Infor and Anthropic Move Claude Into Governed ERP Transactions

What changed

On November 12, 2026, Infor and Anthropic announced a manufacturing-focused partnership and initial availability of Infor Industry AI for Claude to select Infor Velocity Suite customers.¹ The original event date comes from Infor’s announcement, not from subsequent coverage. Anthropic announced Claude for Manufacturing as part of the same launch cycle, describing a package of manufacturing capabilities and connectors intended to support work across operational and business applications.²

The Infor integration is designed to let production planners, buyers and customer-service personnel ask questions about information held in Infor and then initiate approved actions from Claude. Examples disclosed by Infor include creating purchase orders, reallocating inventory and responding to production orders identified as at risk.³ Independent coverage corroborated the move from information access toward user-approved operational actions, while also reporting that the connector initially targets a limited customer population rather than broad general availability.⁴

Infor says the connector uses Model Context Protocol and applies existing Infor identity, role-based permissions, authorization controls and transaction logging.¹ Those are vendor-described controls, not independently validated findings. They matter because the proposed architecture does not give Claude an unrestricted mandate to alter the ERP system. Instead, the assistant is positioned as an interface through which an authenticated user can analyze context and request an action governed by the application’s existing controls.

Human authorization remains central to the announced design. Anthropic describes Claude for Manufacturing as supporting the drafting and analysis of manufacturing work, rather than directly operating production equipment.⁵ Infor likewise says relevant changes remain subject to review or approval.¹ The reported event is therefore not the deployment of autonomous control over factory machinery. It is the introduction of a natural-language interaction layer capable of reaching the threshold of an ERP transaction while retaining a human decision point.

Anthropic’s broader manufacturing offering includes 17 manufacturing skills and connectors for Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite.² Independent reporting also documented the 17 skills and four named connectors.⁶ Infor separately designated Claude as its preferred AI assistant for interaction with Infor products and Claude models as the preferred models for Infor Industry AI solutions.¹

The initial commercial and operational scope is constrained. Infor Industry AI for Claude was announced for select Velocity Suite customers, requires a separate Claude license and must be activated by an administrator.⁴ The available sources do not provide independently verified adoption figures, production-scale performance results, customer return on investment or a general-availability timetable. Those omissions do not negate the architectural significance of the announcement, but they prevent treating the vendors’ product direction as proof of realized business value.

Independent reporting helps bound the announcement. Business Insider interviewed Anthropic enterprise go-to-market leader Lauren Collett, who described manufacturing skills that assist quality engineers and technicians with documentation, diagnosis and other work without directly operating production equipment.[^business-insider] This is firsthand journalistic corroboration of the product positioning, not an independent benchmark of adoption, accuracy or economic results.

Why it matters

The CIO decision model changes when an assistant can cross the boundary between describing a business condition and initiating a change in the system of record. A question-answering tool can be assessed primarily through response quality, data access, privacy and user experience. A transactional interface must also be assessed as a privileged route into operational workflows. The relevant unit of evaluation is no longer just the model or chat experience; it is the complete path from user identity and retrieved context to proposed action, approval, execution and audit record.

Infor’s disclosed examples make that distinction concrete. Creating a purchase order, reallocating inventory or responding to an at-risk production order changes business state rather than merely summarizing it.³ Even when a human must approve the change, the assistant can influence what action is proposed, which records are selected and how the request is framed. The analytical implication is that access governance must cover both what an assistant may see and what it may prepare or initiate on a user’s behalf.

This shifts buying criteria away from generalized measures of model capability. Model quality still matters, but it cannot compensate for an overly broad permission path, an ambiguous approval boundary or an incomplete transaction log. CIOs should ask whether the connector inherits the application’s effective permissions, whether proposed actions are distinguishable from completed actions, whether approvers can inspect the underlying evidence and whether each step can be reconstructed after the fact. Infor says its connector applies existing identity, role-based controls, authorization and logging, but those assertions should be verified in each customer’s configuration rather than accepted as a universal property of the product.¹

Licensing also becomes an architectural concern. The announced implementation requires both qualifying Infor access and a separate Claude license.⁴ That means the cost and control boundary spans more than one vendor relationship. CIOs need to establish which provider is responsible for connector behavior, model access, identity propagation, transaction records and support when a proposed action is incorrect or cannot be completed. This is not evidence that the arrangement is unworkable; it is an inference about the diligence required when one vendor supplies the system of record and another supplies the interaction layer.

The broader connector set reinforces the point. Anthropic identified connections spanning Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite.² If assistants increasingly sit above multiple applications, employees may experience one conversational interface while the enterprise retains several underlying authorization and logging systems. The CIO’s governance challenge therefore moves upward: policies must remain consistent at the assistant layer without obscuring the distinct controls and approval rules of each connected application.

Human review should not be treated as a complete risk control by itself. The announced products retain review or approval for relevant outputs and changes, and Claude is not presented as directly controlling production equipment.⁵ But the effectiveness of review depends on what the approver can see and whether the transaction is reversible. A nominal approval click provides limited protection if the proposed action lacks its source context, affected records or expected consequence. This is an analytical inference from the workflow design, not a reported product deficiency.

Frontier take

The decisive enterprise AI control point is becoming the connector, not the chatbot. Once an assistant can initiate a transaction, its value and risk are determined by the integrity of the route between natural-language intent and the system of record. Infor and Anthropic’s announcement is strategically important because it places that route at the center of the product proposition: authorized users can move from operational inquiry to an approved ERP action without leaving Claude.¹

The governing principle should be delegated execution, not generalized agent access. An assistant should receive only the minimum ability required to prepare or initiate a defined action under the authenticated user’s authority. Approval should occur at the point where business state would change, with the proposed transaction, underlying context and responsible identity preserved in an auditable record. Infor says its implementation uses existing permissions and transaction logging, providing a testable design claim for customers rather than a substitute for customer testing.³

Model Context Protocol is relevant because Infor identifies it as part of the connector architecture.¹ However, use of a protocol does not by itself prove that permissions are narrow, approvals are meaningful or records are complete. CIOs should avoid converting a connector specification into an assumption of governance. The implementation must demonstrate how identity, authorization and transaction state behave in the customer’s own environment.

The strategic assertion is therefore clear: CIOs should classify AI assistants that can initiate business actions as privileged transactional interfaces. Procurement, security review and operating ownership should follow that classification. The winning deployment will not necessarily be the assistant that produces the most fluent answer. It will be the one whose authority can be constrained, whose proposed actions can be inspected and whose completed transactions can be reconstructed without ambiguity.

Three moves for CIOs

— Gate deployment on a permission-delta test Build a test matrix comparing what each pilot user can view, propose, approve and execute directly in Infor with what the same identity can do through Claude. Include the disclosed transaction types—purchase-order creation, inventory reallocation and responses to at-risk production orders—and record every case where the assistant pathway expands, obscures or inconsistently applies the user’s effective authority. Permit transactional use only when the connector introduces no unexplained permission delta.
- Decision trigger: Run the test before enabling any Claude workflow that can create or modify an Infor record, and repeat it whenever a role definition, connector configuration, model access policy or approval workflow changes.
- Why now: Infor says the connector applies existing identity, role-based permissions and authorization controls, but that remains a vendor claim until validated in the customer’s configuration.¹ Initial availability to select customers creates an opportunity to make permission equivalence a launch condition rather than remediate excess authority after wider adoption.
— Define a reversible transaction envelope For each permitted action, specify the records the assistant may touch, the evidence an approver must see, the person authorized to approve, the point at which the system of record changes and the method for reversal or correction. Require logs to connect the user request, assistant-generated proposal, approval decision and resulting Infor transaction. Keep workflows read-only when that chain cannot be reconstructed.
- Decision trigger: Move a use case from analysis to transaction initiation only after a tabletop review can reproduce one approved action and one rejected action from the available identity, approval and transaction records.
- Why now: The announced product crosses from querying operational information into initiating approved ERP actions.³ Human review is part of the disclosed design, but review becomes a meaningful control only when approvers receive sufficient transaction context and the enterprise can distinguish proposals from committed changes.
— Separate assistant adoption from transaction authority Negotiate and architect the deployment so analytical access, transaction initiation and application connectivity can be approved as separate stages. Document ownership for the model license, connector activation, application authorization, transaction logging and support escalation. Preserve the ability to suspend transaction initiation without disabling all analytical use.
- Decision trigger: Apply this staged authority model whenever a deployment requires separate assistant and enterprise-application licenses, or whenever one assistant is connected to more than one system containing operational records.
- Why now: Infor’s initial offering requires a separate Claude license and administrator activation, while Anthropic’s broader manufacturing launch names connectors spanning several enterprise applications.⁴² Separating usefulness from execution authority gives CIOs a practical way to capture analytical value without treating every connected system as ready for transactional AI.

Infor and Anthropic Move Claude Into ERP Transactions, Making Governance the Decisive AI Criterion

Take the argument with you.

What changed

Why it matters

Frontier take

Three moves for CIOs

Sources