HEMA has built HAL, an internal AI assistant that lets staff find technical knowledge through chat, an IDE or tools such as Kiro and Claude. The practical change is not another chatbot window: engineers can ask questions about APIs, teams, services and procedures without rummaging through several disconnected portals.
AWS AI Watch analysis
What happened
The Dutch retailer says HAL uses Amazon Bedrock AgentCore, Bedrock Knowledge Bases and the Model Context Protocol to bring structured service-catalog data, documentation, API specifications and event-streaming information into one governed knowledge layer. HEMA says a question that could previously mean searching three or four portals, sometimes for an afternoon, can now be answered in seconds.
The system began as a standalone web assistant and was later opened to external MCP clients. HEMA uses one gateway for the internal agent and a separate Microsoft Entra-authenticated gateway for tools such as Kiro and Claude. Users authenticate through a browser, receive refreshed tokens automatically and do not need AWS credentials on their client.
Why it matters
This is a concrete example of enterprise AI being used to repair an information problem rather than decorate a slide deck. HEMA says HAL is read-only today, with access controlled through existing Active Directory groups. The company plans to develop it into an action layer, which would let the assistant do more than explain how an internal process works.
That next step is where the stakes rise. Finding the right API is one thing; asking an agent to provision access or change a live system is quite another. A helpful librarian is welcome. A helpful librarian with production credentials needs a rather better diary.
Our read
HEMA’s account is a company-authored case study, not an independent assessment of answer accuracy or productivity gains. Even so, it gives developers useful architecture choices: keep knowledge retrieval separate from live API access, use identity controls already understood by the organisation, start read-only and test with both technical and business users before expanding the system.
The most credible part of the story is its restraint. HEMA says the assistant was tested in staging for a month before production, and admits that existing system-to-system APIs are not always well shaped for agents. That is the sort of detail that makes an AI deployment sound less like magic and more like engineering.
What to watch
- Answer quality:
Whether HEMA publishes accuracy, failure and escalation data from production use. - The action layer:
Which operations become available when HAL moves beyond read-only access. - Permission boundaries:
Whether group-based access remains reliable as more tools and data sources are connected. - MCP durability:
How the separate external gateway behaves as clients and internal APIs change. HEMA’s specific relevance to WittyWires is its documented deployment of an AI assistant using Bedrock AgentCore and MCP to connect governed enterprise knowledge to everyday development tools.
Discussion spark: Should enterprise AI assistants remain read-only by default, or is controlled access to routine actions the only way they deliver value beyond search?
Sources and evidence
- From portal-hopping to instant answers: HEMA’s journey with MCP and Amazon Bedrock (23 September 2026, 18:41 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)