Enabling Ethical AI: A case study in using Ontological Context for Justified Agentic AI Decisions

Enabling Ethical AI: A case study in using Ontological Context for Justified Agentic AI Decisions Primary Authors: Liam McGee 1, James Harvey 1, Lucy Cull 1 Contributors: Andreas Vermeulen 2, Bart-Floris Visscher 2 Reviewers and Editors: Andy Corbett 3, Arielle Bennett 3, Malvika Sharan 3 Affiliations: 1. Kaiasm Ltd, London, UK 2. Avantra, Richmond, UK 3. The Alan Turing Institute, London, UK The version-01 of this article has been uploaded with a separate file with appendices on Zenodo under DOI: https://doi.org/10.5281/zenodo.17809907. Abstract Agentic AI systems, software agents with autonomy, decision-making ability, and adaptability, are increasingly used to execute complex tasks on behalf of organisations. Most such systems rely on Large Language Models (LLMs), whose broad semantic capabilities enable powerful language processing but lack explicit, institution-specific grounding. In enterprises, data rarely comes with an inspectable semantic layer, and constructing one typically requires labour-intensive “data archaeology”: cleaning, modelling, and curating knowledge into ontologies, taxonomies, and other formal structures. At the same time, explainability methods such as saliency maps expose an “interpretability gap”: they highlight what the model attends to but not why, leaving decision processes opaque. In this preprint, we present a case study, developed by Kaiasm and Avantra AI through their work with The Turing Way Practitioners Hub, a forum developed under the InnovateUK BridgeAI program. This study presents a collaborative human-AI approach to building an inspectable semantic layer for Agentic AI. AI agents first propose candidate knowledge structures from diverse data sources; domain experts then validate, correct, and extend these structures, with their feedback used to improve subsequent models. Authors show how this process captures tacit institutional knowledge, improves response quality and efficiency, and mitigates institutional amnesia. We argue for a shift from post-hoc explanation to justifiable Agentic AI, where decisions are grounded in explicit, inspectable evidence and reasoning accessible to both experts and non-specialists. Introduction Agentic AI systems are software systems that leverage Artificial Intelligence to exhibit some degree of autonomy, decision-making capabilities, and adaptability, allowing the systems to pursue complex goals and execute tasks on behalf of users with limited human supervision. Companies across all 1 sectors are actively developing Agentic AI systems to improve their institutional operations. The broadly used technologies often rely on language processing solutions offered by Large Language Models (LLMs) [1]. LLMs are built with a generalised semantic context, deriving a broad understanding of language and relationships from their training data. This enables powerful pattern-matching for language processing, often without explicit instructions. However, this inherent generality poses a significant challenge for enterprise-level Agentic AI architectures. Enterprise data frequently lacks a crucial, inspectable semantic layer that provides specific institutional context, thus severely limiting Agentic AI's real-world applicability. Creating a semantic layer traditionally involves extensive manual "data archaeology" and curation: cleaning datasets, constructing formal knowledge structures like ontologies and taxonomies, and organising information for better availability. Adding to this challenge are the current AI explainability methods, such as saliency maps [2], which exhibit an "interpretability gap"; they can highlight the ‘salient’ features of an input that a model focuses on, but fail to explain the reasoning behind it. This presents the urgent need for Agentic AI decisions to be inherently justifiable, fully inspectable, and easily accessible to both experts and non-specialists. The sheer effort and insufficient tooling for this process hinder an organisation's ability to effectively deploy Agentic AI. An improved solution lies in continuous human-AI collaboration. This involves AI agents generating initial knowledge structure candidates from various information sources, which human experts then validate and enrich. These human-derived corrections, in turn, train subsequent AI models, offering key benefits: 1) bringing tacit and contextual knowledge into a machine readable format; 2) substantially reducing model training cost and improving response quality at lower energy use [3]; 3) preventing knowledge loss (institutional amnesia); and 4) improving the capabilities of Agentic AI. Kaiasm and Avantra AI teams collaborated to investigate this challenge through the enhancement and monitoring of inspectable Agentic AI systems by making knowledge orchestration accessible to their human operators. In their joint work, they tested whether the resultant semantic context improved accuracy, coherence and relevance of model responses. This also created an opportunity for the teams to explore the explainability of AI models, a challenge that currently remains unaddressed by existing approaches such as saliency maps, where an “interpretability gap of explainability methods relies on humans to decide what a given explanation might mean” [4]. This preprint describe a novel collaborative approach to explainability founded on “justification”, ensuring agentic AI decisions can be justified in terms of evidence and reasoning in a way inspectable before and after a decision is enacted, whilst broadening people’s access to formalised underlying knowledge foundations, without requiring highly specialised skill sets. A Hybrid Approach for Adaptability and Explainability Neural networks, including those powering LLMs, excel at pattern recognition and learn from examples, but operate as "black boxes" with limited interpretability. Symbolic AI, conversely, excels at logical reasoning with explicit rules but struggles with unstructured data variability [5],[6],[7]. Neural networks and Symbolic AI are two complementary AI paradigms that can be combined to build neuro-symbolic architecture, creating systems that are both adaptive and explainable. This architecture offers three essential benefits: 1. Pattern recognition with reasoning: Neural components identify patterns in enterprise data while symbolic components provide the logical framework to interpret these patterns meaningfully. 2 2. Explainability by design: Each decision maintains a traceable, justifiable path, a novel approach to addressing the "explainability crisis" plaguing many AI systems. 3. Adaptability with guardrails: The system learns from new data while maintaining consistency with established knowledge structures. Neuro-Symbolic AI in Practice Kaiasm and Avantra have built an integrated technology that combines the capabilities of two independent platforms: 1) OntoKai, a software tool for mapping, visualising, translating and merging detailed knowledge representations such as ontologies and data models. OntoKai is a proprietary platform of Kaiasm Ltd. 1) Avantra AIR, an AI-powered IT Operations software, a proprietary platform of Avantra Services Ltd. This combination can be implemented with conventional Enterprise Resource Planning (ERP) systems (such as SAP systems). This initiative serves to bridge the contextual disparity between enterprise data, artificial intelligence, and human operators. OntoKai functions as a dedicated platform for knowledge orchestration and visualisation for human users, while Avantra AIR extends agentic AI capabilities within the domain of ERP systems. This symbiotic integration establishes a neuro-symbolic architecture, resulting in an AI system with awareness of institutional context encompassing rules, facts, artefacts, probabilities, and processes. This allows the human operators to inspect the system and derive justification for its decisions, whether for internal verification, human-in-the-loop validation, or retrospective auditing. Research shows that context-enriched prompting, informed by structured ontologies from systems like OntoKai, significantly improves AI response quality in experimental contexts across logic and reasoning, learning and inference, knowledge and representation, explainability and trustworthiness [8],[9]. Kaiasm and Avantra to apply this to an as-live operational system to determine the practical commercial value of incorporating formal knowledge structures into agentic AI systems, together with an easy-to understand human readable presentation of agentic decisions and actions that can be used as a governance mechanism to maintain the Agentic system under human oversight. Translating Organisational Knowledge into Structured Information Using OntoKai OntoKai is a general-purpose knowledge orchestration system with extensive ontology management and semantic classification capabilities. It can serve as the knowledge infrastructure for hybrid neuro-symbolic AI systems. It provides an environment in which data can be encoded in both human and machine-readable formats, facilitating the cleaning, organisation, and governance of arbitrary business knowledge and metadata into linked, structured information. OntoKai also makes knowledge accessible to humans through visualisation tools, while maintaining open, standards-compliant data structures suitable for integration with a range of storage and data management systems. In Figure 1, a visualisation for “shipping container” has been provided. Shipping containers can be linked with information such as their physical location, loading unit, shipping details, type and dimensions, contents (types of packages), shipping costs, and relevant legal and financial requirements. OntoKai's merged architecture can logically represent all this information. This merging capability bridges the gap between disparate data sources, facilitating human understanding and data consultation while preserving a machine-readable format. 3 Figure 1: A screenshot from OntoKai platform where the example information about “shipping containers” have been visualised. OntoKai has been deployed in several large-scale operational environments and projects, including: ● National Highways as a Taxonomy & Ontology service: OntoKai provides a visualisation layer across National Highways' complex data landscape, revealing where entity classes are represented across multiple systems, standards and processes. This improves data and knowledge governance by making inconsistencies visible to both specialists and non-specialists, facilitating collaborative review processes, and supporting the resolution of definitional conflicts. By integrating with existing tools, OntoKai is contributing to National Highways' progression towards becoming a data-first organisation with improved decision-making through better data interoperability. ● Howdens Joinery as a discovery capture and knowledge integration environment: At Howdens, OntoKai served as the central knowledge repository for website performance optimisation, creating a unified knowledge graph that exposed previously hidden interdependencies between organisational silos. This enabled teams to identify bottlenecks, prioritise improvements based on evidence, and embed performance awareness across departments. The structured knowledge representation allowed disparate enterprise data to be transformed into actionable insights, while remaining accessible to stakeholders across the business. Additionally, a PoC/PoV has been created showing how the knowledge repository can be queried using an LLM as a semantically dense part of the prompt context. Improving Observability of Enterprise Ecosystem Using Avantra AIR Avantra AIR is an intelligent operations platform designed to support system resilience, automate incident management, and improve the observability of complex enterprise ecosystems. It operates by embedding predictive analytics and automated remediation into critical systems, such as SAP, identifying emerging issues and initiating preventative actions without requiring continuous human intervention. Avantra AIR has already been applied successfully in a range of operational contexts, including: 4 ● Howdens Joinery as an AIOps management solution for SAP ecosystem: Avantra AIR was used to monitor Howdens' SAP environment, enabling early detection of potential failures and automating recovery processes to minimise disruption to business operations. ● Chemical Manufacturing AIOps: In chemical manufacturing, Avantra AIR enhanced system stability by reducing reliance on manual monitoring and intervention, contributing to improved operational continuity. ● Motors, gear units, gear motors Manufacturing AIOps: Within manufacturing environments, Avantra AIR identified emerging faults before they developed into system failures, supporting higher equipment availability and reliability. ● Managed Service Provider AIOps: In managed service provider contexts, Avantra AIR facilitated the automation of system monitoring and maintenance, allowing service teams to focus on higher-value activities while maintaining consistent system performance. A Case for Hybrid Architecture OntoKai integration with AI architecture offers a neuro-symbolic system, which involves three improvement cycles are the following: 1. First cycle - Knowledge Graph: Human-led training and feedback on AI suggestions enhance the encoding of organizational information, empowering staff. 2. Second cycle - Drawing Insights: Contextualises and generates Human-led contextualization of data, enables AI systems to generate better insights. 3. Third cycle - Governance: Human instructed AI with encoded information are better at justifying decisions, particularly in areas like regulatory compliance. Figure 2: The three improvement cycles for the AI system. 5 In the first improvement cycle, existing data and institutional knowledge is brought into OntoKai as knowledge models, which are cleaned, enriched and validated by domain experts across the organisation. Differences in perspective or terms are either retained and clarified, or resolved. General LLM models are also used to ingest raw documents and datasets and turn them into candidate knowledge models in OntoKai. These go through the same audited process of improvement, validation and publication, informing both the LLM as a target training set as well as enriching human-generated models. Here the workflow shifts from ‘human knowing’ to ‘model tuning’. OntoKai can be used to directly re-weight concepts within the LLM, vastly reducing model tuning time, meaning less energy usage and lower cost. A second improvement cycle is based around human queries of the system. Adding ontological/semantic context as part of a prompt context results in higher quality responses [10]. These better responses are part of the next input context for the AI while also leading to improved follow-up questions from the human operator, and so on. Better insights create better contexts and better questions. A third and final cycle evolves into the Agentic Justification Loop. The models in OntoKai can be used by an Agentic AI, Avantra AIR, as inspectable instructions, allowing the agent to explain or justify its decisions through the model. Humans can inspect a decision before an action takes place or simply maintain it as an inspectable record of lower-risk actions. Collaborative Workflow of OntoKai and Avantra Using the REST API, a bridge between enterprise knowledge, enterprise data (held in Avantra’s AIR, structure exposed and corrected in OntoKai) and agentic AI (AIR auto generating reports and suggested actions using prompts from OntoKai) is created. The results involving a series of knowledge artefacts are stored in an OntoKai architecture via a series of three improvement cycles (Figure 2). The system’s workflow (Figure 3) starts by ingesting existing data and institutional knowledge into the OntoKai environment, used as initial knowledge models. These models then undergo a rigorous process of cleaning, enrichment, and validation by domain experts across the organisation. Discrepancies in perspectives or terminology are either retained and clarified or resolved to ensure consistency. Once validated, these knowledge models are published to other systems. This can take various forms, such as schemas for data warehouses, catalogues, formal supplier data requirements, or simply made accessible for general exploration. AI building blocks are then integrated into the process. A general Large Language Model (LLM) can automatically identify and process raw documents and datasets, transforming them into candidate knowledge models within OntoKai. These AI-generated models follow the same audited process of improvement, validation, and publication. This not only provides a target training set for the LLM but also enriches human-generated models. This initial integration represents the first of three improvement cycles, involving collaborative "humans in the loop" approach, which ensures that staff and domain experts work alongside AI to continually enhance and encode the institution's knowledge. 6 Figure 3: OntoKai workflow within Avantra’s ML/LLM architecture, leveraging the functions of a neuro-symbolic system. The second improvement cycle focuses on leveraging improved context for enhanced insights. OntoKai's outputs can be incorporated into an AI prompting context or even directly used to re-weight concepts within the LLM. This cycle significantly reduces model tuning time, potentially leading to reduced energy consumption and lower operational costs. Finally, the Agentic loop is introduced. The models within OntoKai serve as inspectable instructions for an Agentic AI. The agent is required to justify its decisions based on the information encoded in the OntoKai model. This allows humans to inspect decisions before actions are executed, or to maintain an inspectable record for lower-risk actions. This framework ensures the Agentic AI is explainable, controllable, responsible, and compliant with regulations. Knowledge Artefacts as Outputs Within the OntoKai environment, several knowledge artifacts are instantiated. We can explain this using shipping container data, for instance: 1. Human-derived taxonomy and conceptual model for shipping containers. Derived through an interview and discovery process, and captured as a formal concept architecture, as shown in Figure 4.1. 2. Human-merged knowledge models of conflicting knowledge about shipping containers, and arguments about them, as shown in Figure 4.2. 3. Agent-generated Process models showing agentic operations, as shown in Figure 4.3. 7 4. Agent-generated data architectures of the different layers of an agent-created lake-to-warehouse, which were then human-corrected/validated in OntoKai before being fed back to improve the agent’s creation process, as shown in Figure 4.4. 5. Data models for key concepts such as Time, as shown in Figure 4.5. 6. Merged data models to create common data models. These were then validated and corrected in OntoKai, such as how longitude with two different data types from different parts of the data lake (stored as both string and floating point), as shown in Figure 4.6. Agent-generated Systems Architectures are explored and validated by humans. Figure 4: Six types of human and agent-generated data in the context of the “shipping containers” example. Finally, a justification context was created in OntoKai which could use an LLM to justify decisions in terms of the knowledge held in the OntoKai architectures, as shown in Figure 5. 8 Figure 5: screenshot of an architecture in OntoKai showing justification capabilities using a Toulmin argumentation framework. OntoKai provided semantic context and data visualisation in a format understandable by both humans and computers, while Avantra AIR ingested information, suggested models and generated insights, all of which contribute to Agentic AI decision making. This creates an Agentic AI that is explainable, controllable, responsible. The system is designed to support compliance with the EU AI Act and developing UK AI regulatory requirements for risk management, human oversight and explainability. Justification: Improving Explainability for AI-Informed Decisions A wicked challenge of using complex AI is its explainability. We sought to approach it differently, using the philosophy of argumentation and associated analytic approach developed by Stephen Toulmin [11],[12] in the 1950s but still applied today as a means of ‘informal logic’, and further informed by philosopher of science Peter Lipton’s approach to ‘inference to the best explanation’ [13] as a descriptive account of how abductive reasoning is carried out in the sciences. Toulmin’s approach divides an argument into six parts: 1. Claim/decision - the truth claim or decision to be tested 2. Grounds/evidence/data - the statements of fact used to ground the claim. 3. Warrant - the chain of reasoning that sets out why the grounds support the claim, connecting the decision to the supporting evidence. 4. Backing - justification for trusting the warrant (e.g. scientific backing for the warrant) 5. Rebuttals - objections and counter examples. Identifies conditions under which the argument might not hold—it introduces exceptions, counterexamples, or alternative perspectives. “Even if your reasoning usually works, here’s a situation where it might not.” Can attack the grounds, the warrant or the claim, but usually attacks the warrant by setting out a hypothetical situation where the reasoning fails. 6. Qualifier/Exceptions - adjustments to the scope of the claim/decision to account for rebuttals. Lipton’s approach acknowledges that explanation consists of choices between competing explanations based on two factors: likelihood and loveliness. While likelihood concerns how probable the evidence is given a hypothesis, loveliness refers to the explanatory virtues that make an explanation better, not merely more likely. Loveliness covers: ● Explanatory power: How much does the explanation unify and account for the phenomena? ● Simplicity: Fewer assumptions or mechanisms are preferred. ● Coherence: Does it fit well with other established beliefs or theories? ● Depth: Does it reveal underlying mechanisms rather than superficial patterns? ● Breadth: Can it explain more phenomena beyond just the immediate data? This encoding of both Bayesian likelihood and the explanatory power encapsulated in loveliness can be used to inform the prompt to generate the Toulmin ‘warrant’, and to give a rationale for selecting between a set of competing claims that all explain the same set of evidence. A set of prompts were developed to provide the LLM with a human-asserted intent, for the LLM to set out the generated proposed steps for the agent (recorded in OntoKai), and to require the LLM to create grounds, warrants, backing and rebuttals, and resultant qualifiers, against a generated decision (again, recorded in OntoKai for audit/inspection). To the best of our knowledge, the justification approach discussed here is relatively under-explored. 9 However, comprehensive security architecture details, formal compliance assessments against specific regulatory frameworks, and detailed data protection protocols were beyond the scope of this case study. Enterprise deployment would require additional technical evaluation covering network security, data encryption, backup procedures, and formal risk assessment processes. Limitations This study demonstrates the effectiveness of contextual enhancement for large language model performance, but several methodological limitations should be acknowledged. The experimental design did not isolate the specific contribution of ontological structuring versus information quantity or relevance. Future research should include control conditions comparing structured ontological context against equivalent amounts of unstructured but relevant information to determine whether the observed benefits derive from semantic organisation itself or simply from the provision of additional domain-specific context. The response evaluation process used subjective 5-point rating scales, and inter-rater reliability measures were not reported. Future work would benefit from multiple independent raters and formal assessment of inter-rater agreement to strengthen the validity of subjective evaluations. Real-World Implications The integration of OntoKai and Avantra AIR offers a theoretically sound and empirically validated approach to building more robust, inspectable, and ethical agentic AI systems. Our findings have direct implications for enterprise AI implementations: Value of Semantic Context: Our research supports the theoretical claim that providing Large Language Models (LLMs) with ontology-derived context effectively bridges the semantic gap often found in enterprise data. This enables LLMs to generate more accurate, coherent, and useful outputs. Verification of Context-Enriched Prompting: Our work confirms that context-enriched prompting, informed by structured ontologies from systems like OntoKai, significantly enhances both the accuracy and coherence of AI responses. Illustrative Scenario: ACME Chemicals To illustrate how this methodology works in practice, consider this scenario involving a fictional chemical manufacturing company, ACME Chemicals: Initial Knowledge Mapping The project commences by analysing existing data models and process documentation, in this specific case, provided by an SAP system. The system then uses these inputs to generate preliminary ontologies. These initial models undergo validation by domain experts, who identify any gaps in process representation. Structured Knowledge Enhancement A systematic 12-step process is applied for each knowledge domain (e.g., chemical manufacturing processes) to enhance the structured knowledge: 16 1. Prompt an LLM with an OntoKai output and a new knowledge source. 2. Instruct the LLM to consider relationships between elements in the new source and existing entities within the architecture. 3. Where relationships already exist in the architecture, the LLM prioritises these. 4. The LLM identifies entity classes not explicitly present in either the knowledge source or the OntoKai prompt. 5. The LLM provides an archetypical individual of the class, complete with properties and rationale (e.g., Bird: sparrow. It flies, has feathers, builds nests in trees, and sings.). 6. The LLM offers up to three atypical individuals of the class, including properties and rationale (e.g., Bird: penguin. It has feathers and lays eggs, but it cannot fly, swims instead, and lives in aquatic environments.). 7. The LLM presents an exotypical individual of the class with properties and rationale (e.g., Not a bird: bat. It flies, has wings, and is warm-blooded, sharing many functional properties with birds, but it lacks feathers and lays no eggs.). 8. The LLM provides a definition for each entity class. 9. Export the data architecture from the LLM as OWL RDF/XML for input into OntoKai. 10. Human validation of the newly generated architecture. 11. Publish the new data architecture. 12. Merge the new architecture into the overarching knowledge universe. Operational Integration with Avantra AIR The validated knowledge models furnish semantic context for Avantra AIR, enabling the system to accurately interpret monitoring data from chemical processing systems. When anomalies are detected, the system can reason about potential causes using these rich knowledge models. Projected Results ● Before Implementation: System issues would trigger alarms, but required extensive human investigation to determine root causes. ● After Implementation: The system can automatically identify root causes (e.g., ID range issues) and initiate appropriate responses. Furthermore, the system can automate preventative actions based on its reasoning derived from the knowledge models. This hypothetical implementation was tested using increasingly detailed prompts about SAP systems, particularly focusing on logistics execution and delivery scenarios. The testing confirmed that when AI systems were provided with ontological context about specific operational elements (such as "Dispatching Bay 17") and the financial implications of downtime (potential losses of £2.4 million), their responses became significantly more coherent and actionable. Hypothetical Scenario: System 3 Has Lost Contact To further demonstrate the potential difference this approach could make, consider this illustrative scenario: Without AIR/OntoKai: ● An alarm goes off indicating Server 3 has lost contact. ● The incident is triaged and classified as a P1 (highest priority) incident. ● IT staff must determine who to contact for resolution. 17 ● Time continues to pass as the situation deteriorates. ● Eventually, a decision must be made to either: ○ Ship products outside of Standard Operating Procedures (with potential regulatory consequences). ○ Halt operations completely (with significant financial impact). With AIR/OntoKai: ● An alarm goes off indicating Server 3 has lost contact. ● The system interprets the alarm and immediately identifies this as an ID range issue. ● The semantic context of "Server 3 has lost contact" is understood in relation to business operations. ● Preventative actions are automatically initiated based on previous knowledge. ● Automated reporting keeps stakeholders informed. ● The problem is resolved before significant operational impact occurs. Ultimately, once the system is fully implemented, such incidents might be prevented entirely through proactive monitoring and intervention. This demonstrates how the iterative neuro-symbolic loop methodology not only creates more accurate knowledge models but directly translates into tangible business benefits through enhanced issue detection, faster resolution, and proactive prevention. Challenges As with any new technology there are a number of challenges in implementation. 1. Data availability: Complex raw data is crucial for optimal functionality of agentic AI systems, but in real-world scenarios access may often be limited. Concerns such as legal regulations, data security and organisational policies may restrict access to some raw data, and this must be considered on an organisation-by-organisation basis when planning implementation. This also highlights the need to ensure that data within the AI system remains secure: the system must be secure enough to meet the needs of organisations with highly protected data. It must be kept in mind that with greater system complexity comes increased cybersecurity threats. 2. Consensus: When creating AI systems that make important decisions there must be consensus on ethical standards; however, obtaining that consensus is often not straightforward where there are multiple stakeholders and differing views on governance may lead to conflict. This may be further complicated by conflicting business needs and ethical constraints. There is also a need to enforce those ethical standards; in short, a significant amount of governance is required to ensure that these agentic AI systems remain ethical. Any tooling needs to allow for multiple conflicting positions, definitions and intentions, and allow audit of the evolution of policies, processes, regulation and stated intent. 3. Transparency: Human validation is fundamental to this methodology and in order to validate knowledge produced by AI, the provenance of the knowledge must be clear. Lack of transparency is likely to inhibit adoption, particularly by non-technical stakeholders who may not know all of the context behind the technology. Raw data does not, by itself, directly translate to actionable business insights, and the transformation process has the potential to introduce errors, biases or inefficiencies. It is therefore important to reduce the steepness of the learning curve and focus on general usability of any governance tooling such that they are understandable and interpretable to non-technical stakeholders: greater understanding leads to greater adoption, and greater adoption (and ‘many eyes’) will lead to better outcomes. 4. Methodological choice: The choice of hybrid neuro-symbolic processing methodologies is one such question to consider. To date, there is a lack of consensus on which are the best methodologies for Explainability in this area, particularly due to the challenges of blending 18 symbolic reasoning efficiently with neural networks for clear trustworthiness [14]. For instance, symbolic AI excels in logical reasoning but struggles with variability in raw data, whereas neural networks excel at pattern learning but are criticised for their "black-box" nature. Hybrid systems aim to merge these interpretative capabilities with pattern recognition strengths, but determining the optimal integration approach remains an ongoing area of research and development. 5. Scalability: The scalability of processing is another key challenge to overcome. Traditional AI systems often face bottlenecks when handling massive datasets, particularly in real-time scenarios where rapid analysis is required. As these systems grow and handle increasingly large volumes of data, the computational requirements become more intensive. There is a risk that, in larger systems, AI agents may spend more time communicating with each other than completing their assigned tasks, similar to inefficient meetings with too many participants. ‘Hyperscaling’ approaches - using distributed systems for massive parallelisation - offer a potential solution to ensure scalability without compromising speed or accuracy, but the question of what truly defines "hyperscaled" in the justification context remains somewhat ambiguous and depends on the specific organisational requirements. Tooling such as OntoKai needs to work effectively with hyperscalable applications such as AvantraAIR, which will require a synchronising of scale - how to identify which actions can be justified by the system itself without a human in the loop, and how to abstract the justification outputs to enable large sets of similar justifications to be simultaneously reviewed by a single human. 6. Coordination: When multiple AI agents work together in a 'swarm', coordination challenges emerge. Distributed processing frameworks struggle with establishing effective consensus mechanisms among these agents. Different swarm configurations may require different approaches to agreement, raising the question of which consensus model is most appropriate for each specific scenario. This leads to considerations about the optimal level of tolerance within these agentic swarms - if consensus requirements are too strict, the system may become unresponsive; if too lenient, results may be inconsistent. Finding the appropriate balance is crucial for ensuring these multi-agent systems function effectively in enterprise environments. While systems such as Avantra AIR are providing potential solutions to these problems, integrations with knowledge orchestration platforms such as OntoKai need to consider how to make the balancing decisions explicit and governable by human operators. 7. Dynamic requirements: Businesses must contend with dynamic and conflicting requirements as they implement these hybrid systems. As organisations evolve, their technological needs naturally change, requiring AI systems to adapt accordingly. This creates particular challenges for hybrid architectures where neural networks grow increasingly complex while symbolic reasoning components demand more extensive sets of symbols, logic, and rules to function properly. These competing technical requirements often create practical tensions when implementing and maintaining such systems. Organisations must therefore carefully balance these elements while ensuring their AI infrastructure can evolve alongside changing business priorities. Next Steps Building on the success and challenges outlined in this case study, next steps to further develop these technologies have been identified: 1. Expanding the data sources: by diversifying the data inputs from various business systems and external sources, the hybrid architecture can develop more comprehensive contextual understanding, further enhancing the accuracy and coherence of AI responses in complex business scenarios. 2. Expanding the neuro-symbolic AI-powered agentic loop: applying the neuro-symbolic approach to additional business domains will help tackle the dynamic and conflicting 19 requirements challenge, allowing more organisations to benefit from inspectable and trustworthy AI systems even as their business needs evolve. 3. Fine-tuning the neuro-symbolic AI-powered agentic loop: optimising the methodology will improve the efficiency of the human-in-the-loop validation processes, addressing the scalability challenges identified while enhancing the system's ability to manage complex semantic relationships across enterprise data. 4. Adding in an Evidence Framework Approach based on work done at DSTL [15] to determine ontological properties of the evidence. This would cover the quality of the evidence and the extent of its support for the warrant. It is based on a validity dimension, subdivided into equally weighted scales for face, criterion, construct and content validity; and an evidence profile based on equally weighted scales for comprehensiveness, relevance, objectivity, quantity and consistency. It harmonises with the methodology set out in the HMG Aqua Book on analysis and the Magenta book on evaluation. Conclusion The work described in this preprint offers a case study with examples demonstrating the impact of symbolic architectures in bridging the semantic gap between enterprise data and agentic AI. By fostering hybrid neuro-symbolic systems, we have achieved a balance between adaptability and explainability, addressing long-standing challenges in AI deployment within complex organisational environments. Our empirical findings, derived from a real-world use case, show that context-enriched prompting, facilitated by structured ontologies, enhances AI response quality and improves both accuracy and coherence. The OntoKai tooling provides an auditable continuous improvement cycle, assisting human experts to validate and refine knowledge models, thereby safeguarding against institutional amnesia through trustworthy AI systems. The synergy between OntoKai, providing the critical semantic layer for both human and machine accessibility, and Avantra’s AIR architecture, which successfully leverages this context for agentic decision-making, offers practical viability and future potential of this approach. As described in this paper, we advocate for 'justification' approach as an effective means of explainability and human governance over the knowledge models underlying neuro-symbolic architectures. Our collaborative and integrated approach makes AI systems more robust, inspectable, controllable, and ethical, particularly as they are scaled to meet the demands of enterprise-level applications. Conflicts of Interest All authors declare that they have no conflicts of interest. Bibliography [1] Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A. and Agarwal, S., 2020. Language models are few-shot learners. Advances in neural information processing systems, 33, pp.1877-1901. [2] Müller, R., 2025. How explainable AI affects human performance: A systematic review of the behavioural consequences of saliency maps. International Journal of Human–Computer Interaction, 41(4), pp.2020-2051. [3] Baldazzi, T., Bellomarini, L., Ceri, S., Colombo, A., Gentili, A. and Sallinger, E., 2023, September. Fine-tuning large enterprise language models via ontological reasoning. In 20 International Joint Conference on Rules and Reasoning (pp. 86-94). Cham: Springer Nature Switzerland. [4] Ghassemi, M., Oakden-Rayner, L., & Beam, A. L. (2021). The false hope of current approaches to explainable artificial intelligence in health care. The Lancet. Digital health, 3(11), e745–e750. https://doi.org/10.1016/S2589-7500(21)00208-9 [5] ŞAHiN, E., Arslan, N.N. & Özdemir, D. Unlocking the black box: an in-depth review on interpretability, explainability, and reliability in deep learning. Neural Comput & Applic 37, 859–965 (2025). https://doi.org/10.1007/s00521-024-10437-2 [6] Bhuyan, B.P., Ramdane-Cherif, A., Tomar, R. et al. Neuro-symbolic artificial intelligence: a survey. Neural Comput & Applic 36, 12809–12844 (2024). https://doi.org/10.1007/s00521-024-09960-z [7] (2021) Ebrahimi, M., Eberhart, A., Bianchi, F. et al. Towards bridging the neuro-symbolic gap: deep deductive reasoners. Appl Intell 51, 6326–6348. https://doi.org/10.1007/s10489-020-02165-6 [8] Lu, Z., Afridi, I., Kang, H.J. et al. Surveying neuro-symbolic approaches for reliable artificial intelligence of things. J Reliable Intell Environ 10, 257–279 (2024). https://doi.org/10.1007/s40860-024-00231-1 [9] Colelough, B. & Regli, W. (2025). Neuro-Symbolic AI in 2024: A Systematic Review. https://arxiv.org/pdf/2501.05435v1 [10] P. Wang, J. Karigiannis, R.X. Gao (2024), pp. 361-364, Ontology-integrated tuning of large language model for intelligent maintenance, CIRP Ann, 73 (1), 10.1016/j.cirp.2024.04.012 [11] Gabriel, V. O., Panisson, A. R., Bordini, R. H., Adamatti, D. F., & Billa, C. Z. (2020). Reasoning in BDI agents using Toulmin's argumentation model. Theoretical Computer Science, 805, 76. [12] Liu, D., Xiong, M. Keeping balance between loyalty and modification: a Toulminian model as analytical framework. Humanit Soc Sci Commun 11, 639 (2024). https://doi.org/10.1057/s41599-024-03151-w [13] LIPTON P., (2004) Inference to the Best Explanation, Routledge, 2nd Edition, ISBN 0-415-24203-7 [14] Colelough, B.C. and Regli, W., 2025. Neuro-symbolic AI in 2024: A systematic review. arXiv preprint arXiv:2501.05435. https://arxiv.org/abs/2501.05435 [15] Pearce P. A, (2017 , Technical Overview of the Evidence Framework Approach - ISMOR. Defence Science & Technology Laboratory (DSTL) 21