EU AI Act for Automotive: The Full Compliance Guide (2026)

An emergency-braking algorithm that decides whether to slam on the brakes before a collision is high-risk under the EU AI Act. A dealer chatbot that recommends a trim level is not. Why those two sit in different regulatory worlds is the point of this guide.

An automotive group does not run one AI system; it runs dozens — across the vehicle, the cabin, the assembly line, the dealership, and the back office, sharing neither one classification nor one deadline. The task is not to declare the company "high-risk" as a whole; it is to route each system correctly under Regulation (EU) 2024/1689.

---

How the EU AI Act actually applies to a car company

There are two roads to high-risk classification, and automotive mostly takes the first. Most sectors reach high-risk through Annex III — the stand-alone use-case list under Article 6(2) (credit scoring, recruitment, biometric identification). Automotive is different: its core systems reach high-risk through the product-embedded route under Article 6(1), because cars are already regulated products with their own EU conformity machinery. That changes not just the deadline but which AI Act provisions apply at all.

The four buckets

This guide routes the automotive AI estate into four buckets:

1. Driving and ADAS safety AI — the perception, planning, and intervention stack. Article 6(1) product-embedded route.
2. In-cabin attention, drowsiness, or emotion monitoring — can hit Annex III through the biometrics category, on a different clock.
3. HR and recruitment AI for factory, dealer, or office staff — Annex III point 4, regardless of the rest of the estate.
4. The large minimal-risk remainder — infotainment, navigation, marketing, demand forecasting, plant scheduling.

Bucket 4 is the largest and does not become high-risk just because it lives inside a car company; its only duties are the Article 50 transparency obligations (label a customer-facing chatbot; mark synthetic content) and the Article 5 prohibitions. A sibling guide covers the autonomous-vehicle AI obligation stack, and another the factory floor; here, the job is routing and timing.

---

Route 1: AI as a vehicle safety component (Article 6(1) + Annex I)

The central mechanism for automotive is Article 6(1). An AI system is high-risk when two conditions hold together: it is a safety component of a product covered by the Union harmonisation law in Annex I, and that product must undergo third-party conformity assessment before being placed on the market.

Annex I Section B is the legal hook — it lists motor-vehicle type-approval, Regulation (EU) 2018/858. Because cars, vans, and trucks already require EU type-approval, AI safety components inside them fall through Article 6(1), not Annex III.

The safety-component test

The test is functional, not labelled: would failure or malfunction of the AI system directly create a risk to the health and safety of persons? Emergency braking: yes. Lane-keeping: yes. The perception stack deciding whether the object ahead is a pedestrian or a plastic bag: yes. A headrest-adjustment or seat-memory routine: no — its failure annoys, it does not kill. The dividing line is direct risk of death or serious injury; borderline systems should be documented. The glossary entry on what counts as a safety component walks the test in detail.

ADAS, automated driving, and the General Safety Regulation

The Article 6(1) basis spans the whole autonomy spectrum: a Level 2 ADAS feature, a Level 4 automated-driving system, and a freight AV's perception stack all ride the same hook if they are safety components of a type-approved vehicle. Autonomy changes the engineering, not the classification basis. The General Safety Regulation (Regulation (EU) 2019/2144) sits underneath, mandating driver-assistance features on new vehicles and feeding the type-approval framework — but it is the type-approval requirement itself that triggers Article 6(1).

Why type-approval is the conformity gate (the Section B carve-out)

The critical point most automotive guidance gets wrong: vehicles do not run the AI Act's own high-risk machinery. Article 2(2) is the operative provision, worth reading verbatim: "For AI systems classified as high-risk AI systems in accordance with Article 6(1) related to products covered by the Union harmonisation legislation listed in Section B of Annex I, only Article 6(1), Articles 102 to 109 and Article 112 apply." Motor-vehicle type-approval — Regulation (EU) 2018/858 — sits in Annex I Section B. So the provisions binding the manufacturer directly are narrow: Article 6(1) (classification), Articles 102 to 109, and Article 112. Articles 8 to 15, the Article 16 provider obligations, and the Article 43 conformity assessment do not apply to vehicles as free-standing AI Act duties.

The substance does not disappear — it is written into the type-approval regime. Articles 102 to 109 amend the sectoral instruments (Regulation (EU) 2018/858 among them), and the detailed AI requirements are carried through by the delegated and implementing acts under that amended framework. So the manufacturer complies by satisfying type-approval, not by running Article 43 or a standalone Article 16 provider route. (Article 43(3) is the Section A mechanism and should not be applied to vehicles.) The detailed mechanics belong to the autonomous-vehicle guide.

---

Route 2: the Annex III pieces that still catch automotive

The non-obvious part: even though the driving AI is Route 1, some in-vehicle and back-office AI is genuinely Annex III stand-alone — a different clock, a different conformity process. Missing these is the most common misclassification.

Driver monitoring: biometrics vs emotion recognition

In-cabin driver monitoring is the trap. A module that tracks gaze direction, eye closure, or head pose to detect distraction or drowsiness may or may not implicate the Annex III biometrics category, depending on what it infers. The analysis turns on whether the system processes biometric data to categorise the person or infer emotional state — which must be documented, not assumed.

Be precise about the prohibition line. Article 5(1)(f) prohibits emotion-recognition AI only in workplace and education settings — not in a consumer driver-safety context. A driver-attention system that infers fatigue or emotional state in a privately owned car is not caught by the Article 5 prohibition; the question is whether it is high-risk under Annex III, a classification question, not a ban. Do not overclaim that in-cabin monitoring is prohibited — it generally is not, though it may be high-risk, and the reasoning either way must be on file.

Recruitment and workforce AI (Annex III point 4)

The back office catches automakers too. AI used to recruit or select factory, dealer, or office staff falls under Annex III point 4(a) (recruitment and selection); AI used for in-employment decisions — promotion, task allocation, performance evaluation, termination — falls under Annex III point 4(b). Either triggers the full high-risk stack regardless of everything else in the estate. This is route-independent: the HR AI does not borrow the type-approval path; it runs its own Article 43 assessment on the Annex III clock.

Where automakers are NOT the regulated party

Annex III point 2 covers AI used as a safety component in the management of critical infrastructure, including road traffic — adaptive signalling, motorway flow control. But the regulated party is the road operator or public authority, not the carmaker: a vehicle that receives a traffic signal does not provide the traffic-management system.

---

OEM vs Tier-1 supplier: who is the provider?

The most consequential allocation is the provider role, because the provider carries the obligation weight — directly under Article 16 for Route 2 systems, and through the manufacturer's type-approval duties for Route 1. Get it wrong and you discover it at a recall.

Article 3 defines the roles: a provider develops an AI system and places it on the market under its own name or trademark; a deployer uses one under its authority. The provider-versus-deployer guide covers the full split.

The default OEM/Tier-1 allocation

The OEM that places the finished vehicle on the market under its own brand is typically the provider of the AI it develops in-house. A Tier-1 supplier that develops and supplies a named AI system under its own brand — say, an AEB perception stack — is the provider of that system, with the OEM as its deployer.

So a single vehicle can carry AI from several providers: the supplier-branded AEB stack (Tier-1), an OEM-developed parking system (OEM), a third-party infotainment assistant (its vendor, but minimal-risk anyway). Each provider owns the obligation stack for its own system — discharged through type-approval for the safety components, and as direct Article 16 duties for any Route 2 system.

Article 25 role-shift on substantial modification

Article 25 is the trap that catches integration teams. A distributor, importer, or deployer becomes the provider — inheriting the full provider obligations — if it places the system on the market under its own name, makes a substantial modification, or changes its intended purpose.

This bites when an OEM fuses a supplier's perception module into its own decision logic, retrains it, or repurposes it — the OEM may then have become the provider of the modified system, even though a supplier built the original. There is also a dual-hat case: an OEM running its own robotaxi fleet is both the provider of the driving AI and its deployer. The takeaway is contractual — the provider/deployer boundary, modification rights, and documentation hand-over should be settled in the supply contract before integration, not at a recall.

---

The Articles 9-15 requirements — and which mechanism delivers them per route

For any automotive AI that lands in high-risk, the substantive requirements are the same Articles 9 to 15 set: risk management, data governance, robustness, human oversight, logging. But how they bind depends on the route, and conflating the two trips up automotive teams.

• Route 1 (vehicle safety-component AI, Annex I Section B). Under Article 2(2) the only AI Act provisions that apply directly are Article 6(1) plus Articles 102 to 109 and Article 112. The Article 8-15-equivalent substance reaches the manufacturer through the AI-Act-amended type-approval framework — carried by the delegated and implementing acts under the amended Regulation (EU) 2018/858 — not as standalone obligations, and with no separate Article 16 or Article 43 step for the vehicle.
• Route 2 (stand-alone Annex III systems — in-cabin emotion/biometric systems, recruitment and workforce AI). Not vehicle safety components, so the carve-out does not touch them. The full high-risk stack applies directly: Articles 9 to 15, the Article 16 provider obligations, the Article 43 conformity assessment, and Article 49 registration — on the Annex III clock.

What follows describes what those requirements are — type-approval-embedded substance on Route 1, direct AI Act obligations on Route 2.

The Articles 9-15 requirements at a glance

• Article 9 — risk management. A continuous, lifecycle system, not a one-off, addressing AI-specific failure modes: perception failure in degraded weather, distribution shift between validation and the deployed fleet, adversarial fragility.
• Article 10 — data governance. Training, validation, and test data must be relevant and representative. For perception AI that means edge-case coverage — night driving, unusual road markings, rare obstacle classes across the EU markets sold.
• Article 11 and Annex IV — technical documentation. The technical file must cover the nine Annex IV content areas, from system description to validation results to the risk-management report.
• Article 12 — logging. Automatic recording of events over the system's lifetime, for traceability and post-incident analysis.
• Article 13 — transparency and instructions for use. The deployer (OEM or fleet operator) must receive instructions sufficient to operate and oversee it.
• Article 14 — human oversight. The system must be designed so a human can oversee it — driver-in-the-loop for Level 2/3, remote operator or safety design for higher automation, with transition-demand logic.
• Article 15 — accuracy, robustness, and cybersecurity. Declared accuracy levels, resilience to faults and inconsistent inputs, and protection against manipulation.

Where the AI Act layers on UNECE and type-approval

The overlaps matter, but the requirements look to the AI system's own properties. UNECE or type-approval compliance is strong supporting evidence — not a substitute. A type-approved vehicle is not automatically AI Act compliant.

The lifecycle duties round out the stack: technical documentation retained for ten years (Article 18); the EU Declaration of Conformity (Article 47); a post-market monitoring system (Article 72); and serious incidents reported within the Article 73 timelines. On Route 2 these bind directly; on Route 1 their equivalents are delivered through the amended type-approval regime rather than as standalone AI Act articles.

The overlap table

In every row the automotive instrument is supporting evidence and the AI Act requirement is the obligation; one does not discharge the other:

---

Worked example: routing a Tier-1 supplier's AI portfolio

Take Lumida Drive Systems, a fictional 900-person German Tier-1 supplier. It ships an AEB and lane-keeping perception stack to two OEMs, plus an in-cabin attention-monitoring module and an internal recruitment-screening tool. Four systems, four routes.

System 1 — the AEB/LKA perception stack. A safety component of a type-approved vehicle: Article 6(1) plus Annex I Section B. Under the Article 2(2) carve-out, the provisions binding Lumida directly are only Article 6(1) plus Articles 102 to 109 and Article 112 — no standalone Article 16 route and no separate Article 43 assessment. The Article 8-15-equivalent substance reaches Lumida and the OEMs through the AI-Act-amended type-approval framework, satisfied inside type-approval rather than as free-standing AI Act duties. The deadline is 2 August 2028 under the adopted Digital Omnibus deferral (see below), so Lumida plans against that date.

System 2 — the in-cabin attention-monitoring module. Case-by-case analysis against the Annex III biometric category. If it merely detects gaze direction and eye-closure without inferring emotional state, it may stay outside Annex III — but that must be documented, not assumed. If it infers emotional state, the biometric Annex III obligations apply on the stand-alone clock: statutory 2 August 2026, deferred to 2 December 2027 under the adopted Digital Omnibus. It is not prohibited under Article 5(1)(f) because a private car is not a workplace.

System 3 — the recruitment-screening tool. Annex III point 4(a). If Lumida built it in-house, it is the provider; if it licenses a vendor tool, it is the deployer and the vendor the provider. Either way the full high-risk stack applies on the Annex III clock.

System 4 — demand-forecasting and plant-scheduling AI. Minimal-risk. No mandatory high-risk stack, only the baseline Article 5 and Article 50 watch — where most of Lumida's internal AI sits.

One caution on relief: at roughly 900 staff Lumida exceeds the SME threshold, so it cannot rely on the Article 99(6) lower-cap that shields genuine SMEs and start-ups.

---

Deadlines, penalties, and what to do before each clock

The automotive estate runs on several clocks at once; routing tells you which applies.

The automotive compliance calendar

The clocks already running:

• Article 5 prohibitions — in force since 2 February 2025; two new prohibitions — on AI generating non-consensual intimate imagery of an identifiable person and on AI generating child sexual abuse material — apply from 2 December 2026. Article 50, including content-marking, has applied since 2 August 2026.
• Article 4 AI literacy — in force since 2 February 2025.
• GPAI obligations (Articles 51-55) — since 2 August 2025.

The high-risk clocks:

• Stand-alone Annex III (Article 6(2)) — statute originally read 2 August 2026; deferred to 2 December 2027 under the adopted Digital Omnibus. The clock for the recruitment tool and any in-cabin module that infers emotion.
• Product-embedded Annex I (Article 6(1)) — statute originally read 2 August 2027; deferred to 2 August 2028 under the same adopted Digital Omnibus. The clock for the driving and ADAS safety AI.

The Digital Omnibus is now adopted — the European Parliament passed it on 16 June 2026 and the Council on 29 June 2026 — so the deferral to 2 December 2027 for stand-alone Annex III is settled; it was published in the Official Journal on 24 July 2026 and in force since 27 July 2026. It is a set of fixed calendar dates; the Commission's November 2025 "stop the clock" proposal would have linked the dates to standards availability, with backstop dates.

Penalty exposure under Article 99

The tiers are exact:

• Article 99(3) — prohibited practices: up to €35 million or 7% of total worldwide annual turnover, whichever is higher.
• Article 99(4) — high-risk and most obligation breaches: up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
• Article 99(5) — supplying incorrect, incomplete, or misleading information to authorities: up to €7.5 million or 1% of total worldwide annual turnover, whichever is higher.
• Article 99(6) — for genuine SMEs and start-ups, the fine is capped at the lower of the percentage or the fixed amount.

Exposure does not stop at AI Act fines. Market-surveillance action under Regulation (EU) 2018/858 — type-approval withdrawal, recall, suspension of registration — can run in parallel for the same defect, and for an automaker the recall risk is often the more immediate commercial threat.

How Confir routes an automotive AI estate

The hard part of automotive compliance is routing dozens of systems before the work starts. Confir's classification engine is deterministic and rule-based — no model inference. It runs each through the same questions: safety component of a type-approved vehicle (Article 6(1) plus Annex I)? Annex III through biometrics or employment? Minimal-risk with only Article 50 duties? Every conclusion cites the article it rests on.

Once a system is routed to high-risk, Confir structures the Annex IV technical file and the Article 47 Declaration of Conformity for the integrated type-approval assessment or stand-alone Annex III, and surfaces where UNECE R155, R156, and R157 evidence already covers part of the requirement. The GPAI provider workflow under Articles 51-55 remains partial, on the roadmap.

---

Frequently asked questions

Is automotive AI classified under Annex III of the EU AI Act?

Mostly no. AI that is a safety component of a motor vehicle requiring EU type-approval is high-risk under Article 6(1) combined with Annex I (Section B lists Regulation (EU) 2018/858), not under Annex III. Annex III still catches specific cases: in-cabin biometric monitoring can fall under the biometrics category, and recruitment or workforce AI under Annex III point 4. So a single automaker can run systems on both routes at once, each with its own clock and conformity path.

What is the EU AI Act deadline for vehicle safety AI?

The Digital Omnibus date for Annex I product-embedded high-risk AI, including vehicle safety AI, is 2 August 2028. That deferral is now adopted — the European Parliament passed it on 16 June 2026 and the Council on 29 June 2026 — so teams should plan against the adopted 2 August 2028 date; it was published in the Official Journal on 24 July 2026 and in force since 27 July 2026.

Is the OEM or the Tier-1 supplier the provider under the EU AI Act?

It depends on who places the AI system on the market under their own name. If a Tier-1 supplier develops and brands an AI system such as an emergency-braking stack, the supplier is the provider and the OEM is its deployer. If the OEM builds the AI in-house, the OEM is the provider. Under Article 25, an OEM that substantially modifies a supplier's system or changes its intended purpose can become the provider for the modified system, so the boundary should be settled contractually before integration.

Is driver attention or emotion monitoring high-risk?

It needs case-by-case analysis. A module that detects eye closure or gaze direction without inferring emotional state may sit outside Annex III, but that conclusion must be documented. If it infers a driver's emotional state, the biometric/emotion-recognition rules become relevant. Emotion recognition is prohibited under Article 5(1)(f) only in workplace and education settings — not in a consumer driver-safety context — so the in-vehicle case is generally an Annex III classification question, not an outright prohibition.

Does type-approval compliance satisfy the EU AI Act?

For vehicle safety AI, type-approval is the route through which compliance is delivered. Under Article 2(2), because motor-vehicle type-approval (Regulation (EU) 2018/858) sits in Annex I Section B, only Article 6(1) plus Articles 102 to 109 and Article 112 apply directly — no separate AI Act conformity assessment and no standalone Article 16 route for the vehicle. The AI requirements are written into the amended type-approval framework, so the manufacturer satisfies them inside type-approval. It is not a free pass: UNECE compliance such as R155 is supporting evidence, not a discharge. Stand-alone Annex III systems (in-cabin emotion AI, recruitment tools) are different — they run the full Article 43 assessment directly.

Which automotive AI is not high-risk under the EU AI Act?

Most of it. Infotainment recommendations, navigation, manufacturing demand forecasting, plant scheduling, and dealer marketing or adtech are generally minimal-risk with no mandatory compliance stack — their only duties are Article 50 transparency and avoiding the Article 5 prohibited practices. The high-risk stack applies only to safety-driving AI, certain in-cabin biometrics, and workforce-decision AI.

What are the penalties for non-compliance in automotive?

Prohibited practices carry up to €35 million or 7% of worldwide annual turnover (Article 99(3)). High-risk breaches, including failures across Articles 9-15 and the provider duties in Article 16, carry up to €15 million or 3% (Article 99(4)). Supplying incorrect or misleading information to authorities carries up to €7.5 million or 1% (Article 99(5)). For genuine SMEs and start-ups, Article 99(6) caps the fine at the lower figure. Market-surveillance action under Regulation (EU) 2018/858, including recall and type-approval withdrawal, can run in parallel.

---

Related guides

• Article 6 high-risk classification rules
• what counts as a safety component
• the autonomous-vehicle AI obligation stack
• provider vs deployer duties
• manufacturing and machinery AI compliance