What problem does it solve?
Modern cars have many functions behind touchscreens and menus, and drivers should keep their eyes on the road. First generation voice control only understood fixed commands in a fixed order: say it slightly differently and it failed, which leaves drivers with menus, the manual or a question to the dealer.
Large language models change what the assistant can understand, but a car is not a phone. The assistant has to work with patchy connectivity, answer in a split second, never distract, keep vehicle and location data private, and separate harmless requests (a warmer seat, a restaurant nearby) from anything that touches driving. Car makers also want to own the experience and the brand voice rather than hand the cabin to a third party assistant.
How does it work?
- Wake and listen. The driver says the wake word or presses the steering wheel button; speech recognition runs on board for commands and in the cloud for open questions.
- Decide who answers. Vehicle commands (climate, seats, media, navigation) stay in the car maker's own system. Only questions it cannot answer go to a language model, anonymized and without vehicle data, as Volkswagen describes for IDA.
- Ground the answer. Knowledge questions use a web search or a maps platform, and questions about the car use the owner's manual and vehicle status, so answers are current and specific.
- Keep the conversation. The assistant remembers the dialogue for a limited time, so follow up questions work without repeating the context.
- Answer in the brand voice and act. The reply is spoken in the car maker's voice and, where allowed, the assistant sets the function or starts navigation, then confirms what it did.
- Learn safely. Voice data is stored anonymized, answers are screened for harmful content, and usage patterns can suggest routines the driver can accept or ignore.
- Audience
- Customer facing
- Autonomy
- Autonomous
- Adoption
- Early adopters
- Channels
- Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Customer experience, Revenue growth, Inclusion and access, Lower cost to serve.
Indicative value
A car maker with 1 million connected vehicles on the road
USD 50,000 to USD 900,000
Customer service cost avoided on feature questions per year
How this is calculated
Formula: vehicles * featureContacts * deflection * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Connected vehicles with the assistant vehicles, vehicles | 1,000,000 | 1,000,000 | The reference car maker. |
| Contacts to brand customer service or the dealer about how a vehicle function works featureContacts, contacts per vehicle per year | 0.1 | 0.3 | Editorial assumption, replace with your own contact reason data. |
| Share of those contacts the assistant makes unnecessary deflection, fraction of feature contacts | 0.1 | 0.3 | Editorial assumption. None of the car makers on this page discloses usage or deflection figures. |
| Cost of a handled contact costPerContact, USD per contact | 5 | 10 | Editorial assumption for a blended phone and chat contact, replace with your own cost. |
What it leaves out: Covers service cost only. It leaves out the main reasons car makers build this (product differentiation, connected services revenue and brand loyalty), the cost of cloud models and speech processing per vehicle, and the integration and validation work in the vehicle.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
BMW Group
Germany · Automotive · 2026
BMW rebuilt its Intelligent Personal Assistant on Amazon's Alexa Custom Assistant with Alexa+ large language model technology, so drivers can hold natural conversations, ask several questions in one sentence about vehicle functions or general knowledge, and have the assistant recognize context. BMW introduced the German language version in the new BMW iX3 from mid April 2026 production; earlier iX3 vehicles were due to receive it by software update from the end of May 2026. BMW plans other markets and further models on BMW Operating System 9 and X from the second half of 2026. The assistant suggests routines from daily usage patterns, and from July 2026 BMW Operating System X adds new options for creating routines.
No outcome disclosed.
Mercedes-Benz Group
Germany · Automotive · 2024
After a US beta of ChatGPT in 2023, Mercedes-Benz brought a general knowledge function to series production vehicles in December 2024: the MBUX Voice Assistant runs a Microsoft Bing search and answers in natural language with ChatGPT through Azure OpenAI Service, keeps the dialogue for up to one hour for follow up questions, stores voice data anonymized in its own cloud and uses a risk assessment tool to reduce harmful answers. Mercedes-Benz announced it as a free update for over three million vehicles in German and English. In January 2025 it announced Gemini based conversational search for points of interest with Google Cloud's Automotive AI Agent, starting in the new CLA.
No outcome disclosed.
Volkswagen
Germany · Automotive · 2024
Volkswagen added ChatGPT to its IDA voice assistant through Cerence Chat Pro. IDA keeps handling vehicle functions such as infotainment, navigation and climate control; only questions its own system cannot answer are forwarded anonymously to ChatGPT, and the answer is read out in the familiar Volkswagen voice. ChatGPT gets no access to vehicle data, questions and answers are deleted immediately, and drivers can switch the online assistant off. It launched in 2024 in all new ID. models, the new Golf, Tiguan and Passat, in five languages, and Cerence reports that it reached cars already on the road by cloud update across Volkswagen, Cupra, Seat and Skoda.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A catalog of vehicle commands and the functions the assistant may set, per model and market
- Owner's manuals and feature descriptions in every supported language
- Vehicle status signals the assistant may read, with a privacy classification per signal
- Anonymized samples of real utterances to test recognition and answers
Systems to integrate
- Head unit and vehicle operating system for commands and status
- Connected car cloud and over the air update pipeline
- Speech recognition and synthesis with a brand voice
- Language model and web search or maps platform for open questions
- Companion app and customer account for consent and settings
Complexity: High
Hybrid on board and cloud processing, low latency speech in a noisy cabin, many languages and accents, strict separation from driving functions, over the air updates and privacy by design across millions of vehicles make this a platform program rather than a feature.
- 1
Draw the line between commands and conversation
List which requests the vehicle system keeps, which go to the language model and which are refused, and never let the model reach functions that affect driving.
- 2
Design privacy in from the start
Decide what leaves the car, how it is anonymized, how long dialogue memory lasts, how the driver switches the online assistant off and how consent is recorded in the account.
- 3
Ground answers about the car
Load owner's manuals and feature descriptions per model and market into retrieval, so the assistant explains the driver's own car rather than a generic one.
- 4
Test in the cabin, not the lab
Test with road noise, accents, passengers talking and weak connectivity, and measure recognition, latency and task success per language before launch.
- 5
Launch per language and market, then update over the air
Start with a few languages in new vehicles, extend to cars on the road by software update and add functions such as routines only after they pass the same tests.
Guardrails
- No access for the language model to driving, safety or security functions
- Only anonymized text leaves the vehicle, without vehicle identifiers or location unless the driver asked for a place
- Content screening of answers for harmful, illegal or distracting content
- A clear way to switch the online assistant off, and short retention of dialogue history
- Spoken answers kept short to limit driver distraction
KPIs to instrument
- Share of active vehicles that use the assistant each month
- Task success rate per request type and language
- Latency from end of speech to start of the answer
- Share of answers flagged as wrong or unsafe in reviews
- Customer service contacts about vehicle functions per 1,000 vehicles
Human in the loop
There is no human in the conversation, so the human control sits in design and operations: product and safety teams approve every new function the assistant may set, reviewers sample anonymized dialogues for wrong or unsafe answers, and customer service handles complaints and feedback about the assistant.
Common failure modes
- Confident answers about the wrong car
- The model explains a feature the driver's model or trim does not have. Ground answers in the manual for that vehicle and say when a feature is not available.
- Latency that drivers will not accept
- A cloud round trip that takes several seconds makes people give up. Keep commands on board and stream the answer.
- Privacy surprises
- Drivers discover that voice or location data left the car without them knowing. Make data flows, retention and the off switch visible.
- Distraction by design
- Long, chatty answers pull attention from the road. Limit answer length and avoid anything that invites the driver to look at a screen.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(1): people must be informed that they are interacting with an AI system unless that is obvious from the context. Article 50(2): synthetic audio output must be marked as artificially generated. A cabin assistant for comfort, media, navigation and knowledge questions is not an Annex III use. It would move towards the high risk regime if it became a safety component of the vehicle: vehicle type approval legislation is listed in Annex I Section B, and under Article 2(2) the high risk requirements reach those products only through the amendments the AI Act makes to that legislation. Keep driving and safety functions out of its reach.
Rules that apply
Guidance
- Guidelines 01/2020 on processing personal data in the context of connected vehicles and mobility related applications (European Data Protection Board, Europe). Sets out how GDPR applies to data processed in and sent from connected vehicles, including consent, minimization and local processing.
- Visual manual NHTSA driver distraction guidelines for in vehicle electronic devices (US National Highway Traffic Safety Administration, North America). US guidelines to limit the distraction caused by in vehicle devices, relevant to how much the assistant shows on screen and asks of the driver by hand. Auditory vocal interaction itself is outside their scope.
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Requires disclosure that people are interacting with AI and marking of synthetic audio.
Controls to put in place
- Documented allow list of vehicle functions the assistant may control, approved by product safety
- Data protection impact assessment covering voice, location and vehicle data
- Anonymization and retention rules enforced in the cloud pipeline
- Regression test suites per language and model before every over the air release
- Monitoring of harmful or wrong answers with a fast rollback path
Frequently asked questions
- Which car makers use a generative AI voice assistant?
- Volkswagen added ChatGPT to its IDA voice assistant in 2024 through Cerence, for new ID. models, the Golf, Tiguan and Passat. Mercedes-Benz made a ChatGPT and Bing based knowledge feature in the MBUX Voice Assistant available to over three million vehicles in December 2024, and BMW introduced the German language version of its Alexa+ based Intelligent Personal Assistant in the iX3, from mid April 2026 production.
- How do car makers keep voice data private?
- Volkswagen forwards only questions its own system cannot answer, anonymously, gives ChatGPT no vehicle data and deletes questions and answers immediately. Mercedes-Benz stores voice data anonymized in its own cloud. The EDPB guidelines on connected vehicles set the GDPR baseline.
- Is an in car AI assistant high risk under the EU AI Act?
- Usually not: it carries the Article 50 transparency duties. It would move towards the high risk regime if it became a safety component of the vehicle, so keep it away from driving functions.
How to cite this page
Blits.ai AI Use Case Library, "Generative AI voice assistant in the car", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/in-car-ai-voice-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published