Neurosymbolic AI is a framework that combines the core strengths of symbolic AI, which represents explicit enterprise structure, relationships, rules, constraints, and decision logic, with neural AI, which powers capabilities such as language understanding, pattern recognition, and synthesis. This framework builds upon the domain expertise and rules of symbolic AI and the capability of neural AI to find and learn from the complex patterns and relationships found in vast amounts of data. A straightforward conceptual model for describing this architecture is 10-80-10. The first 10% is neural, as the LLM helps interpret the user’s intent and translate it into a structured enterprise question. The middle 80% is where the enterprise foundation carries the reasoning that the LLM should not be asked to improvise. The system resolves enterprise context and can invoke plans, policies, constraints, decision rights, forecasting, causal analysis, simulation, optimization, rules, and solvers to evaluate scenarios and trade-offs. The final 10% is neural again, as the LLM explains the result in clear business language and makes the answer accessible for business decision-making.
As a result, neurosymbolic AI can make sense of large volumes of data at scale and recall complex patterns, while also incorporating structured reasoning and contextual knowledge that is unique to an organization. This grounds outcomes in reality and significantly reduces the chance of hallucinations.
Understanding the business question in context
A crucial component to incorporating neurosymbolic AI into a business’s larger AI strategy is a structured enterprise and decision-model foundation that can connect data, knowledge, plans, policies, constraints, and decision context. For example, the o9 Digital Brain’s Enterprise Knowledge Graph provides a structured and continuously updated semantic foundation that organizes enterprise information and connects it to decision models for capabilities such as forecasting, optimization, scenario analysis, and recommended actions. Better prompt construction is becoming irrelevant with each new frontier model. The system needs to understand the business question in context: the objective, relevant enterprise scope, current plans and actuals, policies and constraints, available decision levers, and who has authority to act. A neurosymbolic architecture allows much of that context to come from the enterprise model itself rather than requiring the user to recreate it in every prompt.
Source credits: