#retail#data#team

One Knowledge Layer: How HEMA Ended Portal-Hopping with MCP

HEMA's Bedrock AgentCore and MCP case shows why agent response speed depends on company system integration, not model choice.

  • The problem: the knowledge exists, but access does not
  • HEMA's approach: an access layer, not a copy of the data
  • How it works technically
  • The model is replaceable; the architecture is not.

Case

AWS published a case study on HEMA, a Dutch retailer with a 100-year history and 750 stores across several countries. For years, the company's engineers, product owners, and business analysts searched for answers to simple questions by jumping between wikis, service catalogs, and IT portals. HEMA replaced this route with a unified knowledge layer built on Amazon Bedrock AgentCore and the MCP protocol.

The trigger is one customer's case, but the broader point is this: the speed at which an employee receives a usable answer (TTU, time to use) is determined by how the agent accesses data sources, not by which model the company purchased.

The problem: the knowledge exists, but access does not

In most companies, the required information physically exists—in a knowledge base, ticketing system, procedure, or code. The problem is the route: an employee must remember which of five systems to search, open it, formulate a query, and match the result to the task context. Every extra portal means minutes multiplied across hundreds of employees and thousands of questions a day. HEMA saw this as friction accumulating across 750 locations and dozens of roles.

HEMA's approach: an access layer, not a copy of the data

The HEMA team did not consolidate everything into a new central knowledge base: that would create a second copy of the data that would need synchronization and could not be trusted by default. Instead, the Bedrock AgentCore agent accessed existing systems through MCP: the protocol describes which tools and resources are available to the agent in each system, while keeping access rights and data in place, at the source.

How it works technically

An MCP server wraps each system—a wiki, service catalog, or IT portal—as a set of tools with clear descriptions of what can be asked and retrieved. The AgentCore-side agent receives an employee's question in natural language, determines which MCP servers are relevant, queries them in parallel, and assembles an answer with source links. The agent selects the system for each request in seconds, so the employee does not need to click through browser tabs.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

The model is replaceable; the architecture is not.

  1. The same AWS package included an article on running an AI coding agent on open models through Bedrock: companies could keep data within their perimeter and switch models to match the task—from data residency requirements to price per token.

  2. GPT-6 Sol and GPT-6 Luna launched at the same time—a pair of models differentiated along the axes of intelligence and efficiency, so each task could be assigned to a model according to its actual complexity rather than handled by one universal model.

  3. The trend is clear: the model is a configuration parameter of the agent system.

  4. The foundation is how the agent connects to the company's data.

The real engineering lies in sizing and response cost.

  1. AWS also showed the third piece of the same picture: concurrency sweeps for selecting an inference endpoint configuration.

  2. Without systematic testing, a company either overpays for idle GPUs or cuts capacity, creating request queues and increasing user latency. A concurrency sweep runs load tests across combinations of instances and parameters to find the point where cost and latency are balanced.

  3. It looks like a single number in the configuration.

  4. Behind that number are dozens of load-test runs and an understanding of the specific product's load profile.

What This Means for Business

  1. Changing models every quarter to follow the next release gives a company the right to say it uses the latest model.

  2. This does not change response speed for employees.

  3. Speed comes from building an agent access layer for internal systems: an MCP wrapper around CRM, a ticket tracker, 1C-Bitrix, or Pimcore; a gateway that separates permissions and logs the agent's access to sensitive data; and an inference configuration tuned to the actual load.

  4. That is the engineering: the route between an employee's question and the system where the answer resides.

  5. The model is a secondary detail in this route.

Conclusion

HEMA built an MCP layer over its systems and reduced the journey from question to answer from portal wandering to seconds. Companies that invest in such a layer will continue to get results through any model change in the market. Companies waiting for the next announcement will keep jumping between browser tabs.

Discuss the article: A unified knowledge layer: how HEMA solved portal hopping…

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Related Cases

All cases

Articles on the topic

Integration roundup

Related videos

All videos
4:46 Chestny Znak and 1C integration | Automatic product labeling 1CIntegrations
5:14 Integration via REST API | Fewer errors and faster processes APIIntegrations
5:14 API architecture and service brokers | Effective integrations APIIntegrationsArchitecture

News on the topic

All news