Automotive AI is moving beyond voice assistants that wait for a spoken command, with Smart Eye demonstrating an in-cabin agent that combines information about the driver, passengers, vehicle, and surrounding environment before deciding how to respond.

Smart Eye, the Gothenburg-based driver-monitoring specialist, unveiled the technology at InCabin Europe in Barcelona on 23 September. Its demonstration combines voice with driver identity, gaze, attention, emotion, driver state, and occupancy, while also drawing information from sensing systems inside and outside the vehicle.

The additional context is intended to narrow the amount of information an AI agent needs to process before taking an action. Smart Eye argues that giving a model a clearer picture of who is speaking, what they are looking at, and what is happening around the vehicle can reduce the computing required to interpret a request rather than forcing a general-purpose model to reconstruct the situation from language alone.

The demonstration can carry out tasks through connected vehicle functions and services, build dashboard widgets from live driver and cabin data, and combine eye tracking with exterior sensing to establish what a driver is referring to outside the car. Smart Eye says the underlying platform can work with different foundation models and voice systems, with processing divided between local and cloud infrastructure.

That architecture pushes automotive AI into a different engineering problem from the conversational assistants already familiar in vehicles. A useful in-car agent has to interpret incomplete instructions while the driver is moving, understand whether attention is on the road or elsewhere, determine which passenger is speaking, and decide whether a requested action is appropriate in the current situation.

Context could reduce the compute burden

Generative AI has given carmakers a more flexible conversational interface, although running large models continuously inside a vehicle creates practical constraints around processor capacity, power use, connectivity, latency, and cost. Cloud inference can provide access to more capable models, but a car cannot assume that a low-latency connection will always be available, particularly where a function depends on an immediate understanding of driving conditions.

Smart Eye’s approach is to supply a more tightly defined package of contextual information before the model reasons about a request. A driver looking towards a building while asking “what is that?” creates a different task from a conventional voice query because gaze and exterior-camera information can help identify the subject without requiring a lengthy verbal description.

The same principle could apply to safety-related interactions. Information about attention, drowsiness, occupancy, identity, or activity in the cabin can change whether a vehicle should deliver information immediately, defer it, simplify an interaction, or issue a warning. The sensing layer consequently becomes part of the AI system rather than a separate product feeding isolated alerts into the dashboard.

Smart Eye already has an established automotive sensing business, giving the company a route into the software architectures where a more capable agent would eventually have to operate. That existing position is commercially relevant because an in-cabin AI system cannot be treated as a standalone application if it depends on access to driver-monitoring sensors and vehicle functions.

A demonstration at an industry event nevertheless remains some distance from a production vehicle programme. Automotive software has to meet development cycles measured in years, operate consistently across hardware variants and environmental conditions, and pass validation processes that are far more demanding than those applied to a consumer chatbot.

The car becomes an integration problem

Adding agentic software also raises questions about authority. An assistant that changes a playlist carries little risk, whereas one that adjusts vehicle settings, accesses personal information, interacts with navigation, or acts on assumptions about driver state requires clearer limits around what the software may do without confirmation.

The number of suppliers involved makes that harder. Modern vehicles already combine software from carmakers, Tier 1 suppliers, mapping providers, connectivity vendors, chip companies, cloud platforms, and specialist sensing businesses, while foundation-model providers introduce another dependency. A model-agnostic layer may give manufacturers flexibility, but every additional interface creates another integration and validation requirement.

Costs will also determine how much intelligence ultimately runs inside the cabin. High-end vehicles can absorb more compute and sensor hardware than mass-market models, while always-on cloud services create recurring operating costs that manufacturers either have to absorb, include in the vehicle price, or convert into subscriptions.

Smart Eye frames richer contextual input partly as a way of improving that equation by reducing unnecessary model processing. The claim will need to be demonstrated under production conditions, where network availability, hardware performance, language variation, unusual driver behaviour, and sensor failures are considerably less controlled than an exhibition environment.

Production will require automotive evidence

The company has demonstrated that several streams of vehicle and human information can be combined into an agent interface, but the commercial evidence will come from integration with a vehicle programme rather than the public demo itself. Carmakers will want to know how much compute is required, which workloads remain local, how quickly the system responds, and how its behaviour changes when individual sensors are unavailable.

Privacy will also become difficult to separate from functionality because the richer the contextual model becomes, the more information it potentially processes about the people inside the car. Identity, gaze, emotion, speech, location, and behavioural data can improve personalisation while also creating sensitive datasets that manufacturers and suppliers will have to secure and govern.

Regulation provides another constraint because driver-monitoring systems already sit inside automotive safety frameworks, whereas general AI assistants have developed under a much looser product model. Combining the two means that a feature marketed as an improved interface may depend on systems whose operation carries safety and compliance consequences.

Smart Eye’s demonstration therefore shows a plausible direction for automotive AI without establishing that the product is production-ready. The competitive question is moving from whether a car can host a conversational model towards whether software can understand enough about the driver and surrounding situation to take useful action without creating another distraction, another privacy risk, or an uneconomic layer of compute.