On November 2, Gartner released the "Magic Quadrant for Product-Centric Enterprise Cloud ERP, 2026" (analysts Tomas Kienast and four others); on November 6, Infor and Microsoft each officially announced their inclusion in the "Leaders" quadrant; and on November 4, Infor had already listed item by item the specific actions of that batch of agents in the October version.
Putting the three things together makes it clear: the ERP industry is rewriting "agentic AI" from a feature term into a market definition. The judgment quoted verbatim from Gartner in Infor's official release is that—agentic AI has become the dominant innovation, driving architectures toward composability and platformization, enabling enterprises to differentiate through "governed extension" while maintaining a "standardized operating core."
Gartner predicts that by 2030, more than half of basic ERP tasks will be performed autonomously by AI, reducing manual involvement in finance, supply chain, and HR processes (forecast value, quoted from Infor's official release dated 2026-11-06)
But what is truly worth a CIO's time is not that 2030 number, but the other half of the question: to hand off half of the tasks, what must be sorted out today first? The answer in Infor's checklist is surprisingly plain—all routine work that is clearly defined, verifiable, and traceable.
1. One Word Changed: From “System of Record” to “Enterprise Resource Execution”
In an announcement on November 6, Infor brought a new term to the fore: ERX (enterprise resource execution), and said that this is the term Gartner uses to describe "intelligent, autonomous ERP." To clarify the basis: this term currently appears in the vendor's paraphrase of the report; this article cites it according to Infor's wording and has not independently verified it.
What is even harder than the naming are two quantitative predictions and the same week's manufacturers' hidden cards:
| caliber | Content | Source and Date | Nature |
|---|---|---|---|
| Market Definition | Agentic AI becomes the dominant innovation; architecture shifts to composable and platform-based; "governed scaling" differentiates, while the "standardized operating core" remains stable | Gartner Magic Quadrant, 2026-11-02 (as cited in Infor's official release) | Institutional judgment |
| Prediction One | By 2030, >50% of basic ERP tasks will be autonomously executed by AI | Gartner, 2026-11-02 (cited in Infor, 11-06) | Predicted value |
| Prediction Two | In 2030, AI tools reduce ERP modernization costs by 40% | Gartner, as cited via Infor's official release dated 11-06 | Predicted value |
| Manufacturer actions | Infor was named to the Leaders Quadrant for the sixth consecutive year; Microsoft announced its inclusion on the same day and listed Dynamics 365 ERP MCP Server (GA in January 2026), Work IQ, Procurement Agent (production-ready preview), and Agent 365 | Infor and Microsoft official blogs, both dated 2026-11-06 | Manufacturer Announcement |
A reminder that must be written into the selection meeting minutes: Gartner's disclaimer explicitly states that it does not endorse any vendor; reports are opinions rather than statements of fact. Two companies announcing on the same day that they were "selected as Leaders" only indicates who is in the quadrant.
Translating "composable + governed extensions + standardized core" into implementation language is really three old sayings: don't mess with the core ledger, put differentiation in a versionable extension layer, and agents can only go through interfaces whose permissions have already been defined. The new term is ERX; the engineering requirements haven't changed.
II. Breaking down "autonomous execution": What exactly do the five types of agents take over?
Infor's November 4 announcement described the 2026.10 release launched in October, covering CloudSuite Distribution, Distribution Enterprise, and WMS. ERP News provided an item-by-item breakdown the same day. When the list is laid out, it becomes clear that so-called "autonomous execution" comes down to very specific actions:
| Agent / Capabilities | Specific actions | Boundary |
|---|---|---|
| Quoting Agent (Quotation) | Retrieve customer and product information, apply pricing and discounts, and convert quotes into orders | Pricing and discounts follow existing rules; the model does not negotiate on its own |
| Purchasing Agent (Purchasing) | Track purchase orders, identify delayed deliveries, and generate supplier follow-up letters | Action invokes underlying ERP services via GenAI Assistant |
| Product Agent (Fulfillment) | Aggregate warehouse availability, supply sources, order history, and costs, and recommend fulfillment paths | Output is a suggestion, not an automatic commitment to delivery date |
| Batch Wave Agent (Warehousing) | Combine orders with shared line items into picking waves to reduce picking travel | Triggered on demand (on request), not continuous autonomous operation |
| User & Task Management Agent | Use natural language to handle onboarding, permission configuration, and task reassignment | Permission changes are sensitive actions and must be logged for audit purposes. |
| Long-tail forecasting and financial reconciliation | Forecast demand for new/slow-moving SKUs by product attributes, and set safety stock based on forecast fluctuations; convert RFQ/PO/PDF into CPQ quotes or ERP orders; reconcile supplier statements and process scanned bank statements with AI | Unmatched items are explicitly left for manual processing. |
The automotive industry 2026.10 release also gives two figures: agents can connect to Infor LN's 6,000+ OData APIs via MCP, and Infor has consolidated more than 100 single-purpose agents into 12 domain agents. The material shortage check agent only provides "suggestions"; after the planner reviews them, it is asked to make changes and then rerun validation. The official wording is "Planners stay in control" (Infor release notes).
6,000+
Number of OData APIs exposed by Infor LN via MCP (per vendor technical documentation, 2026.10)
100+ → 12
The change in the number of single-purpose agents converging into domain agents (Infor 2026.10 note)
These two numbers better illustrate the threshold than "how many Agents there are": whether it can actually do the work depends on whether the underlying layer has structured, callable, permissioned business capabilities; converging from "a pile of small agents" into "a few domain agents" is an architectural regression forced by operational complexity. The prerequisites given by Decision Intelligent (11-04) are the same: whether APIs are available, whether data is clean, and whether freeze codes and order-following rules have been defined.
III. Open Protocols Are Not About Labels: MCP, A2A, and the Dividing Line of "Compliant Implementation"
In this release, what the open-source camp should pay most attention to is the interface layer. In its October release notes, Infor states: it supports the A2A protocol to connect GenAI Assistant to third-party assistants such as Microsoft Copilot and ServiceNow; MCP provides a single governed interface, replacing "each agent doing its own integration once"; and there are more than 100 new or updated APIs for IT teams to connect AI, analytics, and automation tools. GenAI Assistant becomes the default landing page of the Infor OS portal and comes with Knowledge Hub and guardrails.
Why this relates to open source software: Only when ERP exposes business capabilities through open protocols and open APIs does the question of "which model, which assistant, which implementation partner" begin to be decoupled. Enterprises can self-host models, connect their own assistants, and let industry knowledge accumulate in their own repositories, without being locked into a single vendor's interface. This is the same logic that open source ERP has always advocated—"data self-control, no lock-in"—except that this time a commercial vendor has opened a gap from the middle.
But the open protocol is being used very loosely. This week, The Agent Times compiled a set of cases of independent developers welding MCP into accounting/ERP backends: Frihet released the MIT-licensed npm package @frihet/mcp-server, exposing 31 tools across six types of resources: invoices, expenses, customers, products, quotes, and reports; PaperLink exposes 118 tools across 12 domains; one developer forked Intuit's official QuickBooks MCP Server and expanded it to 29 entity types, 143 tools, and 335 unit tests; the MCP-Nest module in the NestJS ecosystem has 40 contributors and 616 stars.
In the same report, developer ShatilKhan was candid about his NestJS accounting app: it is not a compliant MCP server—no stdio/SSE transport, no official SDK, no JSON-RPC negotiation; it is merely a "facade" resembling a function-calling specification. The entire value of the protocol lies in interoperability: an "MCP" that can only work with a self-developed frontend is functionally just another API wrapper layer. Project data is self-reported by the developer and unaudited.
The conclusion is straightforward: open protocols lower the cost of integration, and also lower the cost of "pretending to be integrated". The question is not "does it support MCP," but whether any compliant client can discover and invoke it, whether permissions inherit ACLs, and whether write operations have a human approval gate.
IV. Practical Recommendations for CIOs／Digital Leaders
1. First do the interface accounting, then discuss the autonomy level. Break the processes to be handed to agents into four columns: "which objects to read, which APIs to call, which table to write, who approves." For processes that cannot be listed out, do not enable autonomy yet—what they lack is not model capability, but a structured business definition. You can use the granularity of the 12 domain agents as a reference; don't start by rolling out 100 single-point tools.
2. Write "standardized core + governed extensions" into the contract terms. Require vendors to explain: which capabilities are in the standard core, which belong to extensions, whether extensions need retesting during upgrades, whether MCP/A2A interfaces remain stable across versions, and whether write operations can be disabled by role. The core ledger remains rollback-capable, with differentiation placed in a versionable extension layer.
3. Conduct a compatibility stress test on the open protocol.Use a third-party compliant client to actually connect once, verifying tool discovery, identity inheritance, write operation confirmation, and log export; require manual confirmation to be retained for every high-risk action (payment, posting, bulk changes to master data, external correspondence). This also aligns with the three hard constraints of "controllability / full-process traceability / digital passport" in China's mandatory national standard for agent security in October.
4. In the acceptance criteria, clearly state "who tested it". List the vendor's claimed cost reduction, speed improvement, and inventory improvement separately from your own actual measurements. First run a narrow scenario (for example, quote-to-order for 20 high-frequency SKUs), use four metrics—completion rate, exception rate, number of manual interventions, and cost per transaction—for acceptance, and then decide whether to expand.
The three sets of claims must be kept separate; do not combine them into a single conclusion: the >50% and -40% for 2030 are institutional forecasts; "reducing picking travel, lowering cost per order, and improving inventory turnover" are Infor's self-reported benefits; on November 4, ERP News explicitly stated that this distribution announcement did not provide customer-measured results, nor did it provide a complete product–entitlement/availability matrix. In addition, Gartner reports do not endorse any vendor, and a quadrant position is not the same as fit. Leaving a gap between forecasts, self-reports, and third-party observations is the most cost-saving habit in this round of selection.
This week no model refreshed the leaderboard; instead, a more practical matter was put on the table: whether ERP will be autonomously executed by AI—the answer does not lie with the models, but in whether there is a clean, callable business core with permissions. Only by first organizing interfaces, data, and rules into assets that machines can understand can autonomy levels be meaningfully discussed. Otherwise, that percentage in 2030 is just a number in someone else's report.
Sources: Gartner, “Magic Quadrant for Cloud ERP for Product-Centric Enterprises” (Tomas Kienast et al., 2026-11-02) and disclaimer; Infor official press release and resource page (2026-11-06, including market definition and two 2030 predictions); Microsoft Dynamics 365 official blog (2026-11-06); Infor Distribution and Automotive 2026.10 release notes; ERP News (2026-11-04, including the note that “no actual customer test results were provided”); Decision Intelligent evaluation checklist (2026-11-04); The Agent Times accounting/ERP MCP case compilation (week of 2026-11, developer self-reports unaudited). All year-specific percentages are institutional forecast values; vendor self-reports, third-party reports, and forecast values are listed separately, with no extrapolation. As of 2026-11-08.
Zhongdi Information · Dedicated to Odoo and digital solutions for enterprise management
