Gasgoo Munich- Over the past year, AI agents have rushed into new vehicles at an unprecedented pace. "They can chat, write poetry, and understand emotions" — these capabilities have become recurring talking points at countless launch events.
Yet the buzz at launch events hasn't translated into high-frequency usage. Driven by reliability and safety concerns, many drivers are actively scaling back their use of AI controls. The focal point of complaints has shifted from "not understanding" to "not executing correctly." As supply heats up and usage cools down, a rift is emerging.
Where exactly is this experience bubble coming from? And as the capabilities of large language models converge, where will the next competitive gap open up?
Hot Launches, Cold Usage
Amid a flurry of launches, a broader inflection point in penetration rates is approaching. The installation rate of pre-installed smart cockpits has already topped three-quarters, meaning the penetration dividend is peaking. With nearly every vehicle now sporting large screens and internet connectivity, and large models accelerating their integration, the question of "whether it exists" is largely settled. The issue of "whether it's actually useful" is now taking center stage.
Image Source: Volkswagen
The cockpit is where users perceive intelligence most directly. Wang Chao, a professor at the Beijing University of Technology, noted at the 2026 China Automotive Supply Chain Conference that it is the closest point to the user and the first frontier where software-defined vehicles take shape. Operating systems, over-the-air updates, and large models typically run here first. It also acts as the "translator" for autonomous driving, conveying what the car is thinking and when a human needs to take over. Precisely because of this, when the cockpit starts to "speak inaccurately and perform poorly," trust in the vehicle's overall intelligence suffers.
Yet real-world data paints a different picture. J.D. Power's 2026 study reveals that among user issues with advanced features, "instability and inaccuracy" account for 48% — the highest category. Its 2025 research also found that 56.5% of users have actively reduced their use of AI controls due to reliability and safety concerns. Meanwhile, a quarter of users view the lack of dedicated physical buttons for common functions as a design flaw.
A longer view makes the shift in pain points even clearer. Data from the "Report on the Development of China's Auto Parts Industry (2026)" shows that in 2025, the top user complaint regarding smart cockpit features was "hard to understand," at 36%. By 2026, across all advanced features, "inaccurate execution" had risen to 48%. While the statistical scopes differ, making direct comparison difficult, the trend is unmistakable: interface friendliness has improved significantly, but the reliability of execution has now been pushed to the forefront.
J.D. Power's 2025 study also broke down user evaluations of specific functions. AI-generated art and poetry — features heavily hyped in marketing — received an importance score of just 3.87 out of 5. In contrast, smart traffic-avoidance navigation, a function that "helps when it matters most," scored 4.27.
"Low usage rates are, in effect, users voting with their feet," Wang Chao said bluntly in an interview.
Underlying this is a broader shift in user criteria: moving from "usable" to "useful, reliable, and frequently used." Reliability and trust are becoming the primary demands.
Notably, the ability to "chat" will soon become standard — likely faster than multi-screen setups or voice commands did in the past. Wang Chao anticipates that large models will increasingly be treated as an off-the-shelf capability. It is already becoming common for a single brand to integrate two different models across various markets, or even within the same vehicle.
When conversational ability can be integrated cheaply, merely "having a large model" no longer constitutes a competitive moat. Selling points hyped at launch events today may soon become ubiquitous features that nobody notices.
Cracks in the bubble are appearing in the stark contrast between "hot supply and cold usage." Functions are piling up, yet few are used frequently. Debut numbers keep hitting new highs, but user trust isn't building at the same pace. Now that "having an AI agent" is no longer novel, the industry is forced to answer a more fundamental question: who actually defines the value of the smart cockpit?
Where Are the Roots?
Where does this experience bubble originate? Wang Chao ranks the root causes, putting engineering and product definition at the top.
The first culprit is engineering. "The speed at which features are loaded into cars is outpacing the speed at which stability is improved." Large models can iterate daily, but automotive-grade validation progresses monthly — the entire verification system hasn't kept pace with the rush of new technology.
This is a mismatch of rhythms. The consumer electronics era accustomed us to "launch first, iterate later," but cars are products that carry lives on public roads. Safety-related functions require rigorous automotive-grade validation. When the release cycle far outstrips the validation cycle, "inaccurate execution" becomes almost inevitable.
Another root lies in the starting point of product definition. "Everyone starts from the spec sheet: you have three screens, so I'll put in three. You integrate a large model, so I'll integrate one. Naturally, the results look increasingly alike." Wang Chao argues that the true starting point should be the tasks users need to complete, not a checklist of components.
In his view, marketing and launch cadence are merely "amplifiers," not the root cause. They magnify homogeneity and exacerbate the gap between user expectations and actual experience.
The fundamental difference between automobiles and consumer electronics means the smartphone model cannot be simply copied. Cars have life cycles of 5 to 10 years and rigid boundaries of liability. While experience from consumer electronics interaction layers can be migrated — such as intent-first, edge-first, and open standards — the "fast iteration, weak safety" boundary cannot. Wang Chao's principle is "strong verification for high risk, fast iteration for low risk."
This distinction is particularly critical as AI shifts from "dialogue-based" to "action-based." A chatty error is merely awkward. But once AI can directly control the vehicle body or driving-related functions, the cost of a misjudgment is entirely different. As capabilities extend toward "taking action," the demand for certainty rises — which is why "inaccurate execution" is being magnified.
A popular saying holds that "autonomous driving will converge, but the cockpit will define personalization." Wang Chao only agrees with half of that. The safety baseline for autonomous driving is indeed converging; L2 combined driving assistance penetration has exceeded 70% this year, a mandatory national standard for combined assistance will take effect in January 2027, and many brands are even offering assisted driving for free. However, the cockpit and driving systems are fusing. Explaining the car's judgment to the driver and requesting a takeover is the cockpit's job — the two cannot be separated. This means cockpit personalization cannot detach from the baseline of "accuracy"; no matter how distinct the persona, it must be built on certainty.
A deeper mismatch lies in the industry's habit of defining products based on an "average user" persona, while real users have long since diverged.
The report categorizes users into four scenario types. The "commuter efficiency" type wants low interference and high certainty, treating the cockpit as a reliable driving partner. The "family companion" type seeks to upgrade the cockpit into a mobile family space. The "self-pleasing" type values emotional resonance, personalization, and immersive content. The "business and silver" composite type prioritizes privacy, reliability, and ease of use.
The definition of a "good cockpit" varies almost entirely among these four groups. Starting from a single spec sheet, it is difficult to hit the actual needs of any one group.
The disconnect in business models is also indirectly fueling the bubble. Wang Chao points out that software subscriptions currently serve more as an "entry ticket" to premium projects rather than a profit source. After comparing three listed cockpit companies, he found that none saw an increase in service revenue share in 2025. The sequence of the business loop cannot be reversed: retention and OTA frequency must come first, followed by service revenue.
When subscriptions struggle to monetize and R&D investment must continue, feature launches and marketing buzz often gain visibility faster than actual experience. This creates fertile ground for the bubble to grow quietly.
Shifting Barriers
When the ability to "chat" no longer serves as a barrier, where will the competitive gap emerge? Wang Chao believes this round won't simply replay the old paths of multi-screen setups or voice control, because the real barriers have shifted elsewhere.
The first barrier stands at the execution layer. The real gap isn't in understanding a sentence, but in calling tools and orchestrating vehicle controls and services within controlled permissions to see a task through from start to finish. Moving from "can converse" to "can execute" may seem like a short step, but it tests an entire system of capabilities.
Image Source: Bosch
Equally critical is trustworthiness. "The AI in the car must learn to stay silent when appropriate, speak accurately and clearly when it matters, and ask for confirmation when unsure." It is precisely this hurdle that has caused 56.5% of users to scale back their use of AI controls.
Determining whether a car gets "smarter with use" is long-term operations. The ability to continuously loop back scenario data and safely upgrade versions is what separates continuous evolution from stagnation at the point of delivery.
To rebuild this capability, the industry must first change its metrics. "Stop comparing who has an extra screen; compare function completion rates and interaction success rates." Wang Chao argues that the industry should replace spec sheets with metrics like function completion rates, interaction success rates, and reuse rates. In an era of converging AI capabilities, "doing it right" is the ultimate differentiator.
Differentiation sometimes means subtraction. Keep physical buttons for common functions and return the driver's attention to the road. This aligns perfectly with the shift in HMI primary principle from "showcasing technology" to "reducing distraction."
The adoption of standards is establishing a quantifiable common language for "usability." The "Artificial Intelligence Terminal Intelligence Classification Part 7: Automotive Cockpit" (GB/Z 177.7—2026), released in 2026, divides cockpit intelligence into four levels: L1 Response, L2 Tool, L3 Assist, and L4 Collaborative. This is a promising start, making "usability" measurable and comparable.
Wang Chao cautions that every prompt and status display in the cockpit carries compliance implications; personalization cannot come at the cost of certainty. "Personality must be built on the foundation of accuracy." When an evaluation system allows consumers to understand and buy based on these metrics, while enabling enterprises to escape the spec race, the experience bubble will truly lose its soil.
The business loop also requires a pragmatic path. For the next three years, the bulk of per-vehicle value will remain in hardware and platforms, but growth will come from structural upgrades. High-compute domain controllers and cabin-driving fusion platforms, regulation-driven driver monitoring systems, and evaluation-system-driven HUDs and forward displays represent three areas of high certainty.
Wang Chao also advises that when calculating the economics of the cockpit, one shouldn't fixate solely on feature payments. The value of cockpit definition is currently realized first in scale and retention, not directly in pricing premiums. A more complete ledger must account for how much customer acquisition costs are lowered, how much repurchase rates are lifted, and how much service revenue is generated.
With subscriptions still immature, companies have three realistic paths to support investment. First, rely on platform reuse — reuse rates are the only real economies of scale for parts suppliers. Second, rely on overseas projects, where gross margins for leading companies are nearly 9 percentage points higher than domestically. Third, bundle long-term maintenance and OTA services into the platform price rather than hoping to sell subscriptions separately. In terms of product strategy, "standardizing basic features and charging tiered rates for value-added services" is the safer choice.
"User trust is slow to build and quick to lose."
Solving the experience bubble won't be found in the specs of the next launch event, but in the daily routine of "getting it right the first time." When the industry stops using screens and slogans to generate hype and instead treats function completion rates, user trust, and long-term operations as true competitive advantages, the smart cockpit can finally move from "launch-event intelligence" to "everyday intelligence."
That is the only path forward for the smart cockpit in its next chapter.
