-first, Pay-as-you-go) that avoids unnecessary infrastructure investment while maintaining flexibility to adopt new models as the market evolves. |
| Regulatory Readiness | A governance architecture aligned with EU AI Act requirements, providing the client with a defensible audit trail, data access controls, and agent lifecycle management from day one. |
| Data Enablement | A structured RAG pipeline and AI data mart design enabling agents to work securely with both unstructured policy documents and structured operational data at enterprise scale. |
| Security Assurance | Integration of AI red teaming and security evaluation frameworks (PyRIT, Azure AI Foundry) into the development lifecycle, proactively addressing prompt injection and compliance risks. |
RAG Process Scheme
EXECUTIVE SUMMARY
By partnering with Algomine, this leading global financial services and insurance group gained more than a technical document. It gained a strategic foundation for enterprise AI transformation.
The Agentic AI Architecture Recommendation delivered by Algomine provides the client with a clear, actionable path from AI experimentation to production-scale deployment. By addressing architecture, framework selection, data infrastructure, LLM strategy, on-premises inference, and governance in a single integrated deliverable, we eliminated ambiguity and enabled aligned decision-making across technical, compliance, and business stakeholders.
The result is a client that is now positioned to build AI agents confidently, govern them responsibly, and scale them sustainably, across claims processing, underwriting, customer service, fraud detection, and beyond.
CONTACT US
Ready to build your enterprise AI architecture? Get in touch with us, schedule a meeting where we will showcase the full potential of Agentic AI for your organization.