Chinese automakers are entering 2026 with a new priority: making artificial intelligence work in daily operations, not just showcasing the latest models. According to industry data from Volcengine, its Doubao large model is now installed in more than 7 million smart vehicles across over 50 brands and 145 models, with more than 30 million daily in-cabin interactions. But behind these impressive numbers, a different story is emerging.
Industry research cited by the Chinese outlet Gasgoo suggests that while about 85% of companies have deployed generative AI in some form, only 5%-10% have moved beyond pilots to achieve scalable value. The focus of industry discussions is shifting from 'whose model to use' to 'how to actually make AI useful.'
The 'pilot paradox' hits the automotive industry
A May 2026 study on China's automotive intelligence experience, cited by Gasgoo, found that while the penetration of advanced in-car features is rising, user satisfaction has not kept pace. The report notes that users value 'stable and smooth interaction, accurate and reliable execution' far more than low-frequency feature demonstrations and lengthy learning processes. Model parameters, debut models, and concept videos may dominate press conferences, but they do not necessarily translate into daily value for car owners.
This phenomenon is not unique to the automotive industry. McKinsey has labeled it the 'pilot paradox,' attributing the root cause not to technology but to organizational and human factors. Cui Dongshu, secretary-general of the China Passenger Car Association, said that automakers' digitalization efforts commonly suffer from 'insufficient top-level design, weak data governance, and low scenario penetration.' Industry research further shows that data from smart driving logs, battery charge and discharge curves, charging infrastructure conditions, and annual inspection records are scattered across different entities, making it difficult to form a complete lifecycle digital profile for a single vehicle.
Digital employees: the new frontier
At the AI-Defined Automobile Forum organized by Gasgoo, Liu Xinlong, head of Volcengine's AI application solutions for automakers, noted that as agent architectures mature, AI is evolving from a single-point assistive tool to a 'digital employee' that can deeply participate in business operations. He emphasized that companies should focus on 'compound interest engineering'—continuously structuring business processes and accumulating execution manuals. This involves breaking down business tasks to a level AI can understand and converting human workflows into executable manuals, clarifying which steps are handled by AI and which require human intervention.
In 2026, 'digital employees' are no longer just a buzzword within automakers. China FAW has deployed 29 digital employees and 40 agent applications, with over 20,000 employees collaborating online. Its first digital employee, a 'production plan administrator,' has structured 34 business rules covering 14 Hongqi models, enabling real-time order intake and second-level production scheduling. This has reduced manual approval nodes by 60%, compressed R&D and production cycles by 50%, and lowered manufacturing costs by 40%. Geely Automobile Research Institute reports that after introducing AI digital employees, overall R&D efficiency improved by about 30%, with code-writing efficiency up by 50%. Changan Automobile has integrated office agents across R&D, manufacturing, supply chain, sales, and service scenarios. Earlier, Farizon, a commercial vehicle brand, and Alibaba Cloud jointly launched the industry's first IP-based AI employee, 'Chengcheng.'
Liu Xinlong also highlighted a detail often overlooked: in large-scale deployment, systems do not call the strongest, latest model for every task. Instead, they automatically schedule across different generations of models based on task complexity, 'protecting quotas and costs.' He stressed that a working digital employee requires four elements: a model providing 'basic intelligence,' an enterprise security architecture ensuring connectivity and control, contextual office data to lower the usage barrier, and agents that connect business systems into a closed loop.
As the technology cycle accelerates—with concepts like large models, VLA (vision-language-action), and physical AI refreshing quarterly—the consensus among researchers and practitioners is forming: the gap between automakers is shifting from 'who has larger model parameters' to 'who has turned business knowledge into AI-callable assets.' For a company investing hundreds of billions of yuan annually in R&D and managing dozens of vehicle platforms, data governance and process reengineering are slow efforts that take three to five years. But this slow work creates long-term barriers: when the foundation is solid, every model upgrade becomes a free improvement; when it is not, even the most spectacular launch event is just the beginning of the next 'pilot paradox.'
