Gasgoo Munich- By 2026, embedding large models in cars is old news in China's auto industry.
Volcengine reveals its Doubao large model is now installed in more than 7 million smart vehicles, spanning over 50 brands and 145 models, with cabin interactions exceeding 30 million times daily.
Yet behind those flashy installation figures, industry research points to a different reality: while about 85% of companies have deployed generative AI in some form, only 5% to 10% have moved beyond the pilot stage to deliver value at scale.
The industry's focus is quietly shifting from "whose model to use" to "how to actually make AI work."
Easy to Install, Hard to Implement: The "Pilot Paradox" Behind the Hype
According to the "China Automotive Intelligentization Experience Study," released by a leading authority in May 2026, the penetration rate of advanced in-car features has climbed year over year — yet user satisfaction hasn't kept pace. That intriguing contrast demands the industry's attention.
For users, the report notes, "stable, fluid interaction and precise, reliable execution far outweigh low-frequency flashy tricks and steep learning curves." The stars of press conferences — model parameters, debut models, and concept videos — don't necessarily translate into the value drivers perceive every day.
This isn't unique to the auto sector. McKinsey describes the phenomenon as the "pilot paradox" — a problem rooted not in technology, but in organization and people.
In the automotive world, the bottlenecks are more specific.
Cui Dongshu, secretary-general of the CPCA, puts it bluntly: automakers generally suffer from "insufficient top-level design, weak data governance, and low scenario penetration." Industry research goes further, showing that for a single new energy vehicle, data on autonomous driving logs, battery charge-discharge curves, charging station status, and maintenance records is scattered across different entities. The industry still struggles to form a complete, full-cycle digital portrait of even one car.
At the AI-Defined Vehicle Forum hosted by Gasgoo, Liu Xinlong, head of Volcengine's automotive AI solutions, noted that as Agent architecture matures, AI is evolving from a single-point support tool into a "digital employee" capable of deep business participation.
At the same time, Liu cautions that companies must commit to "compound engineering" — the continuous structuring of business operations and the codification of execution manuals. Businesses need to break down operations to a level of structure AI can understand and translate human workflows into executable manuals, defining clearly which steps belong to AI and which require human intervention. The more solid this groundwork, he emphasizes, the stronger the data and process foundation for scaling AI capabilities later — ultimately unlocking the true value of human-AI collaboration.
The Real Starting Line for Digital Employees: Turning Business Knowledge into AI Assets
By 2026, "digital employees" are no longer a gimmick within automakers.
Image Source: FAW Group
FAW has launched 29 digital employees and 40 intelligent agent applications, connecting over 20,000 employees online. Among them, digital employee No. 001, the "Production Plan Administrator," structured 34 business rules covering 14 Hongqi models. It enables real-time order intake and second-level production scheduling, trimming manual approval nodes by 60%, slashing R&D and production cycles by 50%, and cutting manufacturing costs by 40%.
Similar practices are taking root across multiple automakers.
Image Source: Huaban
Geely Automobile Research Institute reports that after introducing AI digital employees, overall R&D efficiency rose by about 30%, with coding efficiency jumping 50%. Human-machine teams now cover embedded development, cloud applications, and smart cockpits. Changan Automobile has also deployed office agents across five scenarios: R&D, manufacturing, supply chain, sales, and service. Earlier still, Farizon Commercial Vehicle partnered with Alibaba Cloud to launch "Chengcheng," the commercial vehicle industry's first IP-based AI employee.
Liu points out a frequently overlooked detail: in true large-scale deployment, systems don't necessarily call on the strongest, newest model for every task. Instead, they automatically schedule across different generations of models based on task complexity — a move to "protect quotas and costs."
In his view, a functional digital employee requires four elements: the model provides "base IQ," the enterprise security architecture ensures connectivity and control, accumulated office context lowers the barrier to use, and the Agent weaves various business systems into a closed loop.
In other words, once the foundation is solid, rapid model iteration becomes an upgradeable bonus. Conversely, the more you rush to swap out the "engine," the more likely you are to spin your wheels in the mud.
Technology cycles wait for no one. In 2026's auto industry, concepts like large models, VLA, and physical AI are still refreshing on a quarterly basis. Yet a consensus is forming among researchers and practitioners alike: the gap between automakers is shifting from "who has the larger model parameters" to "who has crystallized their business knowledge into callable AI assets."
Conclusion:
For an automaker pouring tens of billions into R&D and managing dozens of vehicle platforms, data governance and process reengineering are slow, grinding tasks measured in three to five years.
But that slow grind builds a lasting moat. When the underlying engineering is solid, every model swap is a free upgrade. Without a firm foundation, even the most glitzy product launch is just the beginning of the next "pilot paradox."
