In the past two years, global enterprises have continued to increase their investment in the field of AI. According to IDC statistics, global enterprise AI spending is expected to exceed $200 billion by 2025, but McKinsey's survey shows that about 70% of enterprise AI projects have not achieved scale implementation. Technology itself is no longer a bottleneck - GPT-4 level language models, mature computer vision APIs, AutoML platforms, and other technological tools are within reach. The real obstacle lies in the organization's internal data governance system and change management capabilities.
##Data governance: the foundation of AI
The performance of AI systems is highly dependent on data quality. Many companies only realize that their data foundation is far from meeting the standards after launching AI projects. Common data governance issues include:
**Data silos are the biggest structural obstacle. The data of large enterprises is usually scattered across multiple incompatible systems such as CRM, ERP, supply chain management systems, customer service platforms, etc., lacking unified interfaces and data standards between each other. A manufacturing customer discovered during the implementation of an AI quality inspection project that their quality inspection data distributed across three business units used completely different coding systems and defect classification standards. Data alignment alone took up half of the project cycle.
**The quality of data annotation directly affects the effectiveness of the model. Many companies outsource their labeling work to low-cost teams, but neglect the development of labeling standards and quality inspection mechanisms. Taking the risk control model of a certain fintech company as an example, the consistency of labeling "high-risk transactions" in its training data is only 75%, which directly leads to a much higher than expected false alarm rate of the model in the formal environment.
**Data privacy and compliance are another constraint. With the successive implementation of the Personal Information Protection Law, the Data Security Law, and the EU AI Act, enterprises must meet compliance requirements at every stage of data collection, storage, processing, and cross-border transmission. In practice, this means the introduction of technology solutions such as data anonymization, differential privacy, and federated learning - each of which significantly increases the technical complexity and implementation cycle of AI projects.
##Organizational Change: A Challenge More Difficult than Technology
Technology selection can be outsourced, but organizational change must be completed internally. The most easily underestimated factor in the process of implementing enterprise AI is the human factor.
**The role conflict of middle-level managers is particularly prominent. Many companies' AI projects are driven from top to bottom by CEOs or CTOs, but middle managers often face a dilemma - on the one hand, they are required to support AI transformation, while on the other hand, they are concerned about their teams and functions being replaced by automation. When a retail enterprise was deploying an AI inventory management system, the regional manager passively cooperated due to concerns about being laid off, resulting in continuous loss of core data in the system and a six-month delay in project progress.
**The friction of cross departmental collaboration cannot be ignored. AI projects naturally require close collaboration between IT departments, business departments, and data teams, but these three often have vastly different work rhythms and performance metrics. The IT department focuses on system stability and security, the business department pursues quick and effective ROI, and the data team tends to spend time polishing model accuracy. When there is a lack of effective coordination mechanisms, projects are prone to a vicious cycle of "IT says data is not enough, business says system is not working well, and data says business is not cooperating".
**The skill gap is a rigid bottleneck for the implementation of enterprise AI. Even if advanced AI tools are purchased, if the team lacks operational personnel who understand basic machine learning principles, engineers who can write high-quality prompts, and business analysts who can interpret model outputs, the effectiveness of the tools will be difficult to unleash. According to LinkedIn's 2025 Global Talent Report, the demand for AI related skills is growing at 3.2 times the supply rate, and the supply-demand imbalance is difficult to alleviate in the short term.
##Experience insights from successful cases
In overcoming the above challenges, some leading enterprises have accumulated valuable experience for reference.
The AI risk control project of a leading commercial bank adopts a strategy of "small steps, fast running+business embedding". At the beginning of the project, the AI team was directly embedded into the credit approval business group, working together with the approval manager, and spent three months polishing a risk control auxiliary model covering a single product line. After the model was launched, the approval efficiency increased by 40% and the bad debt rate decreased by 15%. With this' model room ', the internal promotion resistance has been significantly reduced, and it will gradually expand to 12 product lines throughout the bank in the next two years. The key success factor is to make the business team the leader of the project rather than the passive recipient.
The AI predictive maintenance project of a multinational manufacturing enterprise has taken a "data infrastructure priority" approach. Before starting any AI model development, the enterprise spent 18 months unifying the equipment data standards of 37 factories around the world, establishing a central data platform, and deploying edge computing nodes for real-time data collection. Although there was a huge initial investment, the development and iteration speed of predictive maintenance models significantly increased after the data was ready, covering 80% of critical equipment within two years and reducing unplanned downtime losses by approximately 230 million yuan per year.
##Practical suggestions
For enterprises planning to launch AI projects, the following suggestions have universal reference value:
1. * * Conduct data audit first * * - Before selecting AI tools, evaluate the quality, coverage, consistency, and accessibility of existing data assets. When data preparation is insufficient, AI projects are destined to achieve twice the result with half the effort.
2. * * Choose the right entry point * * - Choose a scenario with clear business value, relatively mature data foundation, and limited impact of failure as the first AI project. Low hanging fruits can not only build confidence for the team, but also accumulate reusable methodologies.
3. * * Set up an AI transformation officer or similar role * * - requires a cross departmental coordinator to break the gap between IT and business
4. **建立渐进式学习机制**——不要等待"完美的大模型"，而是先用现有工具跑通最小可行产品（MVP），在迭代中验证业务假设、积累训练数据、培养团队能力。
【参考来源】IDC全球AI支出指南、麦肯锡AI现状调查报告、LinkedIn 2025全球人才报告