Before an AI assistant can answer a business question, it needs a source to answer from. In most companies, those sources fall into two groups. One is the document wiki, such as Confluence or SharePoint. The other is the set of live systems where pipeline, invoices, and inventory sit, such as Salesforce, NetSuite, and SAP. Both are called a knowledge base.
The live systems are the challenge, because each one carries its own permissions and its own audit requirements. With CData Connect AI Gateway, your team can connect both kinds of source to ChatGPT Enterprise, Microsoft Copilot Studio, Claude, or n8n through one governed endpoint built on the Model Context Protocol (MCP).
What AI access to a knowledge base means: document wikis vs. live business systems
The term knowledge base describes two different things, and AI needs a different kind of access to each. Many guides address only document repositories. The live business system is the harder case, because access depends on who is asking.
OpenAI’s connector updates describe the same difference, with synced connectors and access connectors built on MCP. As of October 2026, OpenAI no longer offers administrator-managed sync connectors. Real-time Google Drive, SharePoint, and Microsoft Teams apps replace them.
Comparison point | Document repositories | Live business systems |
Examples | Confluence, SharePoint, Notion | Salesforce, NetSuite, SAP |
What it stores | Written content such as pages and files | Operational facts such as pipeline, invoices, and inventory |
Matching OpenAI connector type | Synced connectors | Access connectors |
How content is fetched | Indexed ahead of time | Fetched when a user asks |
AI access to a knowledge base: the governed ability of an AI assistant or agent to retrieve approved documents and live system records, filtered by the requesting user’s permissions, and use them to answer questions or complete tasks.
Each source has its own API, identity model, and row-level permissions. If retrieval runs through one shared, overprivileged identity, any user may be shown data that user is not authorized to see, even when sign-in at the application boundary functions correctly, according to an Aembit analysis of the Open Worldwide Application Security Project (OWASP) Top 10 for LLM Applications 2026. Access has to be checked on every request.
Managed MCP vs. DIY MCP servers vs. RAG pipelines: three ways to connect AI to your data
Teams usually pick between a do-it-yourself (DIY) server, a Managed MCP, or a retrieval-augmented generation (RAG) pipeline.
MCP: an open protocol that lets AI clients find and call tools and data exposed by an MCP server, with an OAuth 2.1-based authorization framework for remote (HTTP) servers.
Approach | Best for | Permissions and governance | Maintenance burden |
DIY MCP server | One source, engineering-led proof of concept | The MCP specification makes authorization optional. A remote server that protects itself acts as an OAuth 2.1 resource server, so your team must implement OAuth 2.1, audience validation, and logging | High: often one server per source, updated whenever an API changes |
Managed MCP (for example, CData Connect AI Gateway) | Many sources behind one endpoint | Passthrough authentication where the source supports it, role-based access, and audit logging, managed centrally | Lower: the platform maintains the server and connectors as source APIs change |
RAG pipeline | Large sets of unstructured documents, such as policy libraries and manuals | Embeddings are copies of source content and don’t automatically carry source permissions. An Aembit analysis treats them as sensitive derived data | Ongoing: ingestion, indexing, and refresh |
Connect AI Gateway is one of three valid paths. Its case rests on replacing many per-source servers with one governed endpoint over hundreds of sources, so IT reviews one endpoint instead of many. It reports that its efficient MCP toolset reduces token consumption by 20% to 60% compared with endpoint-specific implementations. Its semantic layer also exposes business metrics, dimensions, and hierarchies from live sources, with consistent definitions across tools. The trade off is reliance on a vendor’s connectors and release schedule. A team with one source and engineering capacity may reasonably choose the DIY route.
Prerequisites for secure, permission-aware AI knowledge base access
Single sign-on (SSO) identity provider: An SSO identity provider using OAuth and Security Assertion Markup Language (SAML) must be in place so that each AI request runs as the real user. The MCP security guidance requires servers to reject tokens that weren’t issued for them and explicitly forbids token passthrough
Admin-enabled AI workspace: In ChatGPT Enterprise and Edu, admins use role-based access control (RBAC) to grant permission to create custom MCP connections, and only admins and owners can publish apps to the workspace, according to OpenAI
Clean, current data: Field names must be clean, metadata clear, and sources current
Permission passthrough: a pattern where the connectivity layer authenticates each AI request as the signed-in user so that the source system’s own role-based access rules decide what data comes back.
Before giving AI access to a knowledge base, confirm SSO identity, admin-controlled connector permissions in your AI tool, and source-level role-based access, because MCP authorization is optional unless you enforce it.
How to give AI access to your knowledge base in 5 steps
Step 1: Map your knowledge sources and AI tools
Build a simple inventory with one row per source. Each row becomes a connection in the next step. For worked examples, see how teams connect SQL Server and ServiceNow.
Source | Data type (documents or records) | Owner | AI tools that need it |
Salesforce | Live customer relationship management (CRM) records | RevOps | ChatGPT Enterprise, Copilot Studio |
Step 2: Connect your sources to one governed MCP endpoint
Separate connections multiply reviews and upkeep. With Connect AI Gateway, you add each source once and give every AI tool the same governed MCP endpoint, https://mcp.cloud.cdata.com/mcp.
Add sources: Open Sources, click + Add Connection, and use derived views to curate what sources expose
Check the transport: The endpoint must use Streamable HTTP. Copilot Studio supports only that transport and no longer supports Server-Sent Events (SSE) for MCP after August 2025
Step 3: Set up OAuth/SAML passthrough and role-based permissions
A connected AI tool can reach whatever its credential can reach.
Identity: Connect AI Gateway passes the user’s identity through OAuth or SAML where the source supports it, so each request runs as that user and the source’s RBAC rules apply
Actions: In ChatGPT Enterprise, admins can turn custom-connector actions on or off, for example allowing read but not write, per OpenAI’s connector updates. Start every rollout read-only
Step 4: Add the endpoint to ChatGPT Enterprise, Copilot Studio, Claude, or n8n
Each client uses the endpoint from Step 2
Copilot Studio: Open your agent, go to Tools > Add a tool > New tool > Model Context Protocol, enter the endpoint as the server URL, then pick None, API key, or OAuth 2.0
ChatGPT: In workspace settings, go to Apps > Create and enter the endpoint, then an admin publishes it
Claude: Add it as a custom connector. Claude connects from Anthropic’s cloud, so the endpoint must be reachable from the public internet
n8n: The MCP Client Tool node supports Bearer, header, and OAuth2 authentication.
Step 5: Test requests, review audit logs, and scale to more teams
Test the same question as two users with different roles. Confirm that each gets a different, correctly permissioned answer, and check that both requests appear in the Connect AI Gateway audit log with each user’s identity. Only then expand to the next team. This test shows whether retrieval is running through an overprivileged identity.
Governance checklist and troubleshooting tips for getting IT approval
Cite the National Institute of Standards and Technology (NIST) AI Risk Management Framework and its Generative AI Profile (NIST-AI-600-1) in your approval request, so reviewers can tie the controls to a recognized framework.
Control | What IT verifies | Evidence for audit |
Token audience validation | MCP endpoint rejects tokens not issued for it | Endpoint configuration and security documentation |
User identity | Each request runs as the user, not a shared account | Identity provider settings and a request log |
Audit logging | Each request records the user and systems reached | Exported log sample |
Approval tip
RevOps and finance teams can ask IT to approve one governed endpoint instead of many per-app connectors. That shortens review and keeps the setup out of shadow IT.
Troubleshooting
Claude can’t connect: A common cause is a server behind a virtual private network (VPN) or firewall. Allow inbound traffic from Anthropic’s cloud.
Copilot Studio agent ignores the tool: Check that the transport is Streamable HTTP and that the server description clearly says when the agent should call it
Give every AI tool governed access to live data with CData Connect AI
Bring live enterprise data into the AI tools your team already uses with CData Connect AI Gateway. As a governed connectivity layer, it provides a single managed MCP endpoint to hundreds of sources, including Salesforce, SAP, NetSuite, SharePoint, and Confluence.
The same endpoint works with ChatGPT, Copilot Studio, Claude, and n8n. IT gets one auditable point of control with OAuth and SAML passthrough, and platform engineers can build a working proof of concept without maintaining custom connectors.
Start your free trial today!
Frequently asked questions
What do I need to set up before giving an AI tool access to my knowledge base?
Set up a single sign-on (SSO) identity provider, an admin-enabled AI workspace, and clean, current data. Confirm source-level role-based access before connecting anything.
Can I connect AI to live business systems like Salesforce, SharePoint, or NetSuite, not just documents?
Yes. An MCP endpoint fetches records at the moment a user asks, so the AI works with current data from live systems.
How do I make sure the AI only sees data each user is already allowed to access?
Use permission passthrough, where the source supports it, so each request runs as the signed-in user and the source’s own role-based access rules decide what comes back. Then test the same question as two users with different roles.
What is the fastest way to connect an AI agent to multiple enterprise data sources without building custom connectors?
Connect every source to one governed MCP endpoint, such as Connect AI Gateway, and add that endpoint to each AI tool instead of building a connector for each one.
How do I get IT or security approval to connect AI to our internal data?
Ask IT to approve one governed endpoint and bring the governance checklist.