The typical ML engineer follows a familiar workflow: choose an architecture, train, evaluate. This mindset leads to searching for a single model that solves the entire problem on its own. In practice, no single computational paradigm fully meets the demands of real-world applications, especially in the public sector and regulated environments. The solution is not a better model but a better composition of models.
Hybrid AI: A design pattern, not an algorithm
Hybrid AI is not a technique but an architectural pattern for composing heterogeneous components into a unified system. The central principle is that every AI technique has characteristic failure modes, and these failure modes do not overlap. A large language model (LLM) generates confident but false answers. A rule engine cannot generalise. A Bayesian network requires prior distributions that are rarely available. An evolutionary algorithm finds solutions but cannot explain why. By combining techniques with complementary failure modes, a system emerges where the weakness of one component is covered by the strength of another. The design question shifts from “which model solves all subproblems” to “how do I orchestrate multiple techniques so that each does what it excels at.” A critical distinction applies here: an ensemble of three transformers is not Hybrid AI, since it deploys the same technique three times. By contrast, a pipeline combining a transformer, a constraint solver and a knowledge graph constitutes genuine Hybrid AI, three different computational paradigms in composition.
Neurosymbolic AI: The most powerful instantiation
Neurosymbolic AI is a subset of Hybrid AI that locks both poles to specific paradigms: the connectionist pole (neural networks, transformers, diffusion models, anything that learns from data via gradient descent) and the symbolic pole (logic programming, knowledge graphs, ontologies, domain-specific languages, theorem provers, anything that operates on discrete structures with formal inference). These two paradigms do not merely complement each other; they exhibit precisely inverse trade-off profiles. Neural systems learn from data, generalise to novel inputs, tolerate noisy inputs and handle ambiguity, but never guarantee correctness and struggle to explain their reasoning. Symbolic systems guarantee correctness within their model, explain their reasoning through derivation traces and handle compositionality, but do not learn from data and break on noisy inputs. Their combination is not additive but multiplicative: the resulting system acquires properties that neither component possessed individually. Henry Kautz (2020) proposed a taxonomy of six integration patterns, from loose to tight coupling. For the practitioner, three are most actionable. First, Neural to Symbolic (pipeline): the neural component handles perception and extraction, then its outputs feed a symbolic reasoner. For example, an LLM extracts structured data from natural language, then a logic program evaluates regulatory compliance. Second, Symbolic to Neural (constrained generation): symbolic rules constrain the search space before or during neural processing. For example, an ontology filters which documents can be retrieved before passing them to the LLM. Third, Unified (tight coupling): neural and symbolic components share representations, such as graph neural networks operating over knowledge graphs, or differentiable logic programming where weights are learned but the structure remains logical. The practical recommendation is to start with pipeline (loose coupling) and tighten the coupling only where required.
RAG: A design family across the Hybrid-to-Neurosymbolic spectrum
Retrieval-Augmented Generation (RAG) is already familiar to most AI developers through the basic pattern: embed, retrieve, generate. However, this is merely the entry point of a design family, and the choice within this family determines whether a system is merely Hybrid or genuinely Neurosymbolic. At one end lies Naive RAG: chunk documents, embed, retrieve top-k via cosine similarity. It works but suffers from irrelevant retrieval and inability to reason over structure. In the middle, Advanced RAG and Self-RAG/Corrective RAG improve metrics through query rewriting, re-ranking and self-critique, but retrieval remains statistical. In Agentic RAG, the system dynamically plans its retrieval strategy. In Graph RAG, retrieval operates over a knowledge graph via SPARQL or Cypher queries, introducing genuine symbolic reasoning. At the far end, Normative RAG feeds retrieved context into a formal reasoning engine such as eFLINT, a domain-specific language developed at Utrecht University for encoding legal norms as executable specifications based on transition systems and Hohfeld’s legal framework. This architecture produces something no pure RAG system can deliver: verifiable normative compliance. The LLM handles natural language understanding and generation, retrieval brings relevant regulatory texts, and the symbolic reasoner produces formally proven conclusions with a complete derivation trace. The difference is critical: instead of a system saying “I believe this health data access is permitted,” it can prove why it is permitted, step by step, with references to specific regulatory articles.
Why this matters for the public and private sectors
In the private sector, neurosymbolic RAG systems reduce hallucinations, improve auditability and produce documented decisions. In the public sector, where every decision carries institutional consequences, formal verifiability is not a luxury but a prerequisite. Regulation (EU) 2024/1689 (the EU AI Act) requires explainability and human oversight for high-risk applications. The new European Health Data Space (EHDS, Regulation 2025/327) establishes a multi-layered regulatory framework (GDPR as horizontal baseline, EHDS as sectoral overlay) that demands precisely this kind of architectural composition. The EHDS example is instructive. Health Data Access Bodies must evaluate secondary-use applications against permitted purposes (Article 53), verify absence of prohibited uses (Article 54), confirm data minimisation, check secure processing environment accreditation and respect data subject opt-out rights. Each of these functions involves normative reasoning that can be formalised. An eFLINT-backed system could provide what no manual process can: formal traceability from every data access decision back through the specific EHDS articles, GDPR provisions and national implementing measures that justify it, with the reasoning chain being machine-verifiable, auditable and adaptable when norms change. The critical recognition for the developer is threefold. Hybrid AI is not a buzzword but the acknowledgement that no single paradigm solves real-world problems alone. Neurosymbolic AI is the most powerful instantiation of this principle, because it combines the two most opposite and complementary paradigms. And RAG is the first practical entry point: the place where the average developer is already doing Hybrid AI, often without realising it.
Sources
Kautz, H. (2020). The Third AI Summer. Keynote, AAAI Conference on Artificial Intelligence: taxonomy of six neurosymbolic integration patterns from loose to tight coupling.
Van Binsbergen, L.T., Liu, L.C., van Doesburg, R. & van Engers, T. (2020). eFLINT: a Domain-Specific Language for Executable Norm Specifications. Proceedings of GPCE ’20, ACM: presents the eFLINT language for executable norm encoding based on transition systems and Hohfeld’s legal framework.
Van Binsbergen, L.T. et al. (2025). Lawful and Accountable Personal Data Processing with GDPR-based Access and Usage Control in Distributed Systems, arXiv: applies eFLINT for automated normative reasoning under GDPR with XACML integration in distributed systems.
Gao, Y. et al. (2024). Retrieval-Augmented Generation for Large Language Models: A Survey, arXiv: comprehensive survey of RAG methodologies from Naive to Advanced, Modular and Agentic forms.
Regulation (EU) 2025/327 (EHDS), EUR-Lex: establishes the European Health Data Space, Health Data Access Bodies and the multi-layered regulatory framework for secondary use of health data.
Regulation (EU) 2024/1689 (EU AI Act), EUR-Lex: Regulation on Artificial Intelligence, with requirements for explainability, human oversight and risk-based compliance.