How HEMA Built an Internal AI Knowledge Layer With MCP and AgentCore
The Dutch retailer connected governed Bedrock knowledge bases and internal APIs to chat and developer tools while keeping external clients off AWS credentials.
Edited by Tyronne Panaino
HEMA and Amazon Web Services described the retailer's HAL internal assistant on September 23, showing how a governed knowledge layer can serve both a dedicated chat interface and external developer tools through Model Context Protocol. The AWS case study co-written with HEMA says the 100-year-old Dutch retailer operates more than 750 stores across several countries and built HAL to reduce the time staff spend moving among internal portals, wikis and service catalogs.
The useful lesson is architectural rather than promotional. HEMA did not replace every source system with one chatbot. It kept structured knowledge and live APIs in their existing systems, then added a governed retrieval and tool layer that different clients can call. That makes identity, authorization and the boundary between reading and acting as important as the language model itself.
One knowledge layer, two access paths
HEMA's internal agent runs in Amazon Bedrock AgentCore Runtime and uses Amazon Bedrock Knowledge Bases for documentation and semantic retrieval. For live service-catalog and organizational data, the agent calls tools exposed through AgentCore Gateway. The reported knowledge set includes technology documentation, API specifications, event-stream schemas, data-exchange channels and mappings between people, teams, services and business capabilities.
The internal agent and external clients do not share one gateway. HEMA says the agent-facing path uses AWS identity controls, while a second gateway accepts Microsoft Entra ID authentication for tools such as Kiro, Claude and HAL's web chat. That separation reflects a product constraint described in the case study: one AgentCore Gateway supports one inbound authentication type. It also creates a smaller failure boundary, because the external surface can change without altering the agent's internal gateway.
The client does not receive AWS credentials
For external MCP clients, HEMA placed an authentication proxy in front of the Entra-connected gateway. Users authenticate through the organization's existing identity system, and authorization follows current Active Directory groups. The client receives access to the approved tools without being given AWS credentials.
This is the practical delta from simply publishing an MCP endpoint. The gateway still needs to decide which tools a user may discover and call, while the underlying knowledge and APIs retain their own access controls. HEMA's current deployment is deliberately read-only, limiting the blast radius while the organization evaluates coverage, retrieval quality and cross-role use.
Rollout evidence is promising but first-party
HEMA says it tested HAL with engineers and business users in staging for one month before promoting it to production. The case study reports that questions which could require several portals and much of an afternoon can now be answered in seconds from chat or an integrated development environment. Those outcomes are not accompanied by an independent study, a measured sample or an error-rate disclosure, so they should be read as a customer and vendor account rather than a general performance benchmark.
The next phase is also not live evidence. HEMA plans to extend the read-only layer into carefully scoped actions such as account provisioning and other operational requests. That step will test whether the identity model that controls access to answers can also constrain changes made on a user's behalf.
What other teams can take from the design
The transferable pattern is to separate the knowledge plane, the client interface and the authorization path. A company can expose one governed set of retrieval and API tools to several compatible clients without duplicating every integration, while still using distinct gateways when identity requirements differ. The important checkpoint is not whether the assistant answers a demo question; it is whether permissions, source freshness, failure isolation and audit evidence remain intact when more users and eventually write-capable tools arrive.
Status
Learning. The architecture and rollout are documented in an official AWS case study co-written with HEMA, but the operational outcomes remain first-party, so internal confidence is medium.
Sources
Update note: Last reviewed 2026-09-24. We will revise this post if HEMA publishes measured adoption, reliability or action-layer evidence.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.