Neuro-symbolic AI: Towards Precise, Governed, Explainable Reasoning

General-purpose language models can draw on broad knowledge, but they have no automatic knowledge of how your organisation defines a client or customer, calculates a metric or handles an exception. That context is often located across separate systems, documents and teams. In organisations that have grown organically or through acquisitions, even familiar terms may carry different meanings in different systems.

Those differences won’t be resolved by more data alone. AI also needs agreed definitions, explicit relationships and reliable sources. A semantic layer captures that shared understanding in a form that can be used by people and AI systems alike.

Neuro-symbolic AI brings neural learning and symbolic knowledge together. Neural models recognise patterns and process language. Symbolic structures explicitly represent concepts, relationships and rules. In an enterprise application, the combination creates an intermediate layer between data and decision.

The Antal-MakoLab report examines how a semantic layer can support this approach. Its value depends on how it is built and connected to the model. Documenting definitions is only the beginning. The system must then use them to retrieve information, apply rules and check results. However, while this can improve traceability, it doesn’t do away with the need for testing and oversight.

The research cited in the report highlights the demand for greater transparency. In a survey by MIT Technology Review Insights, produced in association with EY Studio+, 41% of the responding executives and business leaders in regulated industries identified transparency and explainability as a key concern when using AI for customer experience.¹ Deloitte’s 2026 State of AI in the Enterprise report found that 46% of the companies surveyed identified model quality, consistency and explainability among the AI risks they were most concerned about.²

For decision-makers, the practical questions are straightforward. Which sources did the system use? Which definition did it apply? Can the result be checked against the underlying data and business rules? Although a symbolic layer can help make those checks possible, it does not, in itself, prove regulatory compliance or reveal every internal computation of a neural model.

The foundations are often less ready than AI ambitions suggest. In a study by Harvard Business Review Analytic Services, sponsored by Cloudera, only 7% of the respondents said their organisation’s data was completely ready for AI adoption. Meanwhile, 73% agreed that processing and preparing data for AI had been challenging.³

Neuro-symbolic AI argues for starting with a focused business problem, going on to identify the concepts, sources and exceptions that determine whether an answer is useful and then formalising them with the people who understand the work. A finance metric, for example, needs not only access to the relevant figures, but also an agreed calculation and reporting period.

This is a shared business and technology task. Domain experts define meaning and technical teams make those definitions usable in the system. Both need a way to maintain them as products, processes and rules change.

A plausible response is not evidence of a correct one. Language-model outputs are the result of probability, context and available data, so reliable use requires explicit evaluation criteria and repeatable checks.

Neuro-symbolic AI explores what should be established before deployment, what a correct response looks like for representative cases, which sources should be taken into account, what errors are unacceptable and what level of certainty a given application requires. Those checks need to evolve with the knowledge base. The goal is a process that can detect when an answer no longer meets an organisation’s standards.

In Capgemini’s research, 74% of the surveyed executives expected the benefits of adding human oversight to AI agent-driven tasks to outweigh the costs over the subsequent one to three years, assuming the solutions become more accurate and reliable.⁴ This is an expectation, rather than a measured return, but it underlines the role of oversight in plans for adoption.

Building and maintaining a semantic layer takes expert time. A realistic initial scope helps an organisation test whether that investment improves a specific decision or process before extending the approach. Neuro-symbolic AI considers these operating questions alongside the technical architecture.

Download Neuro-symbolic AI: Towards Precise, Governed, Explainable Reasoning to explore how neural models and symbolic knowledge work together, where the approach can add value and how to assess the data, the evaluation process and the oversight needed for a first use case.

Book a working session with MakoLab to review where AI informs your decisions, identify the definitions and rules those decisions depend on and discuss a practical starting point.

Neuro-symbolic AI: Towards Precise, Governed, Explainable Reasoning is the fourth publication in a research series by Antal in partnership with MakoLab. Designed for IT leaders and decision-makers, the programme maps the operational reality behind enterprise AI adoption.