Short answer
MCP servers give AI agents a standard way to read from and act on enterprise systems such as SAP, SharePoint and ticketing tools. In an enterprise they should be treated like any other integration layer: scoped permissions per user, a gateway with authentication and audit logging, curated tools instead of raw API access, and a review process before a server goes live.
On this page
What MCP changes for enterprise integration
The Model Context Protocol (MCP) is an open standard that defines how AI applications discover and call external tools and data sources. An MCP server exposes a set of tools (actions such as "create ticket"), resources (data such as a document) and prompts. Any MCP-capable client, from coding agents like Claude Code to desktop assistants and agent platforms, can then use them.
Before MCP, every AI tool needed its own connector for every system. With MCP, you build or adopt one server per system and reuse it across clients. That is a real simplification. It also means that a single, poorly designed server can expose a core system to every agent in the company. Governance has to be designed in, not bolted on.
Typical enterprise systems and what to expose
| System | Useful read tools | Useful write tools | Main risk |
|---|---|---|---|
| SAP (ERP) | Order status, material master lookups, open items | Draft purchase requisitions, create notes | Financial impact, authorisation complexity |
| SharePoint and document stores | Search, fetch document, list recent files | Upload draft, add comment | Oversharing through broad permissions |
| Ticketing (Jira, ServiceNow and similar) | Search issues, read history | Create or comment on tickets, change status | Spam, unintended workflow transitions |
| CRM | Account and contact lookup, pipeline overview | Log activity, draft opportunity updates | Personal data under GDPR |
| Data warehouse | Run approved queries, describe tables | Usually none | Data exfiltration, expensive queries |
| Internal wikis and code hosting | Search, read pages, read repositories | Draft pages, open pull requests | Leakage of confidential content |
A principle that holds across all rows: expose business-level tools, not raw APIs. "Get open invoices for customer X" is safer and more useful than a generic "call any OData endpoint" tool. Curated tools constrain what an agent can do, make audit logs readable and reduce the chance that a prompt injection turns into a damaging action.
Reference architecture
A pattern that works well in enterprises has four layers:
- Clients: coding agents, chat assistants and agent workflows that speak MCP.
- MCP gateway: a central entry point that authenticates the user, applies policies (which user or agent may use which server and tool), rate-limits and logs every call.
- MCP servers: one per system or domain, owned by the team that owns the system, deployed like any internal service.
- Systems of record: SAP, SharePoint, ticketing and databases, accessed with the user's delegated permissions wherever possible.
Running local MCP servers on developer laptops is fine for experiments. For production use with business data, remote servers behind a gateway give you control over versions, credentials and logs.
Permission design: the part that decides safety
- Act on behalf of the user. Prefer delegated authentication (for example OAuth flows against your identity provider) so the agent can only see and do what the human could. Avoid shared service accounts with broad rights.
- Separate read and write servers or tools. Many use cases need only read access. Roll out read-only first, add write tools individually.
- Require confirmation for consequential actions. Creating a purchase requisition or changing ticket status should require explicit human approval in the client.
- Limit result sizes. Tools that return whole document libraries or table dumps invite both cost problems and data leakage.
- Treat tool output as untrusted. Content retrieved from documents or tickets can contain instructions aimed at the model. Agents must not follow them blindly, and sensitive actions should never depend solely on retrieved text.
For SAP specifically, our article on connecting AI agents to SAP safely goes deeper into authorisation objects and audit requirements.
Governance process for new servers
Treat each MCP server like an integration going into production:
- Intake: owner, purpose, systems touched, data classes, users.
- Design review: tool list with descriptions, read/write classification, authentication method.
- Security review: dependency check, secrets handling, input validation, logging, prompt injection exposure.
- Registration: add to an internal catalogue so clients only connect to approved servers.
- Monitoring: usage, errors, unusual patterns, and periodic re-certification of access.
Third-party and community MCP servers deserve particular care. Review the code or vendor terms, pin versions and run them in your own environment rather than connecting to unknown remote endpoints.
Where kalAIDoCreate fits
Building a gateway, catalogue, audit trail and permission model yourself is possible, but it is a platform project. kalAIDoCreate provides this layer as an enterprise AI orchestration platform: governed AI agents, audit trails, SSO, integrations with SAP, CRM, SharePoint and Slack, and EU data residency, together with an expert team that helps design the use cases. Instead of every department wiring its own agents to core systems, kalAIDoCreate centralises connections and policies so that security and compliance teams see one controlled surface. For a structured view of that decision, read build vs buy: enterprise AI orchestration platform.
Getting teams ready
MCP is easy to start with and easy to get wrong. Engineers building servers need to understand tool design, authentication and injection risks; architects need to understand where MCP fits next to existing integration platforms. neooutput runs hands-on workshops on building and governing MCP servers, described in our MCP workshop article. A practical first project is a read-only server for one internal knowledge source, rolled out to a pilot group with full logging.
FAQ
Is MCP secure enough for enterprise systems?
MCP itself is a protocol; security depends on how servers and clients are implemented and governed. With delegated authentication, curated tools, a gateway, audit logging and human confirmation for write actions, it can meet enterprise requirements.
Should we use community MCP servers for SAP or SharePoint?
They are useful for learning and prototypes. For production, review the code thoroughly, run it in your own infrastructure and consider building curated servers with business-level tools that match your authorisation model.
Does MCP replace our integration platform or iPaaS?
No. MCP is an interface for AI agents. Existing integration platforms and APIs often sit behind MCP servers, which then expose a curated subset of their capabilities to agents.
Who should own an MCP server?
The team that owns the underlying system, supported by a central platform team that runs the gateway, catalogue and security standards. This keeps domain knowledge and accountability together.