This article is crafted based on publicly accessible information and is intended purely for the purpose of informational exchange. It should not be construed as any form of investment advice.

A pervasive sense of unease grips the global enterprise software sector. Over the past year, BlackRock's index fund, the 'iShares Expanded Tech-Software Sector ETF,' which tracks the S&P North American Expanded Tech-Software sector, has plummeted approximately 28% from its zenith. The capital markets' message is unequivocal and unforgiving: Large language models may render the traditional SaaS business model obsolete.

The crux of the matter is this: If an AI agent can autonomously activate (a Chinese term meaning 'invoke' or 'utilize') tools, follow processes, and complete tasks, why should enterprises continue to foot the bill for a graphical interface that necessitates human interaction through clicks, drags, and form-filling?

Within a mere month of each other, two SaaS behemoths from opposite ends of the Pacific Ocean arrived at a remarkably similar conclusion: Dismantle their existing structures.

On March 17, 2026, in Hangzhou, DingTalk's founder, Chen Hang, announced at a launch event that the DingTalk team had dedicated an entire year to completely revamping the underlying architecture that serves 800 million users. They unveiled the world's first enterprise-grade AI-native work platform, 'Wukong,' and rebranded themselves as Agent OS—an operating system tailored for AI. 'In the past, humans relied on DingTalk for work; in the future, AI will utilize DingTalk for work,' Chen proclaimed.

A month later, on April 16, 2026, in San Francisco, SaaS giant Salesforce, after 27 years in operation, revealed its historic decision at the TDX Developer Conference: Headless 360. The company stripped every capability from its platform, transforming them into API endpoints, MCP tools, or CLI commands, thereby enabling AI agents to operate the entire system without ever opening a browser. This marked Salesforce's most significant and profound architectural transformation in its 27-year history—a veritable disassembly and reconstruction of its business model.

Two companies, one in the East and one in the West, independently chose the same trajectory: Eliminate the graphical interface and grant AI direct control.

01

From Button Clicks to API Calls

The essence of Headless 360 is straightforward: Since agents cannot click buttons, transform all buttons into callable code interfaces. Salesforce simultaneously released over 100 new tools and skills, granting external coding agents—such as Claude Code, Cursor, Codex, and Windsurf—real-time access to Salesforce's complete environment for the first time, encompassing data, workflows, and business logic.

Salesforce Executive Vice President Jayesh Govindarjan put it succinctly in an interview, noting that the decision was made two and a half years ago: 'Instead of concealing functionality behind a UI, expose it so the entire platform can be programmed from anywhere.'

This strategy closely mirrors DingTalk's approach. To position AI, rather than humans, as the primary operator, DingTalk underwent a comprehensive CLI transformation. This means 'Wukong' can natively operate thousands of DingTalk capabilities without simulating human clicks on a graphical interface, truly achieving 'communication as execution.'

Technologically, both companies constructed similar three-pillar architectures.

The first pillar is 'Build Any Way.' Salesforce offers over 60 MCP tools and 30+ pre-configured coding skills. DingTalk established China's largest enterprise-grade MCP capability marketplace, initially covering 6,000+ commonly used enterprise MCP capabilities, and open-sourced its CLI tool.

The second pillar is 'Deploy Anywhere.' Salesforce's Agentforce Experience Layer decouples agent behavior from presentation, enabling native rendering across six different interfaces—Slack, mobile apps, Microsoft Teams, ChatGPT, Claude, and Gemini—with a single experience definition. DingTalk's 'Wukong' similarly supports connections to mainstream IM platforms like WeChat and Slack, allowing users to remotely activate it for work.

The third pillar, the least visible yet most critical, is trustworthy scalable agents. Salesforce introduced Agent Script, a domain-specific language for deterministically defining agent behavior, which it has open-sourced. DingTalk built an AI-native file system from scratch, where every operation is traceable and reversible, with snapshot management enabling users to revert to any previous version instantly.

02

The Engineering Tension Between Determinism and Probability

The synchronization of these product roadmaps is no mere coincidence. Both companies are addressing the same fundamental issue: AI agents are inherently probabilistic systems, yet enterprises demand deterministic outcomes.

In interviews, Govindarjan highlighted a real issue often obscured by hype: Early adopters deploying agents in production environments found the systems brittle. 'They fear modifying agents because they don't know if they'll work 100% correctly afterward. All tests must be rerun,' he said.

This exposes the core contradiction of enterprise AI.

Salesforce's solution is Agent Script, a single-file state machine where enterprises can explicitly mark which steps must adhere to deterministic business logic and which steps can allow large models to reason freely. The language is versionable and auditable. Salesforce open-sourced it, and with clear documentation, Claude Code can already generate it natively.

DingTalk opted for a lower-level yet equally pragmatic approach: Full CLI-fication. In late March 2026, DingTalk and Feishu open-sourced their respective command-line tools on GitHub within 24 hours. This means AI agents can operate systems through standard command-line interfaces—a deterministic interaction method validated by developers for decades—rather than relying on the rapidly evolving MCP protocol.

This directly contrasts with the industry's popular 'ambient programming' approach, which advocates letting AI act freely and driving code generation through natural language. The giants' choice is: Lock down deterministic areas with deterministic languages while leaving flexibility for large models. This isn't compromise—it's stratification.

03

Leaving Choices to the Market: Pragmatic Bets in Open Strategies

When it comes to open ecosystems, both companies' pragmatic orientations are highly aligned.

Salesforce's openness at TDX is noteworthy. The platform now integrates models from OpenAI, Anthropic, Google Gemini, Meta LLaMA, and Mistral AI. The open agent framework supports third-party SDKs. The new AgentExchange marketplace combines 10,000 Salesforce apps, 2,600+ Slack apps, and 1,000+ agents and tools, backed by a $50 million Builder Fund.

Yet Govindarjan's assessment of MCP is refreshingly candid: 'Completely uncertain' whether it will become a standard. 'Many engineers see MCP as just a CLI with a nice wrapper... Some think CLI alone suffices.' Salesforce's strategy is pragmatically flexible: It doesn't bind itself to any single protocol but simultaneously offers API, CLI, and MCP access methods.

DingTalk's strategy is equally pragmatic. It fully opened AI capabilities like computing power, industry models, and MCPs, built China's largest enterprise-grade MCP capability marketplace, and launched the enterprise AI platform DEAP for one-stop creation, management, distribution, and operation of AI agents. DingTalk CEO Chen Hang put it bluntly: 'We will co-build industry models, MCP marketplaces, AI agent ecosystems, and AI hardware markets with partners to advance China's enterprise AI ecosystem toward 'myriad intelligences, unified prosperity.'

By not betting on a single standard or binding to a single protocol, leaving choices to the market, both companies adopt a shared survival strategy.

04

Fundamental Shift in Business Models

The deepest changes may lie not in technology but in fundamental business model shifts.

Salesforce has transitioned from per-seat licensing to usage-based Agentforce pricing. Agentforce charges $2 per conversation or uses Flex Credits for pre-purchased volumes. When work is executed by agents rather than humans, the logic of per-head pricing collapses: If humans no longer click buttons, why count heads to charge?

DingTalk is experimenting with a 'commercially deliverable, outcome-based payment' model, building an enterprise AI agent marketplace on its platform. Chen Hang stated plainly: 'DingTalk hopes all developers can leverage its open AI capabilities to help enterprises solve problems, boost efficiency, and reduce costs in the B2B market, enabling them to truly profit in the AI era.'

One company charges in USD based on usage; the other in RMB based on outcomes. Yet their cores align remarkably: AI-driven work cannot rely on per-head accounting.

Viewing both launch events together: Separated by a month and an ocean, their messages seem scripted from the same document. DingTalk disassembled itself into CLI instruction sets; Salesforce disassembled itself into API endpoints. One manages workflows for 800 million people in China; the other supports the business foundations of hundreds of thousands of enterprises globally. Ultimately, they independently chose the same path: Since humans no longer click buttons, turn buttons into code and let AI click them.

More intriguing is the path they tacitly understood (a Chinese term meaning 'tacitly') and abandoned.

Neither company attempted to build a 'better graphical interface.' No one said, 'We'll make the UI more intelligent.' On the eve of AI devouring software interaction paradigms, these SaaS giants simultaneously accepted a truth: The graphical interface itself may already be a relic of a bygone era.