Salesforce spent $3.6 billion to acquire three AI companies, and Microsoft developed seven models in-house—the infrastructure for enterprise AI is undergoing a major shift.

In the past 30 days, three things have happened in the enterprise software market. They seem independent of each other, but when pieced together, they tell the same story—the infrastructure of enterprise AI is being rebuilt, and ERP and management systems are the core battleground of this restructuring.

First thing: Salesforce has completed 15 acquisitions in one go from May to now, the latest three being the purchase of Fin (AI customer service) for $3.6 billion, m3ter (consumption billing platform), and Contentful (headless CMS). Second thing: Microsoft has released seven MAI self-developed models, covering everything from reasoning to coding to voice, along with the Frontier Tuning enterprise fine-tuning solution. Third thing: IDC released an AI-ERP market report, with the global smart ERP market size at $89.7 billion, up 23.6% year-on-year, and domestic at 158 billion yuan, with AI function R&D investment accounting for over 41%.

I placed these three things within the framework of enterprise management digitalization and connected them, reaching a relatively straightforward conclusion: the infrastructure for enterprise AI is shifting from "externally attached third-party models" to "self-built + open protocols," and ERP and management systems are the tracks most affected by this shift.

Salesforce's Acquisition Stack: Not Buying AI, But Building the Billing Pipeline for AI Agents

Salesforce's recent acquisition logic, upon closer inspection, is not about "buying AI technology," but rather building a complete pipeline for AI agents from operation to billing to content:

Translate this combination: AI agents run via Fin → billed per interaction via m3ter → retrieve accurate company content via Contentful to answer customer questions. The three companies combined form a complete pipeline of "AI agent operation → billing → content supply".

The logic of this stack is important for business managers because it directly points to a change: the billing model of CRM and ERP is shifting from "per-seat pricing" to "per-AI-interaction pricing". Fin's valuation is 9 times ARR, not a per-head valuation method — this is a valuation based on the logic of "charging per AI conversation."

⚠️ There is a contradiction here

Salesforce claims Agentforce has an annualized revenue of $1 billion, but its stock price has fallen 30% this year, with actual usage rates remaining low. Meanwhile, Salesforce is aggressively acquiring companies while laying off staff from Agentforce, Mulesoft, and Marketing Cloud teams. The simultaneous acquisitions and layoffs indicate that Salesforce is betting on the AI agent route, but while making this bet, it is also shedding the baggage of its old businesses.

For enterprises using ERP, the implication of this change is: if your current CRM or ERP charges per seat, you will likely be pitched a new "per-interaction" pricing model in the future. Salesforce is already doing this (Copilot Credits metering system), and Microsoft is as well (Copilot Credits). The shift in billing model directly impacts your IT budget structure—per-seat pricing is a fixed cost, while per-interaction pricing is a variable cost with much greater volatility.

Microsoft MAI-7 Model: The "Decouple from OpenAI" Route for Enterprise AI

Microsoft released seven MAI self-developed models, covering reasoning, coding, images, speech, and transcription. The significance of this move lies not in the models themselves (to be honest, most enterprises won't care about the number of parameters in reasoning models), but in Microsoft reducing its reliance on OpenAI.

But what's even more interesting is the accompanying release of Frontier Tuning—enterprises can train models using their own private data within compliance boundaries, and the trained model belongs entirely to the enterprise, with Microsoft keeping no copy. One striking statistic: the customized Excel model improves efficiency by 10 times.

I translate this change into corporate management language: In the past, using AI meant "sending data to OpenAI's cloud and letting someone else's model do the calculations for you"; now Microsoft offers another path—"within the enterprise compliance boundary, train a model that belongs only to you using your private data, and keep the results within your environment." The difference between these two paths is not a technical issue, but a data sovereignty issue.

For ERP and management systems, the impact path of this change is clear: if your ERP runs on Microsoft Dynamics 365, in the future you will no longer have only the option of "calling the OpenAI API." Instead, you can fine-tune your own model within Azure's compliance boundaries to handle specific business logic for finance, supply chain, and HR. The model's inference results will not go to OpenAI's data center; the data stays within your own Azure tenant.

Behind IDC's 89.7 Billion Figure: The Paradigm Shift of AI-ERP

The IDC report provides a macro framework. Let me pick out a few data points most directly relevant to business managers:

Global smart ERP market size reaches $89.7 billion (+23.6% YoY), with domestic market at 158 billion yuan. AI function R&D investment accounts for over 41%. More than 60% of medium-sized manufacturing enterprises have incorporated AI capabilities as a core evaluation dimension in ERP selection.

The change behind these numbers is summarized by IDC with one concept: from "process online" to "decision online". Traditional ERP is a "system of record" — business runs online, data is searchable and traceable. AI-ERP takes a step forward: the system not only records data but also understands the business logic behind the data, making trend predictions and decision recommendations.

But the figure of 89.7 billion cannot be viewed solely in terms of scale; its structure must also be considered. IDC data shows that Yonyou ranks first in China's AI-ERP market share—not because Yonyou's AI technology is stronger than SAP's, but because local vendors are better adapted to domestic industry scenarios. Yonyou's YonGPT enterprise service large model is specifically trained for vertical fields such as financial accounting, supply chain forecasting, and manufacturing scheduling, clearly distinguishing its positioning from general-purpose large models.

After LIGAO FOODS adopted YonSuite, production order execution efficiency increased by 81.5%, and inventory turnover rate increased by 73%. A hardware processing enterprise with 120 million yuan in revenue reduced its raw material inventory turnover days from 68 to 34, releasing nearly 2 million yuan in working capital. These figures are not projections in a PPT; they are actual measured results already achieved.

To be honest: China's ERP market is undergoing a very clear stratification—domestic vendors (Yonyou, Kingdee, Digiwin) are catching up to or even surpassing international vendors (SAP, Oracle) in AI-native architecture and industry scenario adaptation. The latter's AI implementation effectiveness in deep manufacturing scenarios has only improved by 15%-25%, while the former can achieve 35%-52%. The gap does not come from the technology itself, but from the depth of accumulated industry experience.

The Security Storm of Anthropic Fable 5: Another Hidden Door for AI Entering ERP

One more thing that is easily overlooked but has a significant impact on enterprise AI deployment: Anthropic's Fable 5 and Mythos 5 were forcibly taken down by the U.S. government on national security grounds—jailbreak vulnerabilities could bypass security restrictions through a code review framework. Subsequently, Anthropic released the Fable 5 Commercial Edition (restricted version), which integrates an automatic fallback system: when the security classifier is triggered, it silently routes to the standard Opus model.

Impact on Enterprises

The problem is not the removal itself—most companies use the standard version of Claude and won't encounter Fable 5. The issue lies in overly aggressive security layers: routine code reviews and security code queries are also being blocked. If your ERP custom development team uses Claude to assist with programming, they may frequently trigger security blocks during entirely normal development work, rendering AI assistance ineffective.

The lesson from this is straightforward: when choosing AI tools, companies must consider not only model capabilities but also the controllability of security policies. How aggressive of security filtering can you accept? How many code review and compliance check steps in your business scenario might be falsely blocked? These questions need to be thought through before deployment.

Summary of Three Main Themes: How Business Managers Should View Them

Salesforce builds its billing stack, Microsoft's self-developed model decouples from OpenAI, and IDC says the 89.7 billion market is undergoing a paradigm shift — three lines converge at one point: The infrastructure for enterprise AI is shifting from "renting others'" to "building your own or using open protocols".

To enterprise managers, my advice is divided into three levels:

1. Billing Mode

The shift from seat-based pricing to interaction-based pricing has already begun. When preparing next year's IT budget, list the variable costs of AI interactions separately. Salesforce Copilot Credits, Microsoft Copilot Credits, and Anthropic Agent SDK independent billing — all three companies are moving in the same direction, charging based on "how much AI does," not on "how many people use it."

2. Data Sovereignty

Microsoft Frontier Tuning offers enterprises a new path: training private models with private data, keeping the results within their own environment. If your ERP data is a core competitive advantage (supply chain pricing, customer behavior, production formulas), this path is far more secure than "sending data to a third-party cloud." Odoo's MCP Server takes a different approach—an open protocol that allows multiple models to connect, but you control which data can be read or written by AI. Both paths are better than "handing everything over to a single provider."

3. Selection Strategy

60% of mid-sized manufacturing companies have incorporated AI capabilities as a core evaluation dimension in ERP selection. However, the selection criterion should not be "who has the most AI features," but rather "whose AI integrates most deeply with my business processes." SAP's Joule follows the Copilot route (prompt suggestions), Salesforce follows the Agent route (automated execution), and Odoo follows the MCP route (open protocol read/write)—each route has its own strengths and weaknesses, and the key is which one suits your business scenario.

Enterprise AI infrastructure is undergoing a transformation. This shift is not about chips or servers—it's about the logic of how AI integrates into enterprise management systems. Billing models have changed, data sovereignty options have changed, and selection criteria have changed. These three changes are happening simultaneously, and business managers cannot focus on just one while ignoring the other two. They are interconnected.