Chinese companies adopted GenAI with ambition and speed. The result, analyzed across more than a thousand interactions with AI leaders in the country, is a detailed map of the most common mistakes. Only 8% of Chinese enterprises had deployed GenAI in production by June 2024, compared to a global rate above 20%. The gap is not technological: it is methodological. And the twelve documented pitfalls are transferable to any organization worldwide.
Obsession with short-term ROI
The first mistake is focusing too much on immediate tangible returns while losing sight of the overall strategy. ROI is a legitimate question, but when it becomes the sole evaluation criterion, it leads to excluding investments in talent development and long-term innovation that build real competitive advantage. The DeepSeek R1 case illustrates this point well: a model developed by prioritizing research over commercialization produced results that surprised the industry.
Failing to identify the real bottleneck
One company introduced an AI coding assistant, achieving a 10% improvement in the development process. It then extended the tool to more developers, increasing code output by 10%. The problem: it had to hire significantly more testers to handle the additional output, slowing down the entire process. The real bottleneck was not in code writing. Adopting GenAI without first identifying where the actual operational problem lies produces partial or zero benefits.
Prioritizing technology over business outcomes
One company invested 200 million renminbi in GPU resources upfront, anticipating future uses. Six months later, no concrete use case had been identified. When innovation is driven by IT without business ownership, results remain undefined and transformations stall due to a lack of clear metrics.
POCs that drag on without exit criteria
A proof of concept should not exceed three months. Organizations that invest a year in a POC without defining clear exit criteria find themselves trapped in a costly cycle, struggling to justify the spend to management. The practical rule is simple: if the best available model does not produce results close to expectations within a reasonable timeframe, it is better to abandon that use case and move on to another.
Democratization without governance
GenAI is accessible to everyone. This is both an advantage and a risk. When a company enters confidential business information into public tools like ChatGPT to gather competitive intelligence, it exposes itself to significant data leakage risks. Making GenAI internally accessible requires an AI literacy program and a clear usage policy before the tools are even distributed.
Too much focus on the foundation model
More than half of the Chinese companies surveyed were fine-tuning or developing their own models. The problem is that training a custom model requires enormous GPU resources and often produces marginal benefits compared to using existing commercial models. In RAG architectures, for example, a hybrid approach combining semantic search and keyword search outperforms a solution based exclusively on GenAI. The model is a component, not the entire solution.
