AI's Easy Wins Are Over: Why Enterprise Adoption Now Demands Software Infrastructure

AI's Easiest Battles Are Nearly Won

Recent events signal a shift: Salesforce's stock surged 22% after earnings, with Agentforce and Data 360 combined ARR nearing $3.9 billion, up over 210% year-over-year. Salesforce also partnered with Anthropic to launch Claudeforce, charging API fees for each Claude call into Salesforce. Anthropic CEO Dario Amodei publicly endorsed Salesforce. Simultaneously, OpenAI's Sam Altman stated OpenAI should become a platform company rather than compete with every software vendor. In China, Tencent's WorkBuddy launched an FDE (Forward Deployed Engineer) ecosystem and an open platform for enterprise apps, while ByteDance's Doubao integrated with Feishu and Alibaba's Qwen targeted enterprise directly. These moves converge on one insight: AI is entering a "deep water" phase where large model companies are no longer the sole protagonists; platform and software companies are rising.

Why the Easy Wins Are Saturated

Over the past two years, AI first conquered tech-sector tasks that are highly digitized, creative, and fault-tolerant — where human correction is cheap. But the real mainstream consists of traditional enterprises whose core business runs on SOPs demanding extreme accuracy and stability. Large models alone cannot crack this. Altman recently admitted over-optimism about adoption speed, blaming user attachment to familiar tools. The author counters that software professionals share the same inertia, yet adopted AI quickly because their work is inherently AI-friendly: information is online, fault tolerance is high, and practitioners are skilled with AI. Most traditional enterprises lack these conditions.

Barriers to AI in Traditional Enterprises

Beyond human inertia, the tasks themselves are ill-suited for pure LLMs. A traditional enterprise's core business is not generating text or code but reliably executing cross-department, cross-system SOPs. The article illustrates with "AI query" (AI 问数): business users initially excited to ask "which region's sales dropped most?" without waiting for data teams. Soon, hallucinations appeared — answers looked plausible but lacked traceable table references, metric definitions, or accountability for errors. This is not a prompt or model intelligence issue; querying, calculation, and verification still require deterministic systems like BI. Therefore, AI penetration into core business must rely on the enterprise's software foundation.

Even with a software base, deployment is not automatic. Real SOPs involve multiple departments, systems, and tacit knowledge in employees' heads. Enterprises often don't know how to select scenarios, organize data, or overcome organizational barriers. This explains the sudden popularity of FDEs: they perform on-site exploration and adaptation that traditional enterprises cannot do themselves. The FDE boom further proves AI adoption is entering the deep end.

Software Risks Persist but Incumbents Have Advantages

While "SaaS doom" narratives have been temporarily disproven, uncertainty remains. Model advances continuously lower software development costs — features that took teams months now take weeks. If development cost approaches zero, many software moats could erode. However, the author remains optimistic: AI is not a sudden disruption for software companies. They have had ample time to transform, and possess mature software foundations, scenarios, distribution channels, and sticky customers — clear advantages over AI-native startups. Beisen (北森) exemplifies this: AI interviewer, AI coaching, and AI leadership coach agents have become its most important growth engine. The author predicts AI will become a core revenue source for many listed software companies next year. "Software will be disrupted by AI, but excellent software companies will not."

A larger threat is client self-build (甲方自研). As models improve, enterprises' incentive to build internally grows. Previously they might achieve 30% of a commercial system; now they can reach 60% — cheap, controllable, and customizable. The most dangerous competitor for a software vendor may be its own customer. Smart software companies will not fight this trend but embrace it, becoming the infrastructure for client self-build, just as SaaS vendors built PaaS platforms in the traditional era. Some software companies will fall behind, but the industry as a whole will not be overturned.

The Truth Is Complex: Coexistence of Two Phases

The AI second half will stay chaotic because penetration in different domains proceeds simultaneously. Contradictory views exist: some say FDEs don't need business knowledge and AI doesn't need software infrastructure; others insist FDEs must know business and AI must deeply integrate with software. Both are right because they observe different samples. The first sample represents AI's first half: high information density, high fault tolerance, where LLMs alone satisfy enterprise needs. The second sample represents the second half: SOP-heavy, demanding extreme stability and accuracy, requiring LLMs to embed deeply into business and software infrastructure. Both sample types will coexist for a while, but the second type is the protagonist of AI's next chapter.