Publicité
ERP IMPLEMENTATION
🇫🇷 Lire en français →

MCP and ERP: Connecting AI Agents to SAP, Dynamics, Odoo or Sage Without Opening Every Door

Model Context Protocol and ERP: vendor support at SAP, Microsoft, Odoo and Oracle, target architecture, 7 risks, guardrails and a 3-step pilot plan.

MCP and ERP: Connecting AI Agents to SAP, Dynamics, Odoo or Sage Without Opening Every Door

Connecting an AI agent to your ERP now takes a few hours. Deciding what it is allowed to do there takes longer, and that is where projects succeed or fail. The Model Context Protocol (MCP) is becoming the standard plug between agents and enterprise systems. Several ERP vendors already ship an implementation, at very different levels of maturity. For a CIO or CISO, the question is no longer whether teams will test MCP on the ERP, but under what conditions.

This article does not compare vendor AI assistants: that is covered in our comparison of SAP Joule, Sage Copilot, Odoo AI and Microsoft Copilot. Nor does it detail Sage Intacct, which has its own analysis of the Sage Intacct MCP AI Gateway. It sets out the cross-vendor frame: what MCP is, where it exists, how to architect it, what risks it creates and which guardrails to put in place before letting an agent touch your business data.

MCP in 5 Minutes: What It Is, and What It Is Not

The Model Context Protocol is an open standard describing how an AI application (the MCP client, for example an assistant or agent) discovers and calls tools exposed by an external system (the MCP server). Microsoft’s Dynamics 365 documentation describes it as a standard that “standardizes how applications provide context to LLMs” (Microsoft Learn, Dynamics 365 ERP MCP).

In practice, an MCP server publishes a list of tools, each with a name, a description and parameters. The agent reads that list, picks a tool based on the user’s request and calls it. Three differences matter to an IT organization:

  • Versus a classic API: an API is designed for a developer who knows which endpoint to call. An MCP server is designed for a model that must decide at runtime which tool to use. The contract is no longer purely technical: it contains natural language (tool descriptions) that the model interprets.
  • Versus an iPaaS connector: an iPaaS flow is deterministic, designed in advance, tested and versioned. With an agent, the sequence of calls is not written beforehand. The same request can produce two different call sequences.
  • Versus RPA: MCP does not go through the graphical interface. Tools are called server-side, with an identity and a token, which makes logging possible, provided you configure it.

The key point for what follows: MCP standardizes transport and tool discovery. By itself it does not settle permissions, validation of actions or traceability. That depends on the server implementation and on the architecture you build around it.

Market Status: Native, Community, Home-Built Gateway

The landscape is uneven. The table below summarizes what we could confirm in vendor sources, and what we could not. It is dated November 20, 2026 and should be rechecked before any decision: these products change with every release.

VendorWhat is documentedStatus in our verification
Microsoft Dynamics 365 (Finance and Supply Chain)Dynamic “Dynamics 365 ERP MCP” server, with data, form and action toolsDocumented by the vendor (Microsoft Learn). The older “static” server was due to be retired on October 1, 2026
SAPMCP gateway in SAP Integration Suite to expose SAP and third-party APIs as MCP tools; Joule agents consume MCP toolsDocumented in SAP’s reference architecture, with the caveat that some components are “not yet generally available” (SAP Architecture Center)
Odoo (version 20)Built-in MCP server, OAuth recommended or API key, tools exposed as server actionsDocumented in the Odoo 20 documentation (Odoo documentation). Community modules and servers also exist for other versions
Oracle FusionSearch results point to MCP elements around AI Agent Studio and analyticsNot confirmed: we could not re-read the vendor pages during our verification. Do not infer native support in the ERP applications
Sage IntacctRead-only MCP AI GatewayCovered in the dedicated article

Three takeaways for decision-makers:

  1. “MCP support” covers very different realities. At SAP, the documentation mainly describes an exposure and governance layer (the gateway) in front of existing APIs. At Microsoft, the server gives access to data, forms and business logic. At Odoo, access goes through declared server actions. These are not the same attack surfaces.
  2. Client and server are two distinct roles. A vendor announcing that its assistant “supports MCP” often means it can call MCP servers, not that its ERP exposes an MCP server. Always ask which way the arrow points.
  3. Community servers exist, but the liability is yours. For most ERPs you can find open-source MCP servers published by third parties. They may be fine for a prototype on a disposable environment. They must not touch a production database without code review, version pinning and an analysis of the permissions they request.

Target Architecture

Read-Only or Write: Two Populations of Agents

The most rewarding architecture decision is also the simplest: separate the agent that reads from the agent that acts.

  • The reading agent answers questions (available stock, order status, account balance, supplier payment schedule). Its identity holds read rights only, on a restricted scope. Its worst case is a data leak, which is already serious, but it changes nothing.
  • The acting agent creates or modifies objects (draft journal entry, purchase order, third-party record). It falls under a different level of control: human validation, quotas, enhanced logging.

Do not mix the two in one agent and, above all, not in one identity. The Microsoft example is telling: its server exposes create, update and delete tools alongside the read tools, and it is the security role assigned to the agent that bounds what it can do (Microsoft Learn). A lookup agent should therefore get a role with no write privilege at all, not a broad role “for now”.

A Gateway or Proxy Between Agent and ERP

Avoid a direct connection between a consumer assistant and the production ERP. Insert an intermediate layer, whether provided by the vendor (the MCP gateway in SAP Integration Suite is one example) or built by your teams. This layer is used to:

  • authenticate and authorize each call with your identity provider;
  • expose only an allowlist of tools, with descriptions you control;
  • apply quotas and timeouts;
  • log every call consistently, whatever ERP sits behind it;
  • filter or mask certain fields before they reach the model.

The MCP specification itself warns against two common design mistakes on these proxies: the “confused deputy” (a proxy acting with its own rights on behalf of an unauthorized client) and “token passthrough” (a server accepting and forwarding tokens not issued for it), which the specification expressly forbids (MCP, Security Best Practices). The same page notes that passthrough blurs the audit trail: downstream system logs no longer show who actually made the call.

Agent Identity: Service Account or User Delegation

Two models coexist, and the choice has direct consequences for audit.

Dedicated service account for the agent. The agent acts under its own identity, with its own roles. Advantages: explicit scope, simple revocation, clear separation between agents. Drawback: you lose the link to the user who triggered the action unless you log it separately, and a service account tends to accumulate rights “to make it work”.

Delegation of the user’s identity. The agent acts with the rights of the person using it. Advantage: you reuse the existing permission model, and a user cannot obtain through the agent what they could not obtain alone. At Microsoft, the documentation states that the context returned to the agent depends on the authenticated user’s security role, and that the system rejects calls to objects that role cannot access. Drawback: the agent inherits all of the user’s rights, including those the task does not need.

Our recommendation: delegation for conversational lookup agents, a narrow service account for autonomous agents, and in both cases a role specific to the use case rather than an existing business role. In Odoo 20, the documentation provides for revoking an MCP client’s access by archiving the registered application: build that step into your incident procedures, whatever the ERP.

The 7 Risks to Address

  1. Over-permissioning. A role that was too broad “for testing” and stays in production. The agent does not need to be malicious: an ambiguous request is enough for it to use a right it should never have had.
  2. Prompt injection through ERP data. A comment field, an item description or a supplier attachment can contain text that, when read by the model, looks like an instruction. OWASP distinguishes direct and indirect injection, the latter coming from external sources the model processes (OWASP, LLM01 Prompt Injection). In an ERP, any data entered or imported by a third party is untrusted input.
  3. Exfiltration. An agent that reads sensitive data and also has a tool capable of sending outward (email, web request, another MCP server) can move that data out of the perimeter. The danger comes from the combination of tools, not from a single tool.
  4. Irreversible actions. A payment issued, an entry posted in a closed period, a third party deleted, an order sent to a supplier. Every write tool should be classified by reversibility.
  5. No traceability. If ERP logs show “service user” with no link to the original request, the auditor cannot reconstruct the chain. This is the effect of the passthrough described above, but also of logs that record neither the prompt nor the tool called.
  6. Segregation of duties (SoD) bypassed. An agent that can create a vendor, change its bank details and run a payment rebuilds exactly the combination your SoD matrix forbids to humans. Non-human identities must enter the matrix, just like named accounts. For the method, see our SoD controls to implement before the annual audit.
  7. Personal data. ERPs hold HR, payroll and customer data. Any data returned to a model is data passed to a processor or service whose hosting, retention and purposes you need to know.

Guardrails to Put in Place

Least privilege, per tool rather than per application. Start from zero tools and add them one at a time. The MCP specification recommends a progressive scope model, with a minimal read set at the start and targeted elevation when a privileged operation is requested; it discourages catch-all scopes (MCP, Security Best Practices).

Human validation on every write. OWASP explicitly cites human control for privileged operations. In practice: the agent produces a draft (entry, order, record), never a validated object. Validation goes through a human holding the corresponding right, inside the ERP, through the usual workflow.

Usable logging. A useful log contains, for each call: the agent’s identity, the user behind the request, the tool, the parameters, the result, the timestamp and a correlation ID tying the call to the conversation. The MCP specification recommends logging scope elevations with correlation IDs. Send these logs to your SIEM, not only to the agent vendor’s console.

Quotas and volume guardrails. Limit calls per minute, records returned and writes per session. An agent stuck in a loop must not be able to create five hundred drafts.

Test environment first. At Microsoft, for example, the server is not supported on cloud-hosted environments (CHE) and requires a Tier-2 environment at minimum or a unified developer environment, which forces you to start outside production anyway (Microsoft Learn). Test with anonymized data and deliberate injection attempts.

Client allowlist. Microsoft requires adding the agent platform to a list of allowed MCP clients; Odoo lets you approve specific clients and enable dynamic registration only as an option. Disable anything that lets any client register itself.

Exclude administration areas. Microsoft’s documentation states that certain administration forms, such as security configuration, role assignment or user management, are not accessible through the server. Check the equivalent at your vendor, and otherwise block it at the gateway.

This topic overlaps four frameworks we have already covered, so we do not repeat them here:

  • AI Act: depending on the use case (recruitment, scoring, decisions about people), an agent connected to the ERP may fall under the obligations for high-risk systems. See our analysis of the EU AI Act applied to ERP.
  • GDPR: legal basis, minimization, processors, transfers. See ERP and GDPR.
  • DORA: in the financial sector, an agent and its model provider become an ICT third-party risk. See DORA and ERP.
  • NIS2: access management, logging, incident management. See the NIS2 evidence checklist.

For the identity and privileged access management underpinning all of this, see our Zero Trust, IAM, PAM and MFA guide for ERP.

3-Step Pilot Plan

Step 1: read-only on a narrow scope (4 to 6 weeks). Choose a low-risk use case: stock queries, order status, payment schedule lookups. Test environment with anonymized data, a single agent, a read role limited to a few entities. Measure answer quality, verify that every call shows up in your logs, and run at least three injection scenarios through text fields.

Step 2: read-only in production, with identity delegation. Same use case, real pilot users, rights inherited from their role. Verify that a user cannot see through the agent what they cannot see in the ERP. Connect the logs to the SIEM and have internal audit review a sample.

Step 3: draft entries, with human validation. The agent proposes, a human validates. Start with a low-stakes object type (for example, draft accrual entries or purchase requisitions), with a strict quota and a dedicated service account, integrated into the SoD matrix. No payment, deletion or bank-detail modification tools at this stage. Moving to direct writes, if it ever happens, is a separate decision with a new committee.

Each step ends with a documented go / no-go decision: which incidents, which log anomalies, which requests for additional rights.

Pre-Production Checklist

  • Dedicated ERP role for the agent, with no right inherited “for convenience”, reviewed by the process owner
  • Reading agent and writing agent separated, with distinct identities
  • Non-human identities included in the SoD matrix and tested
  • Tool allowlist, controlled tool descriptions, no administration tool exposed
  • MCP client allowlist, dynamic registration disabled or justified
  • No unvalidated token forwarding (no passthrough) between agent, proxy and ERP
  • Writes only as drafts, human validation traced
  • Per-call log (agent, originating user, tool, parameters, result, correlation) sent to the SIEM
  • Call, volume and write quotas configured
  • Prompt injection tests through text fields and attachments carried out and documented
  • GDPR analysis done: data exposed to the model, hosting, retention, processors
  • AI Act qualification and, depending on sector, DORA / NIS2 recorded
  • Revocation procedure for an agent or MCP client tested, with a named owner
  • Vendor status reviewed (GA, preview, announced) and retirement date of older server versions

Going Further

If your thinking covers the whole of agent deployment (choice of use cases, governance, value measurement), read our agentic AI deployment guide for CIOs. To receive our next analyses on ERP AI and security, subscribe to the newsletter.