For a long time, I held firmly to one belief: AI was premature for most enterprises today. I believed what many companies truly needed was not AI, but "traditional" digital transformation. However, as of 2025, after witnessing three years of explosive LLM development, I have to admit that my perspective needs to change.
Enterprise Evolution Path
In the past, I firmly believed that any enterprise's development must follow the path of "standardization -> automation -> digitalization -> intelligence". Whenever I saw companies attempting to transition to intelligence without even completing their digital infrastructure, I thought it would only lead to losses.
The underlying logic of this view is quite simple: intelligence relies on an enterprise's accumulated knowledge, which is the company's unique asset. To utilize this knowledge, the first step is to transform it into usable digital information through digitalization systems. Therefore, in my understanding, the ideal enterprise technology architecture should be a combination of "Small language models(SLMs) + mature digital systems."
The Disruptor
However, reality often surprises us. Today, we see organizations of all sizes, from startups to large enterprises, actively trying to incorporate AI. Interestingly, even companies in the market that haven't completed their basic digital infrastructure are pursuing AI with undiminished enthusiasm. This phenomenon sparked my deep curiosity: in the absence of a solid digital foundation, what role can AI play in these enterprises?
Surprisingly, most companies aren't even considering this from a technical architecture perspective. They are simply attracted by LLM's demonstrated capabilities and have started implementing it directly in their businesses. From my previous perspective, this approach would be too costly. LLM calls are expensive, the models aren't trained specifically for enterprises, there's a risk of "hallucinations", and companies need to invest substantial resources in building their own RAG systems.
But my observations revealed an unexpected outcome: this combination works surprisingly well.
Redefining Enterprise Knowledge
I believe one reason is that LLM's performance has far exceeded expectations. Consider that enterprise knowledge is essentially the best practices accumulated by companies over years, serving as their moat. But today's LLMs, by learning and synthesizing the world's best practices, combined with powerful reasoning capabilities, can now surpass individual enterprise knowledge in many scenarios. This suggests we need to redefine what we consider "enterprise knowledge."
Humans and Data
Another reason is that LLM's emergence has fundamentally changed how enterprises think about building technical architecture. I used to believe that technology's ultimate goal was to digitize enterprise knowledge and automate processes through digital systems. But in the LLM era, technical architecture has been simplified to a pure "LLM + applications" combination. Under this new framework, employees' work patterns have undergone a fundamental transformation: they primarily focus on designing appropriate prompts and flexibly dispatching LLM or application outputs across different scenarios. This transformation has essentially bypassed the traditional digitalization and process stages, creating an entirely new work paradigm.
Unstoppable Progress
The emergence of LLM has shattered our conventional understanding - it's not just a simple technological iteration, but a revolutionary technological innovation. While LLM hasn't yet brought many tangible benefits to society, it already shows the potential to revolutionize enterprise organization. I'm not sure if this is a good thing; I just feel that AI's development is progressing at an unsettling pace.