MCP for the Enterprise: How to Give AI Governed Access to Data

MCP makes it easy for AI to discover and call tools, but enterprises need to decide what data an AI can see and what it can do. Here’s how to build governed MCP tools that keep sensitive data and execution under your control.

Key takeaways:

  • MCP standardizes how AI clients discover and invoke tools, but it doesn’t govern what those tools can access or change.
  • Treat access as two layers: a connection layer that controls who can reach a server, and an execution layer that controls which data, credentials, and side effects a tool can touch.
  • Scoped, purpose-built tools over clean data produce more reliable AI answers than broad access to every system.
  • Filtering sensitive fields inside the tool means the AI never sees them, even when asked directly.
  • Start small, involve IT and security early, and use separate MCP servers to enforce least-privilege access by team or role.

 

Model Context Protocol (MCP) gives clients a consistent way to discover and call tools, removing much of the custom integration work between AI and enterprise systems. But it raises a harder question: what are those tools allowed to reach, and who decides? Most organizations have data that AI should never access freely, and more that can only be used under specific conditions.

The answer is a layer between the AI and the data. This layer can reach the systems where enterprise data actually lives, decide exactly what each tool exposes, and run the work inside infrastructure you control.

FME acts as a governed connection layer in both directions. You can publish workflows as MCP tools, with each workflow defining what data the AI can see, and the MCPCaller transformer lets workflows call tools on other MCP servers.

Let’s take a closer look at how to build governed MCP tools that keep data and execution under tight control.

What MCP solves, and what it does not

MCP has standardized tool discovery and invocation, but it hasn’t solved governed execution. A server advertises each tool’s name, description, and parameters, while the client reads them, picks a tool, and calls it. Nothing in that exchange decides whether a tool returns a whole table or three approved fields, runs with admin or read-only credentials, or can write back to a system of record.

Rule: the contract can be remote, but execution should stay with the data.

The tool definition can be visible to a cloud AI while the work runs inside infrastructure you control. FME Flow can run in the cloud, on-premises, hybrid, or fully air-gapped, and sensitive workloads can keep the model, client, and server entirely on-premises using local models such as Llama or Gemma.

Separate the connection layer from the execution layer

The connection layer decides who the client is and which MCP server it can reach. FME Flow’s MCP Server supports OAuth 2.0 and can map to your identity provider through SAML, so users sign in with existing credentials and roles.

The execution layer decides which identity runs the work, which systems and records a tool can touch, and which side effects are allowed. In FME, the workspace behind each tool defines exactly what is exposed, down to individual fields, and jobs run under FME Flow’s role-based permissions and audit trail.

Keeping these separate lets you grant access to a server without giving the AI access to raw systems. Clients only ever see the tool interface.

Easy data vs. hard data

Many MCP projects start with easy data, such as Slack, GitHub, Google Docs, and other services that already publish MCP servers. The harder problems are legacy databases with no API, geospatial and CAD/BIM data, proprietary formats, real-time IoT and SCADA feeds, and regulated data. That data is hard for four reasons:

  • Representation: binary, spatial, 3D, and streaming data AI cannot parse directly.
  • Locality: data that cannot leave its environment, like a BIM model that cannot leave the building.
  • Computation: answers that need a multi-step workflow, not a lookup.
  • Authority: identity, credentials, policy, and side effects that must be enforced.

Adding an MCP server to every system creates MCP silos, leaving the AI to stitch fragments together. When an answer spans many systems, a workflow that combines and filters the data into one governed result is more reliable.

Purpose-built tools beat broad access

Connecting AI to everything often hurts accuracy. An AI with access to a whole shared drive wanders into outdated drafts and misses the authoritative source.

Take a marketing team that wants AI to recommend webinars for a sales prospect. The information is spread across a website, a planning spreadsheet, and shared drives, and general AI search rarely picks the right upcoming sessions. Two small FME workflows fix this:

  1. A helper workflow builds a clean source of truth. An HTTPCaller pulls upcoming webinars from the website API, a FeatureMerger adds internal attributes from the spreadsheet, and a ChangeDetector updates only changed records in a SQLite database on FME Flow, keeping the data current and updates auditable.
  2. A tool workflow searches it. Up to three keywords plus a start and end date become the tool’s inputs, SQLExecutor transformers run simple date-range and keyword queries, and a NoFeaturesTester tells the AI to try again if nothing matches. The output is trimmed to title, date, and link, built with the JSONObjectBuilder, and returned through the MCP Writer.

The AI client can then run several searches on its own and recommend an accurate shortlist, with no one fielding requests by hand.

Once clean data exists, exposing it as an MCP tool is quick, and the workflows can stay simple.

Extending a copilot beyond its own ecosystem

Microsoft Copilot understands Microsoft 365 content. MCP tools let it combine that content with live APIs, external databases, and legacy systems.

Suppose a city stores permit locations as GeoJSON in SharePoint, a format that is hard to read directly, and needs to notify permit holders affected by road construction. In Copilot Studio, an agent finds the file with native Microsoft connectors and passes its path and a buffer distance to an MCP tool. The tool reads the file through FME’s SharePoint Online connector, pulls a live highways construction feed, and spatially joins them to find affected permits.

A second tool sends the notification emails, with two details worth copying:

  • Template variables in the tool description. The description tells Copilot which placeholders to use, such as a permit number. Copilot drafts the wording and the workflow fills in the values.
  • Differently scoped tools. For privacy or legal reasons, only the second tool can see contact emails, which it joins from a SharePoint list using results the first tool passes forward in FME’s FFS format. It runs on a separate MCP server with elevated permissions, so only authorized users can send outbound email.

FME also connects to Fabric, Azure, and SQL Server, and anything without a dedicated connector can usually be reached through the Microsoft Graph API.

Bringing live ERP data into a data platform’s AI

Snowflake’s agent interface, CoWork (formerly Snowflake Intelligence), works well with tables inside Snowflake, guided by a semantic layer that describes the business. But the catalog there may be a periodic copy, while live pricing sits in an ERP that is behind organizational restrictions and cannot be copied to the cloud. Laws and contracts may also limit where that data can be processed.

FME Flow’s MCP Server bridges the two without exposing ERP internals. Snowflake’s MCP connectors link an agent to an external MCP server over OAuth, tying each session to the signed-in user. The tool workflow is short:

  1. Read ERP records from the underlying Oracle database. Any source would work, since the agent never sees what is behind the tool.
  2. Accept a date parameter so the agent can ask what changed since then.
  3. Drop fields the AI does not need, such as margins, supplier names, and account manager contacts, keeping only SKU, product name, unit price, and last-updated date.
  4. Shape the result into one structured JSON block with the AttributeCreator, Aggregator, and AttributeManager. Any clean, structured format is easy for an AI to parse.
  5. Return it through the MCP Writer, which sends a text attribute back to the client.

Asked to compare catalog and ERP prices and report margins, the agent splits the work across both systems. It cannot report margins, because that field never left the workflow. The tool enforces the boundary, not the AI’s good behavior.

Designing tools an AI can use well

A few best practices:

  • Write thorough descriptions. The AI reads them to decide when and how to call a tool. Cover purpose, output format, and constraints.
  • Expose only the parameters the AI should control. Use defaults for the rest.
  • Return only what is needed. Less noise, less exposure.
  • Handle empty results. A message telling the AI to refine its search beats silence.
  • Choose the right output. The MCP Writer returns text, image, or audio; a JSON writer also works. With neither, a tool returns only success or failure, which suits tools that act rather than answer.
  • Separate staging from production with different MCP servers.

Existing workflows convert easily: add an MCP Writer, publish, and register the workspace as a tool. In the other direction, the MCPCaller transformer lets a workspace call tools on any external MCP server, so you can chain MCP tools together or pull results from other MCP-enabled systems mid-workflow.

Use multiple MCP servers for least-privilege access

FME Flow applies security at the MCP server level and can host as many servers as you need. A finance server can be limited to finance users and tools, while a marketing server exposes only marketing data. Users authenticate against the server for their role, and tools inherit the permissions configured in FME Flow. Each tool also defines its own outputs, so a discovery tool scanning file shares or SharePoint can omit restricted files entirely, and the AI never learns they exist.

Getting started

Often, the hardest part of a first MCP project is confirming that connecting an AI client to a data source meets security and compliance rules. Involve IT and security early, agree on which data and actions are in scope, and set up OAuth and SAML first. After that, tools are quick to add or remove and inherit existing permissions.

Start with a simple validation tool that confirms everything is connected and approved, then grow into a toolset. FME Flow’s MCP Server is available from FME 2026.2, and the MCPCaller transformer from FME 2026.1.

One bonus: MCP requires every tool to describe itself, its inputs, and its outputs. That metadata is written for AI, but it also gives people readable documentation of workflows that rarely had any.

Conclusion

MCP makes connecting AI to enterprise systems easy, but connection is not control. Treat MCP tools as a governed layer: workflows decide what is exposed, execution stays near the data, authentication uses existing identities, and separate servers enforce least privilege. Built that way, AI can work with legacy, spatial, on-premises, and regulated data while you stay in charge of what it sees and does.

Learn more

Safe product icons
Learn FME in 90 minutes. Get started today!

Real change is just a platform away.