Contents 8 sections
The Legacy System AI Paradox
Your legacy systems hold decades of business logic and data that nothing else can replicate. They also predate modern APIs and were never designed for AI. So AI initiatives stall at the same wall: the model cannot reach the data that matters.
The Model Context Protocol (MCP) removes that wall. It is a standard interface that lets AI work with the systems you already run, without a rip-and-replace modernization project.
What this means for you: if your ERP or line-of-business database can answer a query today, an MCP server can put that answer in front of an AI model. We have shipped more than 200 deployments since 2014. This is the fastest path from AI pilot to AI in production I have worked with.
What Model Context Protocol Actually Is
MCP is an open standard introduced by Anthropic in November 2024. In December 2025, Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation backed by OpenAI, Block, Google, Microsoft, AWS and Cloudflare. It is no longer a single-vendor bet. It is the default way AI agents connect to external systems in 2026.
An MCP server exposes two things to a model: tools, operations the model can invoke through a strict input schema, and resources, data the model can read. The model discovers both at runtime. Authentication lives in the server, never in the model, and the current stable spec (2025-11-25) tightened the OAuth story considerably.
How MCP Connects AI to Legacy Systems
Before MCP, every legacy-to-AI integration was custom work: a wrapper API, a security layer, data transformation, retry logic, then documentation nobody maintains. Months per use case. Brittle on every system change.
With MCP the pattern is fixed. Deploy one MCP server in front of the legacy system. Configure permissions. Any MCP-capable agent can now use it, and one integration serves every AI use case that follows.
The legacy system itself does not change. The 20-year-old ERP stays where it is, behind its existing security controls. The MCP server is the only thing that talks to it. That isolation, more than the speed, is why security teams approve these builds.
Have a legacy system in mind? Describe one workflow to us at querynow.com.
A Worked Example: Order Status from a Legacy ERP
Here is the shape of a real build. The workflow: support staff need order status from an ERP that only speaks SQL on a private network. Today they file a ticket and wait.
- Step 1: stand up an MCP server next to the ERP. It exposes one tool,
get_order_status(order_id), backed by a parameterized read-only query. - Step 2: wire authorization. The server holds the database credentials. The model never sees them.
- Step 3: connect the agent and run evaluation cases built from real orders. Fix what the tests surface.
- Step 4: add audit logging, then run the acceptance tests we signed before the build started.
One tool, deliberately. Narrow tools with strict inputs keep AI access to production data safe and auditable. You expand tool by tool.
Proof This Cadence Works
For a European pharmaceutical regulator, we built a compliance scanner that checks 620+ content assets against 11 rules in about two minutes. The manual review it replaced took two to three hours per pass. For Rockwell Automation, we made 28,000+ assets findable across 80+ countries and improved findability by 60%. Both builds ran against systems the client already had. Neither required a modernization program.
Implementation Pitfalls I Watch For
Legacy performance: old systems were not built for the query patterns agents generate. Cache reads and keep tool queries narrow.
Data quality: AI reads at a scale humans never did and surfaces every inconsistency. Budget for cleanup.
Over-broad access: the tempting shortcut is one tool that runs arbitrary SQL. Refuse it. Scope each tool to one job and log every call. Our deployments are built to SOC 2, HIPAA and GDPR standards.
Stalled governance: a single named tool with a strict schema and an audit trail gets approved by IT. "AI access to the database" does not.
Start with One Workflow, Not a Platform
The protocol is standard, so your first integration is not throwaway work. Everything you add later speaks the same language.
If you are planning an AI program, talk to our team at QueryNow.
If you want agents connected to your own systems this way, see our AI agent services.
QueryNow leads AI transformation for enterprises in the US and Europe. Founded in Germany in 2014, we build the foundation enterprise AI runs on, and the agents that run on it, for pharma, healthcare, manufacturing and financial services organizations on Azure and AWS.
FAQ
Frequently asked questions
What is MCP in healthcare?
MCP (Model Context Protocol) in healthcare connects legacy systems like EHRs, PACS, and lab information systems to modern AI models without replacing existing infrastructure. It acts as a universal translator, allowing LLMs to securely query patient data, clinical records, and research databases while maintaining HIPAA compliance.
How can I migrate from custom integrations to MCP?
Migration from custom integrations to MCP follows three phases: inventory existing APIs and data flows, deploy MCP adapters alongside existing integrations (parallel running), then gradually route traffic through MCP while decommissioning custom connectors. Most enterprises complete migration in 4-8 weeks per system.
What MCP solutions offer built-in security for enterprise AI?
Enterprise MCP solutions include Azure MCP Server (integrated with Azure AD and RBAC), custom MCP gateways with OAuth 2.0 and mTLS, and managed MCP platforms with audit logging, data masking, and compliance controls. All production MCP deployments should include encryption at rest and in transit, plus detailed access logging.
The Monday digest
The three posts worth your time each week, one paragraph each. From the people building production AI.



