Germany - Market Insights

In just a few years, artificial intelligence (AI) has moved rapidly from experimental deployments into business-critical infrastructure across German industry. In our own work, we are seeing that AI systems are increasingly embedded in manufacturing control systems, logistics optimisation tools, medical diagnostics, HR screening, fraud detection, predictive maintenance or knowledge-based professional services. What distinguishes the current phase is not necessarily technological novelty, but operational dependence. Companies are no longer piloting AI; they are building core processes and compliance structures around it. Also, agentic AI systems and other probabilistic tools are considered cost-cutters and have become essential for companies of all sizes in economically and politically challenging times.

This shift leads to a growing tension between technological reality and legal risk allocation. German contract practice is still largely shaped by deterministic quality and warranty concepts. It struggles to accommodate systems that learn, evolve and produce probabilistic outputs. At the same time, regulatory pressure continues to increase through European digital regulation and national enforcement dynamics. Yet many of the most commercially relevant legal questions are not regulatory in nature, but contractual: who ultimately bears the risk when AI behaves unpredictably, produces biased results or degrades over time?

In German commercial practice, these questions are increasingly shaping procurement negotiations. In the absence of fundamental court decisions on AI contract law, creativity is required in contract drafting and negotiation practices in this area. In our experience, the soundest approach is to gradually transform the structure of performance obligations and acceptance testing. This will also create a shift in warranty and liability regimes. Another important issue to consider when negotiating AI-related contracts is how to allocate compliance risks associated with the EU AI Act, the General Data Protection Regulation (GDPR) or other applicable legislation.

Why traditional contract concepts fail for AI systems

Many of the AI systems currently on the market are hybrid offerings that combine software licences, cloud infrastructure, data pipelines and continuous model training. These systems are often heavily integrated with other hardware or software and are implemented through complex IT projects. Thus, contracts must accommodate complex multi-stage project arrangements, an unpredictable regulatory landscape and often huge expectations concerning the commercial benefits.

Although European digital regulation understandably attracts a lot of attention, we still consider the allocation of project risks to be the most important mechanism for distributing risk. Regulators intervene ex post; contracts shape incentives ex ante. Only well-designed contracts can also mitigate regulatory uncertainty by allocating documentation duties, audit cooperation, update responsibilities and decision authority. Regulation increases pressure, but contracts determine who absorbs that pressure operationally and financially.

However, increasing operational integration of AI systems collides with a contract culture that remains heavily influenced by traditional paradigms: clearly specified deliverables, objective acceptance testing, fixed functionality warranties and mostly fault-based liability concepts. These instruments have existed since the advent of modern German contract law in the late 19th century, and they remain “hard-coded” into the German Civil Code. As a legal practitioner specialising in IT, fitting these concepts to modern IT projects has always been challenging. What could still be applied to deterministic software systems already posed increasing challenges to an industry that was becoming ever more dynamic and agile over the previous decade. However, these paradigms fit poorly with machine learning models, the outputs of which are probabilistic by nature and whose behaviour changes dynamically through retraining, data drift and evolving environments.

Many disputes in the field show that this mismatch often remains hidden at contract signing and only becomes visible once systems operate at scale. While the legal framework promises predictability, technology can only offer statistical confidence. Acknowledging this paradigm shift, strict concepts of work acceptance and solid warranties may still appear favourable on paper. Our experience shows, however, that such comforting clauses will not work as intended unless they are tailored to the specific characteristics of AI.

Adapting contract law to the needs of a new era

When drafting and negotiating AI-related agreements, clients are likely to prioritise compliance issues. Regulatory risk transfer is indeed a sensitive area, largely triggered by the EU AI Act, associated fines and regulatory oversight. Although these obligations typically apply to operators or deployers, customers often attempt to pass on all compliance responsibilities, including documentation, audit support and transparency obligations, to suppliers. Suppliers push back, arguing that they cannot control how customers deploy and contextualise AI outputs. Instead, a level playing field should be established, with strong information obligations on the part of the supplier, particularly with regard to system documentation and operational responsibilities for the customer.

However, certain other AI-specific factors are often not sufficiently considered. Is underperformance a defect or an inherent system characteristic? Can learning behaviour be contractually frozen without undermining functionality? How should acceptance regimes operate when systems evolve?

Classical German IT contracts assume that a system can be objectively assessed against a defined specification and that deviations constitute a defect triggering remediation and liability. However, AI systems challenge both of these assumptions. The traditional “Werkvertrag” concepts of acceptance at handover and warranty obligations appear inappropriate for handling the implementation and operation of AI systems.

In AI, functional correctness is rarely binary. Even well-designed systems inevitably produce false positives and false negatives. At any rate, many contracts still attempt to translate probabilistic performance into absolute contractual guarantees. Often, neither the vendor nor the buyer has adapted their licence and project agreements to the requirements of AI-related contracts, treating AI as “just another feature” of traditional software systems or cloud services. Consequently, customers seek measurable guarantees for accuracy and sometimes even business outcomes. System suppliers counter with best-effort clauses, statistical tolerance thresholds and exclusions for data quality and operational misuse. This conceptual discrepancy gives rise to new arguments in classic contractual battlegrounds.

First, parties should acknowledge that the behaviour of AI systems evolves over time. Even when the underlying software remains unchanged, continuous retraining and model updates can alter output characteristics. A system that satisfies the acceptance criteria at launch may deteriorate several months later without any obvious defect. This raises difficult questions about ongoing conformity and responsibility for performance drift, since — under traditional German contractual models — this risk would usually lie entirely with the customer after acceptance. Conversely, subscription models would likely be treated as rental contracts under German law, shifting the risk back to the vendor. Thorough contract design seeks to balance these risks rather than allocating them to just one party.

Second, questions surrounding the causality of AI behaviour further complicate warranty regulations and liability attribution under AI contracts. Harmful outcomes produced by an AI system may result from the interaction of various factors, such as the quality of the training data, the model architecture, the parameter choices, user behaviour, external data sources and the operational context. Traditional fault concepts presuppose traceable causation, which AI systems are often unable to provide in a legally robust manner. This would put customers in a situation where it would be almost impossible to prove a system defect. To build sustainable customer relationships and avoid disputes, we advocate service models that provide continuous support.

Third, instead of applying rigid acceptance and defect concepts, parties should experiment with service-level regimes linked to performance metrics, continuous benchmarking, adaptive adjustment mechanisms and structured escalation workflows. These models resemble outsourcing and cloud governance frameworks more than classical licence agreements. Such probabilistic key performance indicators (KPIs) must, however, be carefully drafted to remain legally operable under German contract law. Remediation mechanisms must balance flexibility with predictability. Governance structures must be clearly allocated to avoid implicit risk shifts.

In light of the strict interpretation of standard contractual clauses by German courts, even in business-to-business (B2B) contexts, AI providers in particular face the challenging task of operationalising the AI-inherent uncertainty within contractual structures originally designed for deterministic performance and defect concepts. Nevertheless, this evolution is a necessary adaptation to technological reality. It is a reality that is not always acknowledged by procurement departments, which often stick to the legacy of previous IT projects rather than embracing the opportunities for novel contract design. What should also be kept in mind is that AI drafting tools fed with the same contract history are unlikely to generate the innovative clauses that this paradigm shift requires.

Contract design as competitive advantage

In brief, currently both vendors and customers often misplace confidence in traditional contractual regimes. Suppliers overestimate the protective effect of broad disclaimers. Courts may ultimately interpret aggressive exclusions narrowly, especially where core performance expectations are undermined. This risk may particularly arise when the standard terms and conditions of AI providers are assessed against traditional legal concepts codified in German civil law. Conversely, customers sometimes place too much trust in rigid liability provisions to eliminate technical uncertainty.

Effective risk allocation requires technical literacy, realistic expectations and continuous governance — not merely stronger clauses. In the German AI market, contract design is becoming a strategic differentiator. Organisations that invest in balanced risk allocation, transparent governance and scalable liability models achieve faster deal velocity, stronger trust relationships and improved insurability. Those relying on outdated templates risk friction, disputes and hidden exposure. AI forces German contract practice to evolve — not by abandoning legal rigour, but by applying it honestly to probabilistic systems.