Anuradha Weeraman is Co-Founder and CTO at Verdentra.
Artificial Intelligence (AI) is undeniably reshaping industries. As generative AI models demonstrate increasingly impressive capabilities at a breakneck pace, their transformative potential is clear. However, for enterprise adoption to achieve its full potential, particularly in mission and safety-critical applications, several significant hurdles must be addressed.
The most advanced frontier models still operate as “black boxes.” Understanding their internal workings is an active area of research, particularly with Anthropic’s research on mechanistic interpretability to “open up the hood” on how models work. This opacity remains a critical area of concern for leaders who need to trust and verify AI-driven decisions, particularly in highly regulated and safety-critical environments.
The current wave of AI innovation is fueled by Transformer-based Large Language Models (LLMs). These deep neural networks, often featuring billions of parameters and trained on vast data corpora, are adept at recognizing patterns and exhibiting emergent properties of intelligence. Post-training and fine-tuning enables them to unlock further capabilities and thought processes hinting at an innate capability for logical reasoning.
Key strategically important challenges persist, however:
Explainability And Interpretability: The difficulty in dissecting the decision-making process and achieving consistent, verifiable outcomes is a key area of concern. This lack of transparency is particularly a barrier for safety-critical applications such as those in healthcare and autonomous robotics.
Factual Accuracy And Hallucinations: LLMs can generate highly coherent yet factually inaccurate or “hallucinated” responses due to their fundamentally probabilistic underpinnings. This poses challenges to decision-making quality and can introduce safety and reputational risks.
Bias And Fairness: Trained on vast internet-scale datasets, models can reflect and amplify existing societal biases. This can produce unfair or discriminatory outputs, requiring extensive guardrails and posing risks of regulatory and reputational damage.
Safety And Robustness: In safety-critical systems, ensuring robust and reliable performance with out-of-distribution data can be a challenge for the current generation of models. This poses significant safety, liability and reputational risks for businesses.
Scalability Of Reasoning: While techniques such as Chain-of-Thought offer insights into model “thinking,” flaws in reasoning often rooted in “hallucinated” data limit the adaptability of these models to novel situations not covered in training. This hinders true understanding and generalization.
Composite AI: A Strategic Solution
To navigate these complexities, Composite AI emerges as a strategic solution. This approach synergizes multiple AI techniques and disciplines to architect more robust, reliable and ultimately more valuable solutions. Leading frontier models are already integrating Composite AI elements, such as tool-calling, to deliver state-of-the-art performance benchmarks unattainable through a singular methodology.
One promising approach to Composite AI is Neurosymbolic AI. It strategically combines the strengths of deep neural networks (“Neuro”) with classical Symbolic AI approaches, bridging their respective gaps to unlock capabilities neither can achieve in isolation.
Consider the distinct strengths:
Neural AI: Excels at pattern recognition and learning from vast unstructured datasets, often in a largely unsupervised manner, leading to emergent capabilities in language and reasoning. Decision-making, however, emerges from an opaque “black box.”
Symbolic AI: Leverages explicit knowledge representation, rules and logic, often encoded in formal language. This makes it strong in formal, verifiable logical reasoning, providing transparent and explainable decisions.
Key Strategic Advantages
Improved Interpretability And Trust: Crucial for regulatory compliance and adoption in high-stakes, safety-critical domains like healthcare and finance.
Enhanced Reasoning And Generalization: Enables learning from significantly fewer examples than purely neural approaches by reasoning over symbolically represented knowledge, leading to more generalized understanding.
Reduced Data Dependency: Can effectively handle tasks with limited data by incorporating pre-defined rules and domain knowledge, reducing the need for vast datasets to achieve the same level of generalization through neural approaches alone.
Greater Robustness And Reliability: Better equipped to handle out-of-distribution data and novel scenarios safely and reliably by incorporating safety parameters, constraints and logical invariants.
Integration Of Domain Expertise: Allows for direct incorporation of expert knowledge (e.g., via knowledge graphs) into the learning and reasoning process, making AI solutions more contextually aware and effective.
However, realizing these advantages requires navigating several implementation complexities.
Implementation Considerations and Challenges
While employing composite AI techniques unlocks significant potential, organizations must navigate several implementation challenges:
Technical Complexity: Integrating neural networks with symbolic reasoning engines requires sophisticated architectural design and specialized talent. These hybrid systems often necessitate separate training phases, making end-to-end training considerably more complex than traditional machine learning pipelines.
Knowledge Engineering: Converting domain expertise into formal, machine-readable representations demands substantial time investment and specialized skills. This knowledge acquisition process often becomes a bottleneck in project timelines and requires ongoing maintenance as business rules evolve.
Performance Trade-Offs: Real-time applications face inherent constraints due to the computational overhead of symbolic reasoning, knowledge graph traversal and the general reduction in inference speeds compared to pure neural approaches. Organizations must carefully evaluate whether the advantages of interpretability and robustness warrant these performance costs.
Production Readiness: Most Neurosymbolic successes remain in research or pilot phases, making it difficult to assess real-world scalability, ROI and long-term viability.
Conclusion
Adopting a multidisciplinary composite AI approach is gaining attention beyond academic circles, as organizations look for ways to address the limitations of current AI systems. While not a silver bullet, composite and neurosymbolic AI offer potential for improving reliability, interpretability and reasoning—areas where traditional machine learning can fall short.
However, these approaches are still evolving, and integrating them into real-world systems involves significant technical and organizational challenges. For teams working on long-term AI strategies, this is an area worth monitoring, experimenting with and considering where appropriate—but with a clear-eyed view of its current maturity and limitations.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
