In the middle of this year, McKinsey threw out a striking judgment: ERP, as we know it, is coming to an end. At first glance, this sounds like a consulting firm's clickbait, but a set of Gartner data gives it credibility—in 2026, 62% of cloud ERP spending went to AI-native platforms, whereas two years earlier this proportion was only 14%. In twenty-four months, a forty-eight percentage point shift in share.
If this trend is true, then the logic by which enterprises choose ERP will have to be completely rewritten. Today I won't talk about module lists; instead, I want to explain clearly what this "center migration" really means through three signals that are currently occurring.
Signal One: Architectural Reversal — the Model Becomes the System's Reasoning Hub
What is the essence of traditional ERP? It is a business system built around a relational database. Table structures are designed for efficient storage, screens are designed for manual data entry, and business logic is wrapped around the tables. After AI came along, the traditional approach has been to attach a chat box on the side—it queries accounts and summarizes them, but it cannot understand the relationships between one account and another.
AI-native ERP reverses this relationship. ArcBeta describes it as "model-centric": the database recedes to the background, becoming a byproduct that must be documented for compliance and audit; what actually performs analysis, judgment, and recommendations is a Large Accounting Model (Large Accounting Models, LAN) trained on accounting-domain data. It reads not only transaction records, but also account structures, tax classification logic, reconciliation conventions, and multi-entity consolidation rules.
The difference is like that between a calculator and an analyst. Both produce numbers from inputs, but only the latter can look at the numbers and say: "This looks a lot like last year's Q3 restructuring; let me flag the variance for you to review first."
The most direct experience is in month-end closing. Batch-style month-end closing used to take 5 to 15 business days, because transaction accumulation, reconciliation, and report generation were serial. Under AI-native, reconciliation becomes a continuous process—transactions are monitored as soon as they occur, anomalies are flagged in real time, and the trial balance rolls continuously. Early adopters report EBIT improvement of more than 5%, not through layoffs, but by freeing finance from manual reconciliation to do analysis that truly requires human judgment.
The implication for ERP selectors is straightforward: when evaluating an ERP, don't just ask "Does it have AI features?" Ask "Is its AI a bolt-on Q&A add-on, or does it truly understand my business context?" The former will sooner or later have to catch up at the architecture level; the latter will create a generational gap in continuous reconciliation and anomaly alerts.
Signal Two: Delivery Reversal — a Pre-Built Agent Library Wrapped Around the Legacy ERP
The architecture is being inverted, but most enterprises' ERP systems cannot be replaced tomorrow. The realistic delivery path is to put a "layer" over existing systems.
9 Dots Consulting's PRIME AI Agents is a typical example. It is an ERP operations agent library built on Microsoft AI, with more than 20 agents covering ten centers from customer interaction to finance and HR, three of which are already built into SAP Business One. Its delivery model is interesting: 80% prebuilt, 20% configured according to the customer's systems and rules.
What does it actually look like? The RFQ agent at the procurement center pulls quotes in various formats sent by suppliers onto the same baseline to compare prices and lead times, gives recommendations, and then hands them to buyers to start decision-making from an "already compared" state; the Stock Sense agent at the inventory center checks demand and on-hand stock in real time, warns of material shortages in advance, and pushes replenishment or transfer suggestions to the team for confirmation; the visual grading agent at the factory center uses a photo to grade against the company's own standards, and after human confirmation, directly records it in ERP.
Note one common point: every agent hands the "results" back to humans for review instead of posting them to the ledger itself. This "wrapped outside, human-in-the-loop" model is especially friendly to small and medium-sized manufacturers—it costs less and delivers results faster than building agents in-house from scratch, and it also brings the "permission boundary between agents and ERP" to the fore: which actions agents can take autonomously and which must wait for a human to approve need to be clearly defined at deployment.
This also echoes the value of the open-source approach. An open-source foundation like Odoo 20 is itself MCP-native and gives you 100% ownership of your data. Whether agents are self-developed or third-party, they can be connected in an environment where you can see them, modify them, and retain logs. Only with a clean foundation can agents connect reliably.
Signal Three: Commercialization reversal—AI has moved from PPT into cash flow
The first two signals are about capability and delivery, while the third signal is about money—and money that is verifiable by a third party.
CSC Financial's November 2 research report reviewed the latest quarterly results of overseas cloud vendors from a sell-side perspective: SAP already has more than 110,000 Joule Work Desktop users internally, productivity in core functions has increased by more than 20%, and year to date the number of PRs merged into code repositories by developers has risen by about 70%; ServiceNow's FY26 Q3 revenue was $4.106 billion, up 20.5% year over year, and it raised its end-2026 AI ACV target from $1 billion to $1.7 billion, while consumption of agentic AI in production environments rose 13x year over year.
The research report's conclusion is restrained yet critical: overseas cloud revenue, model invocations, and Agent deployment are growing in sync, and enterprise AI demand is spreading from office and development to core business processes; existing customers, business data, and workflows constitute the real support for AI commercialization. In other words, AI in ERP is no longer just a demo; it has begun to generate measurable contracts and consumption.
The research report also made a reminder: the capex growth of the four major manufacturers is faster than operating cash flow, and there is a time lag between capacity building and cash recovery. This is actually two sides of the same coin as the "acquisition bill" mentioned earlier—manufacturers are investing heavily in AI, and returns take time; as a buyer, what you should earn is the part of efficiency that belongs to you in its ability to deliver, rather than paying for its story.
The selection logic needs to be rewritten
Three signals point to the same thing: the center of ERP is shifting from the database to the model. For those making selections or planning today, there are three practical suggestions.
First, treat the "reasoning layer" as a first-class citizen in evaluation. Ask the vendor: Does its AI truly understand business context, or can it only put a Q&A layer on top of the database? For capabilities such as continuous reconciliation and anomaly alerts, can it produce verifiable results on your real data?
Second, use a prebuilt agent library to lower the barrier to implementation, but first draw clear permission boundaries. An agent library layered outside the legacy ERP is a pragmatic path, but on day one of deployment you must specify clearly: which actions can be autonomous, which must be human-reviewed, and who is responsible for data integrity.
Third, judge by verifiable results. Whether looking at SAP's 110,000 users, ServiceNow's 13x consumption, or your own project's month-end closing days and reconciliation manpower, recognize only evidence that can be aligned to specific processes and numbers, not vendor vision diagrams.
Return to the simplest statement: ERP has not disappeared; it is being rewritten. Whoever first puts away the old map of "database-centric" and takes out the new ruler of "model-centric" will, in the next round of digitalization, hit fewer pitfalls and realize more value.
