China EV makers pivot to AI agents amid bruising price wars

by ZHOU Shuqi

Chinese electric vehicle makers, squeezed by prolonged price wars and narrowing margins, are turning to artificial intelligence to recast cars as autonomous "agents", shifting away from incremental upgrades to driver assistance or in-car entertainment in a bid to reset growth expectations among investors.

Executives at several Chinese carmakers argue that as hardware features become commoditized, future differentiation will increasingly depend on software systems that can anticipate user intent and act in the physical world, rather than simply respond to commands.

LI Xiang, chief executive of Li Auto, said recently that artificial intelligence is evolving from chatbots toward systems capable of acting. He likened cars to mobile machines operating in the physical world, saying they should be able to understand user intent and provide services proactively.

A similar view has been voiced by HE Xiaopeng, chief executive of XPeng, who has told staff that the industry is entering a phase of deeper AI integration. He has said that in-car interaction systems and autonomous driving technologies to converge into what he calls a "super intelligent agent", according to Chinese tech media reports.

The approach involves linking two systems that have largely evolved separately: user-facing AI responsible for interaction and services, and safety-critical driving systems that handle perception, planning and vehicle control. Executives say connecting decision-making directly to physical execution is necessary for vehicles to move beyond experimental features and operate autonomously in real-world conditions.

Both Li Auto and XPeng have begun reorganizing teams to support this approach. Li Auto has reshaped its autonomous driving division to mirror AI development workflows, while XPeng has merged its cockpit and driving units into a single AI organization to share data, tools and computing resources.

Industry experts said such convergence is likely to become more common. They noted that the vision-language models used in in-car assistants and the vision-language-action models explored in autonomous driving rely on overlapping capabilities, allowing some capabilities to be shared. New generations of automotive chips are also beginning to offer sufficient computing capacity to support both workloads on a single platform.

Cost pressure is another driver. With margins under strain, integrating systems could reduce hardware redundancy and lower overall costs, analysts said.

Tesla has taken early steps toward linking in-car AI with driving functions. In North America, spoken navigation requests can be interpreted by an in-car assistant and passed to Tesla's full self-driving system for execution. Engineers working on similar technologies in China caution, however, that such systems remain largely reactive, relying on explicit driver commands rather than inferred intent.

They say the next stage would involve vehicles drawing on multiple inputs — including behavioral patterns — to anticipate user needs and align those predictions with real-time traffic and environmental data, allowing cars to adjust their level of intervention dynamically.

Significant constraints remain. AI specialists warn that combining systems with very different risk profiles poses fundamental challenges. In-car assistants can tolerate occasional errors that affect user experience, but autonomous driving systems are safety-critical and require deterministic performance, stable latency and full traceability. Failures in such systems translate directly into safety risks.

A key concern is computing resource management. Conversational AI models can place heavy, variable demands on processing power, while autonomous driving requires consistent latency for every frame of sensor data. Without strict isolation, spikes in cockpit AI workloads could interfere with driving functions — a risk widely viewed as unacceptable in production vehicles.

Development cycles also diverge. In-car software is typically updated frequently as new services are added, while changes to driving systems must undergo lengthy testing and validation. Keeping both release cycles aligned remains difficult.

Most industry participants therefore expect progress to be incremental. In the near term, cockpit and driving systems are likely to remain partially independent, converging first at the infrastructure level through shared computing platforms and data pipelines, while maintaining strict safety isolation.

Analysts say a pragmatic path is to deploy AI autonomy first in low-risk, reversible scenarios, while retaining human confirmation for safety-critical actions. The approach may be less dramatic than the science-fiction vision often invoked by carmakers, but it could prove essential to building trust among regulators, consumers and investors as the industry adapts to the AI era.