AI liability and operator duties in Germany: 2026 guide.

Germany has designated the Bundesnetzagentur (BNetzA) as its market surveillance authority under Regulation (EU) 2024/1689 and adopted the KI-MIG draft in February 2026 to formalise enforcement powers. An operator deploying an AI agent in Germany sits at the intersection of EU AI Act Article 26 deployer duties, German product liability law under the Produkthaftungsgesetz, the revised Product Liability Directive transposition due by 9 December 2026, and one of Europe's most active GDPR enforcement landscapes. Article 50 chatbot transparency obligations apply from 2 August 2026. This guide maps each layer.

Germany's national competent authority: BNetzA and the KI-MIG

Article 70 of Regulation (EU) 2024/1689 required Member States to designate their national competent authorities by 2 August 2025. Germany designated the Bundesnetzagentur as its market surveillance authority for the EU AI Act and DAkkS as its notifying authority within that deadline. This designation placed Germany among the minority of EU Member States to have formally designated authorities by the required date.

The Bundesnetzagentur is an established federal regulatory agency with existing market surveillance responsibilities in telecommunications, energy, postal services, and railway infrastructure. Its designation for AI Act market surveillance extends its mandate into a new technical domain. BNetzA coordinates with BaFin (Bundesanstalt fuer Finanzdienstleistungsaufsicht) for AI systems used in financial services, with the Bundesnetzagentur's media and communication divisions for AI systems in broadcasting, and with the BfArM (Bundesinstitut fuer Arzneimittel und Medizinprodukte) for AI as medical device. This coordination architecture reflects the EU AI Act's sector-specific supervisory design, in which the general market surveillance authority defers to product-safety-specific bodies for AI embedded in regulated product categories.

DAkkS, Germany's national accreditation body operating under Regulation (EC) 765/2008, takes the role of notifying authority. It assesses and accredits the conformity assessment bodies that audit high-risk AI systems before market placement. For operators seeking third-party conformity assessments of high-risk AI systems in Germany, the conformity assessment body they engage must hold DAkkS accreditation or equivalent recognised accreditation under the European accreditation framework (EA).

To provide BNetzA with clear procedural authority, the federal cabinet adopted the KI-Marktaufsichts- und Implementierungsgesetz (KI-MIG) in February 2026. The KI-MIG does not create substantive AI obligations beyond those in the directly applicable EU AI Act; instead, it establishes the administrative infrastructure: BNetzA's investigative powers, its ability to order market surveillance measures and market withdrawal, the procedural rules for complaints filed by operators and affected parties, and the coordination mechanisms between BNetzA and sectoral supervisors. The law also addresses penalties, building on the Article 99 penalty regime in the EU AI Act, which sets maximum fines of EUR 35 million or 7 per cent of global turnover for prohibited-practice breaches, EUR 15 million or 3 per cent for other breaches, and EUR 7.5 million or 1 per cent for incorrect information. [VERIFY the exact KI-MIG parliamentary status as of June 2026 and whether any amendments were made during legislative passage.]

EU AI Act deployer obligations in Germany

The EU AI Act distinguishes between providers and deployers. A provider places a system on the market; a deployer uses it under its own authority in a professional context. Most German businesses deploying AI agents are deployers, not providers. They carry the obligations set out in Article 26 rather than the full Article 16 provider obligations.

Article 26 requires deployers to use a high-risk AI system in accordance with the provider's instructions, to implement appropriate human oversight measures, to monitor the system for performance against its intended purpose, to suspend or disable it if there is risk of serious incident, to inform the provider of any serious incident within the reporting timelines, and to retain Article 12 logs where technically possible. Deployers in Germany who are subject entities under the EU's AI liability exposure must also assess whether their specific use case triggers the Article 27 Fundamental Rights Impact Assessment requirement. Article 27 applies to deployers that are bodies governed by public law or operators of critical infrastructure.

A deployer that substantially modifies a high-risk AI system, or deploys it for a purpose outside the use case the provider documented, acquires provider-level obligations under Article 25. This is a significant risk for German businesses customising large language model-based agents for specific workflows. If the customisation amounts to a new system with different risk characteristics, the deployer becomes the provider for regulatory purposes.

The current legally binding compliance date for high-risk Annex III standalone AI systems is 2 August 2026. On 7 May 2026, the Council, Parliament, and Commission reached a provisional political agreement under the Digital Omnibus on AI package that would defer this to 2 December 2027. That agreement has not been published in the Official Journal of the EU as of June 2026 and does not have legal force. German operators should plan for 2 August 2026 as the operative date and treat the deferral as conditional on formal adoption.

Article 50 transparency: the obligations that are not deferred

Article 50 of Regulation (EU) 2024/1689 imposes transparency obligations that apply from 2 August 2026 and are not subject to the Digital Omnibus deferral. For operators in Germany, these are the most immediately pressing EU AI Act obligations.

Article 50(1) requires deployers of AI chatbots to inform individuals, in a clear and timely manner before the interaction begins, that they are interacting with an AI system. This applies to any consumer-facing or business-to-business AI agent that engages in natural language dialogue. A disclaimer buried in terms of service is insufficient; the disclosure must be visible at the start of the interaction.

Article 50(2) requires providers and deployers to ensure that AI-generated or manipulated audio, image, video, or text content is marked in a machine-readable format. A grace period until 2 December 2026 is proposed for this machine-readable marking obligation for systems already on the market, but this is a proposal under the Digital Omnibus and is not yet adopted. The core disclosure duty of Article 50(1) has no proposed grace period.

Article 50(3) requires deployers of emotion-recognition systems to inform individuals when such systems are operating. Article 50(4) requires that AI-generated deepfakes of real individuals are disclosed as artificially generated or manipulated. The European Commission published Article 50 implementation guidance and a Code of Practice in June 2026, providing detailed compliance templates for each of these obligations.

In Germany, BNetzA enforces Article 50. The penalties for Article 50 breaches fall under the general AI Act penalty structure: up to EUR 15 million or 3 per cent of global annual turnover for breaches of operator obligations other than the Article 5 prohibited practices. For a business with significant global revenue, this is a material fine exposure from a relatively straightforward technical disclosure requirement.

German product liability law and the revised PLD: parallel civil exposure

Independent of the EU AI Act, operators in Germany face civil liability exposure under German product liability law when an AI agent causes harm.

Current position: Produkthaftungsgesetz

Germany implemented the original EU Product Liability Directive (85/374/EEC) in the Produkthaftungsgesetz (ProdHaftG) of 1 January 1990, most recently amended in 2002. The ProdHaftG imposes strict liability on the producer of a defective product for personal injury and property damage caused by the defect, without any requirement to prove negligence. A product is defective when it does not provide the safety a person is entitled to expect.

The principal challenge under the ProdHaftG for AI liability has been whether software qualifies as a product at all. German courts and academic commentary have historically treated software embedded in hardware as part of the product, but treated standalone software with more caution. The emerging consensus among German legal academics and the view of the German Federal Government in its KI-MIG legislative materials is that standalone AI software constitutes a product within the meaning of the ProdHaftG when supplied commercially. Under this reading, a commercially deployed AI agent that causes harm by providing defective outputs could in principle expose its developer (the provider) to ProdHaftG strict liability.

For deployers as opposed to providers, the ProdHaftG strict liability falls primarily on the producer, defined as the manufacturer or importer. A deployer is not typically the manufacturer of the AI system. However, a deployer who modifies the AI system, rebrands it as their own product, or presents themselves to customers as its source may acquire manufacturer status under the ProdHaftG. This mirrors the EU AI Act's provider-by-modification rule in Article 25 and creates a coordinated commercial risk: a deployer who customises an AI agent and presents it to customers under their own brand may simultaneously become the EU AI Act provider and the ProdHaftG manufacturer, with both regulatory and civil liability consequences.

Beyond the ProdHaftG, deployers in Germany may also face liability under general tort law (section 823 BGB, the Buergerliches Gesetzbuch) for harm caused by negligent deployment of an AI agent. Section 823(1) BGB provides a claim for breach of a protected right or interest. Section 823(2) BGB creates liability for breach of a statutory duty, which includes the EU AI Act's deployer obligations once they are in force. This means that a deployer who fails to implement the Article 26 human oversight requirement, resulting in harm to a user, may face a concurrent BGB section 823(2) claim grounded in breach of the EU AI Act obligation.

The revised Product Liability Directive: what changes from 9 December 2026

Directive (EU) 2024/2853, the revised Product Liability Directive, was published in the Official Journal of the EU on 18 November 2024 and entered into force on 8 December 2024. Member States, including Germany, must transpose it into national law by 9 December 2026. It applies to products placed on the market or put into service after 9 December 2026. The existing ProdHaftG and the original Directive 85/374/EEC continue to apply to products placed on the market before that date.

For AI operators, the revised PLD introduces three changes of significance. First, it expressly includes software and AI systems within the definition of product. This removes the doctrinal ambiguity about standalone AI software and places AI agents firmly within the scope of product liability law from 9 December 2026.

Second, it introduces rebuttable presumptions that ease the claimant's burden of proof in complex cases. Under Article 9 of the revised PLD, a product is presumed defective where the claimant faces excessive difficulties in proving defectiveness due to technical or scientific complexity. This is directly relevant to AI agent cases: where the opacity of an AI model's decision-making makes it practically impossible for a claimant to demonstrate the defect, the court may presume defectiveness and place the burden on the defendant to rebut it. For operators of opaque AI systems, this shifts the litigation risk materially.

Third, the revised PLD expands compensable damage to include medically recognised psychological harm and, critically, data loss or corruption. This last category is directly relevant to AI agent deployments: an AI agent that corrupts, deletes, or exfiltrates data belonging to a third party may trigger a revised PLD claim from that third party against the provider.

Germany must transpose the revised PLD by replacing or amending the ProdHaftG by 9 December 2026. The legislative process is underway but completion before that date is required. German operators should monitor the transposition bill and assess how the rebuttable presumption mechanism is implemented in the German text.

GDPR and automated decision-making in Germany

Germany's GDPR enforcement landscape is distinctive in European terms. Unlike most EU Member States, which have a single national DPA, Germany has 16 Laender-level (state) data protection authorities with independent enforcement powers for private sector entities within their geographic jurisdiction, plus the federal BfDI (Bundesbeauftragte fuer den Datenschutz und die Informationsfreiheit) for federal public bodies and certain cross-border matters. The Datenschutzkonferenz (DSK) coordinates interpretive positions but does not remove independent enforcement authority.

Article 22 GDPR prohibits solely automated decisions that produce legal or similarly significant effects on individuals, unless the individual has given explicit consent, the decision is necessary for a contract, or it is permitted by law with appropriate safeguards. An AI agent making consequential autonomous decisions about individuals in Germany without a lawful basis for Article 22 processing is exposed to enforcement action by the relevant Laender DPA, independently of any EU AI Act enforcement by BNetzA.

The Hamburg DPA (HmbBfDI) and the Baden-Wuerttemberg Landesbeauftragter fuer Datenschutz und Informationsfreiheit (LfDI) have been among the most active European DPAs on AI and data-processing enforcement. Their prior enforcement actions against AI-related data processing under GDPR provide signal for how German DPAs approach novel AI deployment questions. Operators should treat GDPR compliance for automated decision-making as a separate compliance track running in parallel with EU AI Act compliance, not a subset of it.

The BfDI has issued AI-specific guidance applying GDPR principles to AI systems, covering requirements for transparency notices, data minimisation in training data, rights of access and erasure for AI-processed data, and the section 22 GDPR Bundesdatenschutzgesetz rules on processing sensitive categories. The interaction between GDPR's Article 22 right to human review and the EU AI Act's Article 14 human oversight requirement creates complementary obligations that must be satisfied together.

BaFin's expectations for AI in financial services

For German financial services operators, BaFin (Bundesanstalt fuer Finanzdienstleistungsaufsicht) supervises AI use within the regulated financial sector independently of BNetzA's general AI Act market surveillance. BaFin published supervisory guidance on algorithmic trading and automated credit decisions in prior years and has updated its expectations in the context of the EU AI Act and DORA (Regulation (EU) 2022/2554, the Digital Operational Resilience Act, applying from 17 January 2025).

Under DORA, which applies directly to regulated financial entities in Germany, ICT-related incidents including AI agent failures in operational systems must be reported to BaFin within defined timelines. A serious AI agent malfunction that disrupts a payment service or a trading system triggers both DORA incident reporting obligations (to BaFin) and EU AI Act serious incident reporting obligations (to BNetzA under Article 73). These are parallel obligations to separate authorities on potentially overlapping timelines. German financial operators should map both reporting chains before deploying AI agents in operational financial systems.

EIOPA's Opinion on Artificial Intelligence Governance and Risk Management (EIOPA-BoS-25-360, published 6 August 2025) addressed to national competent authorities for insurance provides the supervisory baseline for AI used in German insurance underwriting, pricing, and claims. The BaFin, acting as Germany's insurance supervisor under Solvency II, is expected to implement EIOPA's supervisory expectations. German insurers using AI for pricing decisions face both BaFin supervisory scrutiny and EU AI Act deployer obligations if their pricing AI falls within an Annex III category.

Insurance and liability implications for German AI operators

German businesses deploying AI agents face a liability exposure that spans three legal frameworks simultaneously: the EU AI Act (regulatory penalties and deployer duties), German product liability and tort law (civil claims from harmed third parties), and GDPR (fines and compensation claims from data subjects). Standard German commercial insurance lines were designed before autonomous AI agent deployment was common and may not address these exposures adequately.

Coverage gaps to verify

A Betriebshaftpflichtversicherung (commercial general liability policy) covers damage caused to third parties in the course of business operations. The question for AI deployments is whether the policy covers harm caused by an autonomous decision made by an AI agent acting under the business's instructions. Older policy wordings typically do not contain affirmative AI coverage; newer wordings may exclude algorithmic or AI-generated outputs explicitly. German operators should request policy language confirmation before assuming coverage extends.

A Berufshaftpflichtversicherung (professional indemnity policy) covers claims arising from professional errors. If an AI agent is deployed to perform a professional function (legal research, financial advice, medical triage, technical analysis), and it produces a defective output that causes harm, the professional indemnity policy is the relevant coverage vehicle. The exclusion question is whether the policy treats an AI-generated professional error as covered or excludes it as an automated or algorithmic output. This requires policy-level review, not general assumption.

A Cyberversicherung (cyber liability policy) covers data breaches and cyber incidents. An AI agent that autonomously processes, transmits, or exposes personal data and causes a data breach may trigger a Cyberversicherung claim. However, policies written before 2024 often did not anticipate AI-autonomous data handling and may contain gaps on AI-specific breach events.

Affirmative AI insurance options

The market for dedicated AI liability coverage has advanced since 2025. Munich Re's aiSure product provides parametric AI insurance that settles based on measurable performance data rather than litigation outcomes. Armilla, a Lloyd's coverholder with reported limits of up to approximately USD 25 million [VERIFY current limit], provides AI liability and performance coverage and conducts AI governance evaluations as part of the underwriting process. The ElevenLabs AIUC-1 policy, announced in February 2026, represents the first standalone AI agent insurance policy backed by Munich Re, establishing a benchmark for what covered AI agent perils look like: data and privacy exposure, safety failures, security incidents, reliability failures, and accountability gaps. German operators building AI governance documentation consistent with ISO/IEC 42001 (the AI Management Systems standard) are likely to find that documentation useful in insurance submissions, as underwriters increasingly require evidence of structured AI risk management.

For a detailed analysis of the European AI insurance market and carrier capacity, see agentinsured.eu. For the EU AI Act deployer obligations guide covering all Member States, see the Article 26 deployer obligations guide on agentliability.eu.

What German operators should do now

The immediate compliance priority is Article 50 transparency disclosure, which is binding from 2 August 2026 without any proposed deferral. Every customer-facing and business-to-business AI chatbot or agent operating in Germany must display a clear disclosure at the start of the interaction that the user is interacting with an AI system. This is a technical implementation task, not a strategic one, and it should be completed before 2 August 2026.

The second priority is the Article 5 prohibited practices audit. The prohibited practices under Article 5, including subliminal manipulation, exploitation of vulnerable groups, real-time biometric identification in publicly accessible spaces (with narrow law enforcement exceptions), social scoring, and certain emotion inference uses, have applied since 2 February 2025. Any AI agent deployment that could fall within these prohibited categories should already have been reviewed and cleared or discontinued. BNetzA has enforcement authority for these provisions in Germany.

The third priority is building Article 26 deployer compliance infrastructure for high-risk systems. Even if the proposed Digital Omnibus deferral is adopted before 2 August 2026, the prudent position is to treat the deferral as contingent and to build the Article 9 risk management system, Article 12 logging capability, and Article 14 human oversight mechanism in the current period, when deferral would make these available for quality review rather than rushed for a compliance deadline.

The fourth priority is reviewing commercial insurance coverage for AI agent deployments. Request explicit AI affirmative cover confirmation from your current commercial liability, professional indemnity, and cyber insurers. Document the request and response. If existing policies do not extend affirmatively to AI agent errors, consult a broker about endorsements or dedicated AI liability coverage options. Build your AI governance documentation to a standard (such as ISO/IEC 42001) that supports an insurance submission.

For GDPR compliance, coordinate with your data protection officer or external privacy counsel on Article 22 GDPR compliance for any AI agent making consequential automated decisions about individuals. Identify the relevant Laender DPA for your primary business location and monitor its AI-specific guidance. For financial services, map both the DORA incident reporting chain to BaFin and the EU AI Act serious incident reporting chain to BNetzA before deploying AI agents in operational systems.

Frequently asked questions