VERSICH

ERP MCP Servers: Compare Read and Write Capabilities Safely

erp mcp servers: compare read and write capabilities safely

A common assumption is that an ERP MCP server is simply an API wrapper with a conversational interface. That is incomplete. ERP MCP servers expose structured tools and resources to AI applications, but read and write capabilities require different permission models, validation rules, approval paths, and audit controls. Read tools retrieve approved ERP data. Write tools create or change records, so they need stronger safeguards before an action reaches a system of record.

Model Context Protocol, or MCP, defines a way for an AI application to interact with external tools and resources. It does not replace ERP authorization, business rules, transaction controls, or human approval. A safe comparison therefore focuses on more than whether a server supports “read” or “write.” It asks what data the tool can access, which records it can change, how arguments are validated, whether the action is reversible, and how the organization proves what happened.

What Are ERP MCP Servers?

ERP MCP servers are integration components that expose approved ERP data and operations through MCP-compatible tools or resources. An AI assistant, agent, or development environment can use those interfaces to retrieve information, analyze records, prepare a draft, or request an operational action.

The server sits between the AI client and the ERP system. That position matters. It gives the server an opportunity to enforce identity, scope, field filtering, validation, rate limits, logging, and approval requirements before a request reaches the ERP.

A well-designed ERP MCP server does not expose the entire ERP API as one unrestricted function. It provides narrow, understandable operations such as:

  • get_invoice_status

  • search_open_orders

  • draft_purchase_order

  • update_customer_address

  • submit_journal_for_approval

The exact tool names are an implementation choice, but the principle is important. A narrow tool contract makes the permitted action easier to review than a generic operation such as run_erp_query or execute_erp_command.

MCP itself is not an authorization boundary. The underlying connector still needs to evaluate the authenticated identity, tenant or subsidiary scope, record permissions, field access, and action-specific controls. Our guide to building a secure AI agent layer with MCP explains why the agent-facing interface and the ERP safeguards should remain distinct.

ERP MCP Read vs. Write Capabilities

The central difference is simple: read capabilities retrieve information, while write capabilities create, update, submit, approve, or delete ERP data. The risk difference is not simple. A read request can still expose payroll, payment, customer, or personally identifiable information, while a write request can alter financial records, inventory, fulfillment, or approvals.

Read access should therefore be narrow rather than automatically trusted. Write access should be treated as a controlled transaction, not as a more advanced version of search.

CapabilityTypical ERP actionPrimary riskMinimum control approachApproval expectation
ReadRetrieve invoice status or order detailsUnauthorized data exposureRole checks, record scope, field filtering, audit logsNot always required, but access must be authorized
AnalyzeCompare records against a policy or ruleIncorrect interpretation or incomplete contextSource references, data freshness checks, confidence or exception statusHuman review for consequential recommendations
DraftPrepare a purchase order, credit memo, or journal entryIncorrect data prepared for later submissionRequired-field validation, duplicate checks, draft-only permissionReview before submission
WriteCreate or update an ERP recordIncorrect or unauthorized business changeServer-side validation, idempotency, transaction loggingRequired for material or sensitive changes
ApproveApprove a payment, order, or journalBypass of segregation of dutiesSeparate approval identity, workflow enforcement, step-up controlsAuthorized human approval
DeleteRemove a record or related dataData loss and audit disruptionDeny by default, retention rules, explicit confirmationStrong approval, with deletion rarely exposed

The “analyze” and “draft” stages deserve attention because they prevent a false choice between read-only automation and unrestricted writes. An agent can retrieve an invoice, compare it with approved rules, and prepare a proposed action without receiving permission to commit the change.

For organizations evaluating ERP AI agents, this staged model is more useful than a binary read/write label. Our article on giving ERP AI agents identity and boundaries before they act describes the progression from read to analyze, recommend, draft, and commit.

Why Read-Only ERP MCP Tools Are Not Automatically Safe

Read-only ERP MCP tools reduce the risk of unauthorized changes, but they still create data-access risk. A tool that can search every customer, employee, vendor, payment, and transaction record is not safe simply because it cannot edit them.

The first control is scope. The server should evaluate which legal entity, subsidiary, business unit, department, location, or account the requesting identity is allowed to view. Scope should be enforced on the server side, not left to the model to remember in its prompt.

The second control is field minimization. A user may need invoice status and due date without needing bank instructions, tax identifiers, employee compensation, or a complete customer profile. Returning fewer fields reduces the impact of a compromised session, a misleading prompt, or an overly broad question.

The third control is query design. A dedicated get_invoice_status tool has a clearer security contract than a general-purpose SQL-like query tool. Search tools should constrain filters, result size, sorting, and pagination. They should also prevent the model from using a broad search to reconstruct sensitive datasets through repeated calls.

A practical read policy evaluates:

  1. Identity: Which user, service account, or agent is making the request?

  2. Record scope: Which subsidiaries, departments, customers, or transactions are in scope?

  3. Field scope: Which attributes are necessary for the task?

  4. Time scope: How long should the credential or session remain valid?

  5. Purpose: Does the requested data support the stated workflow?

Data freshness also matters. An AI response that cites an order status retrieved several minutes ago should identify the retrieval time when a user could make a consequential decision from it. For high-value workflows, the server should return source identifiers and relevant record timestamps alongside the business data.

Versich approaches AI planning with governance and risk controls before implementation. Our AI consulting service is a relevant starting point for teams deciding which ERP data and actions belong in an agent workflow.

What Makes ERP MCP Write Tools Different?

Write tools change the state of the ERP, so they need controls that are enforced before execution. A prompt instruction such as “only update approved orders” is not sufficient. The connector must validate the order status, user permission, field values, and business conditions on every request.

Write operations should be divided by impact. Adding an internal note is not equivalent to changing a sales order total. Drafting a purchase order is not equivalent to submitting it. Approving a vendor payment deserves more protection than updating a nonfinancial description.

The most important design details include:

Explicit input schemas. Every argument should have a defined type, format, length, and allowed value set. A currency amount should not arrive as unconstrained text. Dates, identifiers, quantities, and status fields need strict parsing.

Server-side business validation. The connector should verify required fields, allowed status transitions, duplicate risk, subsidiary alignment, currency rules, tax treatment, and other ERP conditions. Validation must occur after the request reaches the trusted integration layer.

Idempotency. A retry should not create two purchase orders, duplicate a journal, or submit the same request twice. Write tools should accept an idempotency key or equivalent request identifier where the ERP workflow supports it.

Transaction boundaries. If an operation updates multiple records, define what happens when one update succeeds and another fails. A partial update can create reconciliation work or leave an approval process in an inconsistent state.

Confirmation and approval. The user should see what will change before a material action commits. For higher-risk actions, the agent should create a draft or approval request rather than execute the transaction directly.

Audit evidence. Logs should record the requesting identity, tool name, arguments after sensitive values are protected, validation results, approval identity, ERP response, timestamps, and correlation ID.

Write access also needs a clear response contract. “Done” is not enough. A useful response identifies the record ID, resulting status, timestamp, and any downstream workflow initiated by the action.

How Should ERP MCP Servers Handle Approval?

Approval should match the business impact of the action. A low-risk internal note may need authorization but not a separate approval step. A payment release, journal posting, vendor bank change, or order cancellation should use a stronger path.

A safe approval model separates recommendation, drafting, submission, and approval. An AI agent can recommend an exception without having write permission. It can prepare a draft without being allowed to submit it. It can submit a request without being the identity that approves it.

This separation supports segregation of duties. The identity that creates a transaction should not automatically approve the same transaction. The MCP server should not weaken ERP workflow rules merely because the request originated in a conversational interface.

Approval controls should answer specific questions:

  • What exact fields and values are being changed?

  • Which business rule triggered the approval?

  • Who is authorized to approve this action?

  • Does the approver have a conflict of interest?

  • Does the request expire after a defined period?

  • What happens if the underlying record changes before approval?

  • Can the approval be revoked or rejected with a reason?

A strong pattern is to bind approval to the exact proposed payload. If a user approves a purchase order for one amount and the agent later changes the amount, the original approval should no longer apply.

For US organizations, this approach also supports audit and internal-control objectives without treating MCP as a substitute for established ERP controls. Organizations operating across other markets should map the same design to their applicable privacy, financial, and records requirements rather than assume one universal compliance configuration.

ERP MCP Server Security Controls to Compare

A useful evaluation compares the controls behind each capability, not just the number of tools exposed. Two servers might both offer invoice updates, but one could enforce record-level access and approval while the other forwards broad API credentials.

Control areaRead capability questionWrite capability question
AuthenticationHow is the requesting user or service identified?Is the writer tied to an approved identity rather than a shared credential?
AuthorizationAre records and fields filtered by role and scope?Are create, update, submit, approve, and delete permissions separated?
ValidationAre filters and query parameters constrained?Are values, status transitions, duplicates, and dependencies checked server-side?
Data exposureAre sensitive fields excluded by default?Are secrets and unnecessary response fields omitted from logs and replies?
ApprovalIs access itself authorized?Is human approval required for high-impact changes?
AuditabilityAre retrievals logged with identity and purpose?Are proposed values, approvals, results, and correlation IDs retained?
ReliabilityAre pagination, timeouts, and stale data handled?Are retries idempotent and partial failures recoverable?
MonitoringAre unusual search patterns detected?Are repeated failures, abnormal volumes, and privilege changes investigated?
RecoveryCan exposed data access be revoked quickly?Can a bad change be reversed or quarantined safely?

The information-gain detail here is correlation across the request lifecycle. A tool call log, approval record, and ERP transaction log should be connectable through a shared request or correlation ID. Without that connection, a team may know that a record changed but struggle to determine which agent request, user identity, validation result, and approval caused the change.

How to Test Read and Write Capabilities Safely

Testing should begin in a sandbox or other nonproduction environment with synthetic, masked, or minimum-necessary data. Production access should not be the first test of a new MCP server.

Start by defining representative read and write cases. A read case might retrieve the status of a permitted invoice and confirm that restricted fields remain absent. A negative read case should attempt to access a record outside the user’s subsidiary or role scope. A write case might draft an update with valid values, while a negative write case submits an invalid status transition or changes a protected field.

A practical test sequence is:

  1. Verify authentication and identity propagation.

  2. Confirm allowed records, fields, and subsidiaries are visible.

  3. Confirm unauthorized records and fields are denied or filtered.

  4. Test malformed arguments, missing required fields, and unexpected data types.

  5. Retry a write request and verify that it does not duplicate the transaction.

  6. Change the underlying record before approval and confirm stale approvals fail.

  7. Inspect logs for secrets, excessive personal data, and complete correlation details.

  8. Test timeouts, API errors, partial failures, and rollback or recovery behavior.

Threat testing should include prompt injection inside ERP text fields. A customer note, invoice memo, or order description might contain instructions intended to manipulate an AI agent. The server must treat ERP content as data, not as trusted instructions. Tool descriptions and system policies should also be protected from unauthorized modification.

Rate limits and result limits are practical controls that are easy to overlook. A read agent that can issue unlimited broad searches creates an extraction path even when each individual response appears authorized. Write tools need volume limits too, especially for bulk updates or repeated retries.

Common ERP MCP Design Mistakes

The most serious mistakes come from treating MCP as the control layer rather than as an interface layer.

One mistake is exposing a generic “execute action” tool with a large argument set. This makes it difficult to review permissions, predict behavior, and distinguish a harmless update from a financial commitment. Narrow tools are easier to test and monitor.

Another mistake is using the same credential for reads, drafts, writes, and approvals. Separate permissions make it possible to limit blast radius and enforce separation of duties.

A third mistake is trusting the AI application to apply business rules. The model may produce a plausible request that violates a required field, accounting period, credit limit, or approval path. The connector and ERP must enforce those rules.

Teams also overlook response data. A write tool can safely update the intended record but still return confidential fields that the user did not need. Response minimization belongs in the design, not only in the prompt.

Finally, organizations sometimes begin with complex autonomous actions before proving a focused read-only workflow. A safer rollout starts with data quality, permission behavior, observability, and user feedback, then expands to recommendations, drafts, and approved commits.

How to Choose Between Read-Only, Draft, and Write Access

Choose the lowest capability that completes the business task. If a user needs to answer “Which invoices are overdue?”, read access is sufficient. If the user needs a proposed collection action, recommendation or draft access is more appropriate. If the workflow requires an approved status change, write access should be added only after the earlier stages are controlled.

The right choice depends on four dimensions:

  • Business impact: What happens if the action is wrong?

  • Reversibility: Can the change be undone cleanly?

  • Data sensitivity: Does the workflow involve payment, payroll, personal, or regulated information?

  • Approval structure: Does the business already have a defined reviewer and segregation-of-duties path?

A useful implementation plan defines each tool with an owner, purpose, permitted identity, record scope, field scope, validation rules, approval level, audit requirements, and retirement process. That documentation gives security, finance, operations, and engineering a shared basis for review.

Teams that need help evaluating an existing ERP integration can contact Versich for an initial discussion. The useful outcome is not simply deciding whether to enable MCP. It is identifying which actions deserve read, analyze, draft, submit, or approval capabilities.

The Next Step for Comparing ERP MCP Capabilities

The safest path is to compare capabilities by consequence, not by feature count. Start with the smallest useful read workflow, verify identity and data boundaries, then introduce analysis and drafts before enabling approved writes. That sequence reveals permission gaps and data-quality issues while the blast radius remains limited.

A well-designed ERP MCP server makes every action understandable: who requested it, what records were in scope, which fields were returned or changed, what validations ran, whether approval was required, and what the ERP ultimately recorded. Use that standard to evaluate your architecture, document the capability boundary, and decide which next step belongs in a controlled pilot.

Frequently Asked Questions

What is an ERP MCP server?

An ERP MCP server exposes approved ERP data and operations to an AI application through Model Context Protocol. It acts as an integration layer that can enforce identity, record scope, field filtering, validation, approvals, and audit logging before requests reach the ERP.

Are read-only ERP MCP servers safe?

Read-only ERP MCP servers are safer than write-enabled servers for preventing unauthorized changes, but they are not automatically safe. They still need authentication, record and field-level authorization, sensitive-data filtering, query limits, monitoring, and access logs.

Is MCP required to connect an AI agent to an ERP?

No. MCP is one interface option for connecting AI applications to structured tools and resources. An organization can use direct APIs, workflow platforms, custom services, or other integration patterns, but every approach still needs ERP authorization, validation, monitoring, and governance.

Should ERP MCP write tools require human approval?

High-impact ERP write tools should require human approval or an equivalent controlled workflow. Payment approvals, journal posting, vendor bank changes, order cancellations, and material financial updates deserve stronger review than low-risk internal notes or drafts.

What is the difference between ERP MCP read and write access?

Read access retrieves ERP information without changing the system of record, while write access creates or modifies records and may trigger downstream workflows. Write access therefore needs stricter validation, idempotency, approval, transaction handling, and audit controls.

How much does it cost to build an ERP MCP server?

The cost depends on ERP APIs, number of tools, data sensitivity, authentication design, approval workflows, testing, monitoring, and support requirements. A narrowly scoped read-only connector requires less work than a production system that supports financial writes and segregation of duties.

Can an ERP MCP server replace ERP permissions?

No. An ERP MCP server should reinforce and narrow ERP permissions, not replace them. The ERP remains the system of record, while the MCP layer adds a controlled interface with tool-specific validation, data minimization, and agent-aware monitoring.

Tags: