C3AN: Custom, Compact and Composite AI Systems - A NeuroSymbolic Approach: 4th-Generation Evolution of Intelligent Systems
Amit Sheth
NCR Chair & Prof; Founding Director, AI Institute at University of South Carolina
Amit P. Sheth, Kaushik Roy, Revathy Venkataramanan, Venkatesan Nadimuth. C3AN: Custom, Compact and Composite AI Systems - A NeuroSymbolic Approach: 4th-Generation Evolution of Intelligent Systems, AI Institute, University of South Carolina, Columbia, United States, 10 Feb 2025. Available at: https://scholarcommons.sc.edu/aii_fac_pub/625/
What appears below is an abridged version for preview (about 50-60% of the above).
ABSTRACT
Artificial Intelligence (AI) systems continue to evolve rapidly. From the architecture perspective, it is evolving from large, monolithic models trained on massive internet data to complex, multi-component “compound” systems and “agentic” frameworks capable of semi-autonomous decision-making. These systems show immense promise yet face numerous challenges in reliability, consistency, transparency, and alignment with user goals. In this article, we propose Custom, Compact, and Composite AI with a Neurosymbolic (C3AN) approach, a framework that paves the way to 4th-generation of AI that integrates data, knowledge, and human expertise to build robust, intelligent, and trustworthy AI systems defined by 14 foundation elements. Custom emphasizes high-quality, domain-specific data and knowledge, along with tailored workflows. Compact highlights resource-conscious implementation that does not require extreme scale to achieve reliable domain adaptation. Composite refers to the integration of multiple AI modules that collaboratively perform domain-specific tasks, handling data, knowledge, and human expert feedback within a cohesive Neurosymbolic framework. Together, these qualities address longstanding issues in large, monolithic, or purely black-box models. We illustrate C3AN in a complex AI system with demands representative of enterprise-class and/or mission-critical applications: Nourich, a disease-specific diet management system recommending recipes based on users’ health conditions and food preferences. We conclude by outlining practical challenges and future research directions for robust, multi-domain adoption of C3AN.
Key Words and Phrases: Neurosymbolic AI, Compound AI, Trustworthy AI, Agentic frameworks, Enterprise Grade AI.
INTRODUCTION
AI has reached an inflection point. From the rise of massive language models that can generate human-like text to the emergence of tool-integrated, multi-component solutions, AI has demonstrated remarkable capabilities in content generation, writing assistance, and conversation-based search. Despite these advances, a significant gap exists between research demonstrations and robust, reliable, production-grade solutions. AI systems must satisfy enterprise-ready requirements in many real-world scenarios, such as healthcare, education, finance, and governance, especially in high-stakes environments. Large, monolithic models often struggle with domain specificity and alignment required in these contexts. This shortfall stems partly from the broad nature of their statistical representations, making retrieving or integrating contextually relevant information difficult. Agentic approaches that automate decision-making processes by delegating tasks to individual agents raise concerns about accountability, safety, and performance consistency due to their often opaque mechanisms.
We posit that such limitations demand AI systems that are Custom: focusing on domain-specific knowledge, workflows, or constraints from domain expertise, Compact: maintaining domain-appropriate performance without indefinite scaling of data or model size, and Composite: orchestrating multiple components such as neural modules, knowledge-bases, and decision processes within a coherent framework. A Neurosymbolic AI approach unifies data-driven neural networks with symbolic knowledge representations and human expertise in one framework. Neural models excel at learning from raw data, while symbolic approaches facilitate the incorporation of domain knowledge in a transparent, adaptable manner. By establishing robust feedback loops for expert oversight, Neurosymbolic systems can deliver stronger guarantees of consistency, accuracy, and alignment with enterprise requirements [1, 2].
We present Custom, Compact and Composite AI Systems with Neurosymbolic Approach (C3AN) as a structured way to integrate data, knowledge, and human expertise leveraging Neurosymbolic AI. It is structured on three core pillars – intelligent, robust, and trustworthy, built on fourteen foundation elements as shown in Figure 1. We first review the four generations of AI systems—monolithic, compound, agentic, and copilot—and motivate a new 4th generation: C3AN. We then detail its 14 foundation elements that unify data-driven learning, explicit domain knowledge, and iterative expert feedback. Next, we show how Nourich clarifies C3AN’s approach. We conclude with practical challenges and future directions. An unabridged version of this article is at: https://tinyurl.com/unabridgedversion
BACKGROUND: WHY AI NEEDS A PARADIGM SHIFT
Generative AI has transformed knowledge-worker tasks, enabling text generation, summarization, retrieval, and media creation. Large models like ChatGPT have facilitated engaging conversations, search assistance, and content creation. Analysis of ChatGPT prompts [3] shows frequent use for writing, search, and research. However, generative models rely on vast training data and can produce unreliable outputs, exhibit ethical misalignment, and are costly with limited value for enterprise needs. Many domains (e.g., manufacturing, or nutrition assistance) require targeted models rather than generalized large models. Substantial costs ($100M+ for trillion-parameter LLMs) limit retraining to large corporations [4], while smaller optimized models suffice for specialized tasks [5–7].
Enterprises demand accuracy, transparency, and alignment with organizational goals. Generic models risk hallucinations or misalignment, making them unreliable for mission-critical tasks. Complex real-world problems also require modular AI systems that integrate specialized models. Combining multiple AI models into cohesive systems enables domain-specific decision-making. While deep learning models have delivered strong performance in various tasks, they introduce new challenges like black-box models, data-intensive training, hallucinations, and limited domain adaptability. Classical symbolic AI excels at explicit reasoning and domain knowledge but struggles with unstructured data. Neurosymbolic AI [8] bridges this gap. Yet many approaches do not emphasize the custom, compact, and composite nature demanded by enterprise scenarios. C3AN explicitly integrates domain customization, resource-conscious implementation, and composite orchestration of neural and symbolic modules, with humans providing feedback.
GENERATIONS OF AI SYSTEMS
Figure 2 summarizes the evolution of AI systems across four “generations.” Each builds on insights from predecessors while introducing new capabilities.
0th Generation: Monolithic AI. Large Neural Networks or LLMs (e.g., BERT) trained on massive internet-scale data showed success in tasks like machine translation but required extensive data and lacked transparency.
1st Generation: Compound AI. Tool-integrated, multi-model systems allow generalist language models to be instructed across diverse tasks but can be sensitive to minor prompt variations, leading to overhead and reliance on external knowledge sources [9].
2nd Generation: Agentic AI. Autonomous agents leverage prompting for planning/orchestrating workflows involving multiple language agents. They promise automation of simple workflows but raise concerns regarding safety, debugging, and domain adaptation overhead [10].
3rd Generation: Copilot. Humans rejoin the loop for oversight, verifying agent decisions and refining workflows for improved reliability. This approach is safer but can be slower and interface-heavy, prompting the question: why not incorporate domain knowledge and human expertise in the core?
4th Generation: C3AN. This paradigm systematically embeds domain knowledge (custom), employs targeted data rather than massive scale (compact), and unifies multiple modules (composite) within a Neurosymbolic framework that combines data-driven learning with symbolic reasoning and expert feedback.
THE 14 FOUNDATION ELEMENTS AND 3 PILLARS OF C3AN
C3AN supports 14 foundation elements leading to intelligent, robust, and trustworthy AI systems. These elements require combining symbolic knowledge with data-driven learning and expert input to remain custom, compact, and composite. The callout box titled “Neurosymbolic AI: 14 Foundation Elements to Achieve 3 Pillars of C3AN Framework” provides a summary. Humans remain essential for providing domain context and iterative oversight.
Neurosymbolic AI: 14 Foundation Elements to Achieve 3 Pillars of C3AN Framework:
Neurosymbolic AI combines the pattern recognition and generalization strengths of neural networks with the abstraction and reasoning capabilities of knowledge graphs for symbolic inference. Our proposed Neurosymbolic approach enhances the C3AN framework by integrating 14 foundation elements, ensuring it is intelligent, robust, and trustworthy. By leveraging the right knowledge alongside data-driven models, Neurosymbolic AI systematically captures these foundation elements to build reliable and adaptive C3AN systems.
Reliability ensures dependable, error-free operations (Robust) while building trust by meeting expectations (Trustworthy).
Consistency ensures coherence in outputs (Robust), boosting trust by avoiding contradictions (Trustworthy).
Alignment ensures outputs meet user goals (Trustworthy) while aiding the system’s ability to reason about these goals (Intelligent).
Analogy enables the system to see patterns across contexts (Intelligent), fostering adaptability.
Abstraction elevates low-level data to high-level constructs (Intelligent), improving stable processing (Robust).
Causality uncovers cause-effect relationships (Intelligent).
Instructability enables the system learn from user input (Intelligent), enhancing trust through alignment (Trustworthy).
Reasoning covers logical inference with domain knowledge (Intelligent).
Planning arranges steps to reach goals (Intelligent), boosting trust with actionable plans (Trustworthy).
Grounding ties decisions to real entities (Trustworthy) while maintaining context relevance (Intelligent).
Attribution clarifies data sources, promoting trust and accountability (Trustworthy).
Interpretability fosters user understanding of the system’s logic (Trustworthy), reflecting advanced modeling (Intelligent).
Explainability offers clear reasons for outputs, fostering user trust (Trustworthy).
Safety adheres to ethical guidelines and avoids harm, building trust (Trustworthy) while operating within stable boundaries (Robust).
Example Scenario: Nourich
We illustrate C3AN in a system called Nourich, a disease-specific diet management approach recommending meals by analyzing recipes against users’ health conditions and preferences. The system showcases custom domain knowledge with tailored workflows, compact resource usage, and a composite approach integrating AI modules, diverse data, and human feedback.
Nourich: Disease Specific Diet Management System.
Objective: The goal is to analyze whether a recipe is suitable for a chronic condition (e.g., diabetes) and provide alternative suggestions through ingredient or cooking method substitutions.
Challenges:?Analyzing recipe suitability requires extracting ingredients, cooking methods, and their interactions from unstructured data and then inferring health impacts. This involves compositional reasoning [11], integrating nutrition, disease impact, and glycemic index. Incorrect or incomplete recommendations can be harmful, so high accuracy and transparent reasoning are vital.
Framework: C3AN meets these needs through modular pipelines for knowledge curation, data processing, and reproducible workflows. Nourich (Figure 3) has five tiers that include: (i) Custom data and knowledge tailored to dietary requirements, (ii) Compact size leveraging curated domain data, (iii) Composite Engine integrating neural parsers, knowledge graphs, rule based filters, symbolic reasoning, and feedback modules for refining knowledge. Other modules ensure explainability (Tier 5) and generate recipe modifications (Tier 4).
Knowledge Types: Nourich accommodates (i) Taxonomy: Hierarchical ingredient and cooking method classifications, (ii) Causal: cause-effect relations, e.g. grilling meat and carcinogens, (iii) Logical Constructs: formalizing reasoning, e.g. “high-cholesterol ingredient is unsuitable for diabetes,” and (iv) Rules: for implicit knowledge, e.g. “carbohydrate-to-fiber ratio of 10:1 implies whole grain.” Neurosymbolic methods ensure accountability and explainability, as domain experts can override or update rules [12].
Illustration of the 14 Foundation Elements in Nourich. Reliability ensures outputs are correct. Nourich reliably classifies recipes by combining knowledge-based filters and neural parsers. Consistency avoids contradictory recommendations, enforcing global constraints such as excluding peanuts. Alignment ensures compliance with user and domain needs (vegan, diabetic) through embedded healthcare guidelines. Analogy supports structured reasoning, e.g., mapping tofu to chicken as protein substitutions, validated by domain knowledge. Abstraction groups details into high-level concepts (“low-carb”) so experts can refine borderline cases. Causality identifies causeeffect (sugar intake → glycemic spike) via neural patterns and medical guidelines. Instructability adapts to user directives (“avoid peanuts”) in the knowledge graph. Reasoning merges data-driven insights and symbolic constraints. Planning outlines multi-day meal plans with knowledge-based checks. Grounding ties suggestions to real-world guidelines like USDA, with expert override. Interpretability provides step-by-step breakdown of how nutrition factors affect recommendations. Explainability clarifies rationales (“High sugar causes blood glucose spikes”), citing medical research. Attribution cites sources like nutrition databases or dietician oversight. Safety enforces constraints for at-risk users with symbolic checks and expert sign-off.
IMPLEMENTATION CHALLENGES FOR C3AN AND FUTURE DIRECTIONS
Though promising, incorporating all 14 foundation elements into a custom, compact, and composite framework can be demanding. In particular, Complex Integration requires the Neurosymbolic layer to unify data-driven insights with symbolic knowledge for domain customizations, and to incorporate human expertise without overwhelming users [13, 14]. Scalability vs. Compactness poses a balance between building large models and focusing on relevant data. Model distillation or
“just-in-time” retrieval helps maintain a compact footprint while preserving performance. Knowledge Base Maintenance is also key: when domains evolve quickly, domain constraints must be updated without retraining entire subsystems—an advantage of the custom approach, provided overhead remains minimal. Moreover, User Trust and Adoption requires that Explainability, Attribution, and Safety are emphasized. Because the solution is composite, partial explanations from different modules need coherent integration, while iterative feedback must be supported by a well-designed UI. In a Multi-Stakeholder Environment, conflicting objectives and diverse domain experts demand robust alignment; reconciling these viewpoints without losing efficiency often hinges on a custom architecture. Validation and Benchmarking go beyond accuracy: C3AN must also be assessed for reliability, alignment, resource consumption, and compliance with domain regulations to ensure real-world effectiveness.
Practical Implications for Enterprise AI
C3AN incorporates:
? Custom Implementation: Update domain-specific constraints and knowledge with minimal overhead.
? Compact Deployment: Rely on targeted data and relevant knowledge, reducing operational costs.
? Composite Framework: Combine domain experts, symbolic reasoning, and neural modules for complex tasks.
? Neurosymbolic Intelligence: Leverage data-driven inferences refined by symbolic com- ponents and human feedback.
? Agile Model Development: Apply iterative data augmentation and user feedback to maintain alignment with changing needs.
? Secure Model Hosting: Deploy in high-assurance environments (e.g., healthcare, finance) to ensure compliance with strict regulations. Keeping the model and inference layers within the security perimeter enhances security and benefits enterprises.
Future Directions.
To broaden the real-world adoption of C3AN, several areas call for deeper investigation. Formal Semantics for Composite Workflows can specify clear protocols for how neural and symbolic modules exchange information to yield coherent outputs. Adaptive Customization via Continual Learning allows incremental integration of new domain insights or policies without retraining entire subsystems. Scalable yet Resource-Efficient Deployments uphold a compact footprint, even at enterprise scale. In parallel, Multi-Stakeholder Governance and Ethical Compliance can embed regulatory requirements and ethical norms directly in symbolic layers. User Interfaces for Explainability and Feedback enable experts to adjust system outputs with transparent rationales. A Platform-Centric Approach empowers organizations to adopt composite AI solutions through templated modules rather than relying solely on large technology providers. Emphasizing AI for Everyone ensures accessible systems that function with limited resources or connectivity, extending benefits to underserved communities. Lastly, Re-evaluating Intelligence and Expanding Benchmarks looks beyond simple question-answer metrics to gauge deeper cognitive skills such as abstraction and contextual understanding [15].
CONCLUSION
C3AN offers a coherent 4th-generation framework that addresses the limitations of monolithic, compound, agentic, and copilot AI systems. By systematically embedding domain knowledge and workflows (Custom), maintaining targeted resource usage (Compact), integrating specialized modules (Composite), and merging neural and symbolic reasoning with human feedback (Neurosymbolic), C3AN provides a pathway to more reliable, transparent, and efficient AI solutions. These capabilities are grounded in its 14 foundation elements spanning Intelligent, Robust, and Trustworthy aspects of AI. We illustrated these concepts through Nourich, a disease-specific diet management system that applies compositional reasoning and domain-specific knowledge to recommend healthier recipes. This example shows how the custom, compact, and composite pillars shape an end-to-end AI pipeline: from structured curation of medical and nutritional information to rule-based and neural modules, all of which benefit from feedback loops with domain experts. By anchoring recipe recommendations in explicit knowledge resources (such as nutrient thresholds and cooking method constraints), Nourich demonstrates higher reliability, clarity, and alignment with user health goals. Beyond Nourich, the C3AN approach—combining data-driven learning and formalized knowledge with human oversight—can be adapted to various enterprise and missioncritical domains. Its modular design makes it suitable for tasks where requirements evolve quickly, while the emphasis on targeted datasets curbs the explosive growth in model size. This blend of custom focus and compact operation helps organizations achieve cost-effective, trustworthy AI solutions. We encourage researchers and practitioners to adopt and refine C3AN for real-world use cases. By weaving together data, knowledge, and expertise under a unified framework, future AI deployments can deliver intelligent, robust outcomes that remain accountable to industry standards, regulatory frameworks, and end-user needs. As AI continues its transformative role, C3AN sets the stage for the next wave of mature, domain-centric AI systems grounded in explicit knowledge and agile enough to handle the most pressing demands of modern enterprises.
ACKNOWLEDGMENT
This work was partially supported by the National Science Foundation (NSF) under the following grants: NSF Grant #2335967 "EAGER: Knowledge-Guided Neurosymbolic AI with Guardrails for Safe Virtual Health Assistants", NSF Grant #2119654 "RII Track 2 FEC: Enabling Factory-to-Factory (F2F) Networking for Future Manufacturing", and NSF Grant #2350302 "SaTC: CORE: Small: Enhancing Security and Mitigating Harm in AI-Generated Vision-Language Models". The authors also benefited from discussions with research team members and collaborators.
REFERENCES
[1]???? Amit Sheth, Kaushik Roy, and Manas Gaur. Neurosymbolic artificial intelligence (why, what, and how). IEEE Intelligent Systems, 38(3):56–62, 2023.
[2]???? Amit Sheth and Kaushik Roy. Neurosymbolic value-inspired artificial intelligence (why, what, and how). IEEE Intelligent Systems, 39(1):5–11, 2024.
[3]???? Shayne Longpre, Robert Mahari, Ariel N Lee, Campbell S Lund, Hamidah Oderinwale, William Brannon, Nayan Saxena, Naana Obeng-Marnu, Tobin South, Cole J Hunter, et al. Consent in crisis: The rapid decline of the ai data commons.
In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track.
[4]???? Kate Crawford. Generative AI’s environmental costs are soaring — and mostly secret. Nature, 626(8000):693, 2024.
[5]???? Janakiram MSV. All about deepseek: The chinese ai startup challenging the us big tech. Forbes, January 2025.
[6]???? Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In The Twelfth International Conference on Learning Representations, 2024.
[7]???? Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. A survey on knowledge distillation of large language models, 2024.
[8]???? Tarek R. Besold, Artur S. d’Avila Garcez, Sebastian Bader, Howard Bowman, Pedro M. Domingos, Pascal Hitzler, Kai-Uwe Kühnberger, Luís C. Lamb, Daniel Lowd, Priscila Machado Vieira Lima, Leo de Penning, Gadi Pinkas, Hoifung Poon, and Gerson Zaverucha. Neural-symbolic learning and reasoning: A survey and interpretation. CoRR, abs/1711.03902, 2017.
[9]???? Matei Zaharia, Omar Khattab, Lingjiao Chen, Jared Quincy Davis, Heather Miller, Chris Potts, James Zou, Michael Carbin, Jonathan Frankle, Naveen Rao, and Ali Ghodsi. The shift from models to compound ai systems. https: //bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024.
[10]?? Shomit Ghose. The next “next big thing”: Agentic ai’s opportunities and risks, 2024. Accessed: 2025-01-12.
[11]?? Sania Sinha, Tanawan Premsri, and Parisa Kordjamshidi. A survey on compositional learning of ai models: Theoretical and experimetnal practices. arXiv preprint arXiv:2406.08787, 2024.
[12]?? AIISC. Nourich co-pilot: Custom, compact and neurosymbolic diet ai model(abridged version), 2024. Accessed: 2025-01-28.
[13]?? Manas Gaur, Ugur Kursuncu, Amit Sheth, Ruwan Wickramarachchi, and Shweta Yadav. Knowledge-infused deep learning. In Proceedings of the 31st ACM Conference on Hypertext and Social Media, HT ’20, page 309–310, New York, NY, USA, 2020. Association for Computing Machinery.
[14]?? Amit Sheth, Manas Gaur, Kaushik Roy, Revathy Venkataraman, and Vedant Khandelwal. Process Knowledge-Infused AI: Toward User-Level Explainability, Interpretability, and Safety. IEEE Internet Computing, 26(5):76–84, 2022.
[15]?? Ramesh Jain. Advancing from Science of AI to Intelligence Science. Communications of the ACM, 2025. In print.
FOOTNOTES
C3AN Use Cases:
Project Pages - AIISC Wiki:
Tutorials/Talks:
Senior Lead Architect at HSBC creating innovative sustainable solutions
1 周Thank You Sir. As usual you simplified complexity and focused on adaptation.
AI Advisor and Trainer of Leaders | Investor, Builder, Speaker, Executive Coach
3 周This is exactly the shift AI needs to make to be truly enterprise-ready. Moving from generic, data-hungry LLMs to targeted, knowledge-infused AI models is the only way to ensure reliability, cost efficiency, and real-world applicability. The neuro-symbolic approach is particularly compelling, balancing deep learning with structured reasoning could be the key to unlocking AI that’s not just powerful but also trustworthy. The Gartner and Acemoglu predictions make sense; enterprises need AI that works within their constraints, not just the latest cutting-edge black box. Looking forward to seeing how this evolves!
Amit, This read is even more stimulating than my morning caffeine today, commend the direction and approach ?? ??
Senior Managing Director
3 周Amit Sheth Very Informative. Thank you for sharing
Helping founders & security leaders automate compliance—get SOC 2 & ISO 27001 with Comp AI (Open Source Drata & Vanta Alternative)
3 周Love the focus on trust and interpretability. Do you think enterprises will prioritize composite AI systems over LLMs primarily for cost efficiency, or is it more about control and precision?