How HEMA Replaced Portal-Hopping With Conversational Enterprise Knowledge Using Amazon Bedrock and MCP

Enterprise developers and internal support teams at century-old Dutch retailer HEMA spent years navigating fragmented documentation portals before querying fragmented internal wikis. According to recent architectural insights shared by the AWS Machine Learning Blog, the organization eliminated this friction by deploying HAL, an internal conversational assistant powered by Amazon Bedrock AgentCore.

Key Takeaways

• Built on Amazon Bedrock AgentCore to unify fragmented internal documentation and operational portals.
• Implements the Model Context Protocol (MCP) to standardize communication between LLMs and enterprise data stores.
• Secures client interactions via Microsoft Entra ID without exposing raw AWS credentials on end-user devices.

Scaling Enterprise Knowledge Retrieval Across Legacy Infrastructure

Internal engineering teams struggled with severe cognitive overhead when searching across isolated knowledge silos, reducing overall developer velocity. The legacy workflow required manual navigation through multiple disparate repositories, increasing latency for simple technical inquiries. HEMA addressed this bottleneck by architecting a centralized agentic layer capable of executing contextual lookups across authorized company systems.

Architectural Integration of the Model Context Protocol and Bedrock

The core engineering challenge involved exposing secure enterprise data tools to large language models without compromising corporate perimeter security. By leveraging the Model Context Protocol (MCP), HEMA established standardized client-server interfaces that decouple the LLM reasoning engine from underlying data connectors. The following table highlights the architectural evolution from legacy portal-hopping to the new MCP-driven Bedrock agent topology:

Securing Client Workflows With Microsoft Entra ID Federation

Security compliance demanded that user identities be strictly maintained throughout the entire inference and tool-execution lifecycle. HEMA anchored all authentication procedures in Microsoft Entra ID, ensuring that the AI agent only retrieves documents and executes tool calls authorized for the specific user context. Client applications communicate with the Amazon Bedrock infrastructure without holding long-lived AWS IAM keys, drastically minimizing surface area for potential credential leakage.

Engineering Takeaways for Enterprise Agentic Deployments

Deploying production-grade enterprise agents requires treating tool connectors as first-class microservices rather than ephemeral prompt wrappers. Standardizing on protocols like MCP allows engineering organizations to swap out underlying vector databases or backend APIs without rewriting orchestration logic in Amazon Bedrock AgentCore. Organizations aiming to eliminate developer portal-hopping must prioritize secure identity propagation and modular tool abstractions from day one.