A post recently went viral in China. An HR professional spent three minutes building a recruitment agent. In thirty minutes, it screened six candidates from 310 résumés. Within a day, the post had racked up 300,000 views.
What made it striking was that the same company had previously spent 2.75 million RMB on a full-process AI office system, developed by a professional team and pushed hard for six months. It ended in chaos and was scrapped.
A 2.75-million-yuan professional system failed to deliver, while a three-minute personal tool worked. That contrast made one thing clear: the difficulty of enterprise AI adoption is not about technology. It is about people.
Put bluntly, the pattern is: leadership decides → outsiders step in → everything falls apart.
1. The Numbers Do Not Lie
Gartner reports that 85% of AI projects fail to deliver expected value or are abandoned before deployment. MIT is even more direct: 95% of AI pilot projects never reach production. McKinsey research also shows that while many companies are using AI, only 39% of respondents feel it has a meaningful impact on profit.
The problem is not whether AI can work. It is that these projects start going off track at the human level.
2. Leadership Decisions Are Often About Face
Many AI projects do not start from business needs. They start from leaders wanting to "adopt AI."
Take the 2.75-million-yuan full-process AI office system. At its core, it was a top-down decision to implement AI. The system was built around Workflow automation, aiming to put every business process in AI's hands, including recruitment. In practice, many steps could not be fully automated and still required manual intervention: collecting résumés, screening them, coordinating interview slots. After six months of forced rollout, error rates soared. At times it was less reliable than experienced staff doing the work manually. Eventually, the project was cancelled.
Why? Because leaders treated AI as a plug-and-play political achievement tool. It was not meant to solve a specific problem; it was meant to prove "our company has AI too." They saw others building large models at conferences and came back saying, "We must build one too." AI projects became vanity projects, not substance projects.
DingTalk's ONE project followed a similar path. It carried four goals at once: reducing employee workload, serving as an AI showcase, boosting morale, and exploring commercialization. These goals conflicted with one another. Internal organizational tension was constant, and the product positioning kept shifting. In a post-mortem, the departing product manager said ONE did not fail because of AI technology. It failed because there were too many goals and too much organizational weight. A tool meant to lighten employees' load became a new source of work pressure.
A side-by-side comparison makes the result clearer:
| Dimension | DingTalk ONE | SoloEngine |
|---|---|---|
| Model | Top-down "organizational engineering" | Bottom-up "personal tool" |
| Goals | Grand, multiple, and conflicting | Specific, singular, and clear |
| Users | Enterprise managers | Frontline employees |
| Design | "Chef-style" top-down decision-making | User-oriented, on-demand customization |
| Iteration | "Daily package," high-pressure, surface-level | Community-driven, flexible, fundamentals-focused |
| Cost | Huge upfront investment | Free, low barrier |
So the 2.75-million-yuan system failed because it tried to use AI to solve a complex organizational and management problem, and in doing so created even more problems. SoloEngine succeeded because it returned to the essence of a tool: focusing on the specific pain points of individual users. This confirms a trend: the value of AI lies not in how much it costs or how many features it has, but in whether it can truly empower people.
3. Programmers Understand Code, but Not Business
The real trouble begins when leadership's vision lands in the hands of programmers.
Technically minded people and business-minded people operate with fundamentally different brains. A young programmer can learn PyTorch and Transformers in three months. But three months cannot replicate a veteran employee's thirty-year case library. Technical skills can be learned quickly. Experience cannot.
What is worse, much business knowledge cannot even be articulated.
Again, take the 2.75-million-yuan system. It fell short of the HR professional's three-minute tool because even a skilled professional team could not grasp the unwritten rules of recruitment: HR knows which days of the week bring the most applicants, when in the year salaries are easier to negotiate, and which résumés are clearly mass applications. These things never make it into requirements documents, but they directly determine whether a system is usable.
Even if programmers could learn the unwritten rules and experience of HR, what about finance? Legal? Administration? Industries that are more specialized, more complex, and more experience-driven?
Today's AI Agent is like an intern with a stellar transcript but no feel for how a company actually works. Ask it to do something, and it follows instructions neatly. But the moment a critical judgment is needed, it falls short.
4. Not Replacement, but Augmentation
So what is the answer?
SoloEngine points the way: let the people who understand the business use AI as a tool.
Traditional AI projects aim to "replace experts": programmers encode expert experience into code, then use the system to replace the expert. But expert experience can never be fully encoded. SoloEngine takes a different approach: empower experts, and let them arm themselves with AI.
HR does not need to learn programming. They only need to express rules in natural language. The unspoken "feelings" and "unwritten rules" become executable filtering conditions. This bypasses the dead end of "knowledge transfer" and lets business knowledge create value at the source.
5. At the End of the Day
The difficulty of AI adoption in enterprises is not, in the end, a technology problem. It is an organizational problem. A people problem.
Leaders treat AI as a status symbol. Programmers treat business as data. In between, no one translates real needs clearly. The result is that a 2.75-million-yuan system is less useful than a three-minute tool built by an HR professional.
The future of enterprise AI adoption is probably not "throw money at programmers to write code." It is letting the people who understand the business shape AI into a tool that fits their hands.
As the viral post concluded: people who can use AI will replace those who cannot. The era of Vibe Everything has arrived.
Top comments (0)