{"slug": "ontology-in-the-age-of-ai", "title": "Ontology in the Age of AI", "summary": "A developer built an ontology-driven, validated prediction model orchestrated by an AI agent to answer accommodation-pricing questions for Australia's aged care industry without fabricating data. The system was reportedly used by aged care providers to produce evidence supporting prices above the A$789,686 threshold, a legal requirement in the sector. The approach pairs formal ontology design—classes, slots and rules—with agentic AI to deliver transparent, verifiable answers for regulated audiences such as boards, executives and government regulators.", "body_md": "As the capability of agentic AI is increasing significantly, how can it assist with offering answers for audiences that seek transparency, validation and evidence? Such audiences may include boards, executives and decision makers in organisations that operate in highly regulated sectors such as aged care, health, banking and finance, childcare and education. Such answers may also be sought by external regulators such as state and federal government departments and agencies.\n\nThis article is based on my experience and journey of finding and communicating those very answers. The focus will be on how ontology and validated models, orchestrated by an AI agent, can produce transparent answers. The process included building an ontology that led to a validated prediction model, which an AI agent uses to answer questions on accommodation prices without fabricating or inventing anything.\n\nFollowing this process, I was able to produce those answers for the aged care industry of Australia. As I was not the main applicant, I cannot verify exactly how the information was used, but I was advised that such answers assisted aged care providers in meeting a legal requirement of producing evidence to support prices over the A$789,686 threshold.\n\nThis piece will start with the definition of ontology, followed by a simple example, ontology in regulated sectors, ontology design journey, ontology and agentic AI, where and how to use AI agents and key takeaways.\n\n## What is ontology\n\nThomas R. Gruber (1993) describes ontology as ‘an explicit specification of a conceptualization. The term is borrowed from philosophy, where an ontology is a systematic account of Existence.’\n\nGruber argues that each term in an ontology has three parts:\n\n- a name;\n- a plain-language definition of what it means;\n- formal rules limiting how it can be used.\n\nThough there may be many definitions of ontology, Noy & McGuinness (2001) define ontology as ‘a formal explicit description of concepts in a domain of discourse (classes (sometimes called concepts)), properties of each concept describing various features and attributes of the concept (slots (sometimes called roles or properties)), and restrictions on slots (facets (sometimes called role restrictions)). An ontology together with a set of individual instances of classes constitutes a knowledge base.’\n\n## A simple example\n\nNow let’s try to understand this with an example. Suppose a tech electronics store wants to achieve a knowledge base; what process does it need to follow to get there?\n\n- The first step would be to define the ‘Terms’ (name, meaning and rules). These may include: \n  - Customers: people who have made at least one purchase\n  - Location: the suburb where the customer lives\n  - Product: an item sold by the store\n  - Laptop: an item under the product category ‘Computers’\n  - Product category: a group of products sold at the store\n  - Purchase date: the date when the transaction was completed\n  - Purchase: a completed transaction/sale\n  - Price: the amount paid by the customer to obtain a product\n  - Lives in: connects a customer with their location\n- The second step would be to describe the classes/concepts and arrange classes and subclasses in a hierarchical order. In our example, a class would be ‘Computers’, a subclass would be ‘Laptop’, and further subclasses of ‘Laptop’ may be ‘Office laptop’ and ‘Gaming laptop’. Other classes include ‘Customers’, ‘Location’, ‘Purchase’ and ‘Product’.\n- The third step would be to define slots. In our example, slots may be ‘Price’ and ‘Date’, associated with the class ‘Purchase’. The relationship slots include ‘makes’ between ‘Customer’ and ‘Purchase’ and ‘of’ between ‘Purchase’ and ‘Product’.\n- The fourth step would be to define the facets of the slots (‘rules’). Some of the rules may include: \n  - Price: value in Australian dollars and only one value is assigned per transaction (sale)\n  - Purchase date: a numeric date using the DMY (Day–Month–Year) structure and only one value is assigned to each transaction (sale)\n\nLinking classes with the help of slots restricted by defined rules leads to designing an ontology. However, this is not yet a knowledge base; we will need to link the ontology with individual instances to achieve a knowledge base.\n\n## Ontology in regulated sectors\n\nNow the question is why we need to understand ontology and, particularly, its application in more regulated sectors such as banking, health and education.\n\nTo better address the question, let’s use a real example that was developed and implemented in the aged care/healthcare sectors in Australia.\n\nAs in many western economies, the healthcare industry must comply with well-defined regulations to provide safe and quality care to the elderly. Regulations apply to other areas of aged care and healthcare organisations, such as the price determination process.\n\nAccommodation prices (the price of a room customers pay to move into a home) in the aged care industry over the amount of A$789,686 require approval from a government agency. In order to obtain approval for prices over A$789,686, part of the process is to provide logic/reasons backed by evidence to support the application, so transparency is paramount to justify the ask.\n\nTo address that, I developed a pricing model that predicted prices in line with the market situation. However, price prediction was the destination, and to ensure I did not get lost along the way, the ontology allowed me to map out the journey and offered clear directions to follow.\n\nThe process below demonstrates that journey; this is by no means the best or the only process to follow to achieve actionable insights based on predictions without compromising the validation of the predictions or the transparency of methods.\n\nI am going to share my experience in the light of concepts defined by Noy & McGuinness (2001).\n\n## Ontology design journey\n\n- The first step was to define ‘Terms’. The defined ‘Terms’ included: \n  - Customers: people who are entering the aged care facility\n  - Room size: the internal room area represented in square metres excluding bathrooms and external areas such as balconies\n  - Buys: links ‘Customer’ and ‘Room’\n  - Location: where the customer currently lives (origin) and where the aged care facilities are located\n  - Room type: describes room characteristics from the perspective of the number of occupants\n  - Price: the amount paid by customers for accommodation in Australian dollars\n  - Stock condition: highlights the building age and accommodation state \n    - New: brand-new aged care facilities developed within the last two years\n    - Newer: facilities that were developed or significantly refurbished in the last 2–10 years\n    - Older: facilities that were developed or significantly refurbished between 10 and 20 years ago\n    - Old: facilities that were not developed or significantly refurbished in the last 20 years\n  - Financial capacity: the ability to pay and afford the accommodation costs \n    - Income: average yearly income of customers over the age of 65 in the catchment\n    - Median house price: median house price of the origin suburb\n    - Age pension ratios: the percentage of people over the age of 65 who receive the pension in the Statistical Area 2 (SA2) the customers are coming from\n- Now that we have defined our ‘Terms’, the next step is to define classes and subclasses in hierarchical order as identified above. Classes and subclasses are represented below in a diagram.\n\n- From our process above, the third step is to define slots. ‘Income’, ‘Median house price’ and ‘Age pension ratios’ are associated with ‘Financial capacity’. ‘Room size’ and ‘Price’ slots are associated with ‘Room’. There are relationship slots in our ontology structure including: \n  - ‘buys’ between ‘Customer’ and ‘Room’\n  - ‘lives in’ between ‘Customer’ and ‘Location’\n  - ‘is in’ between ‘Room’ and ‘Facility’\n  - ‘draws from’ between ‘Facility’ and ‘Catchment’\n  - ‘made of’ between ‘Catchment’ and ‘Location’\n  - ‘located in’ between ‘Facility’ and ‘Location’\n- Following our order of steps, the fourth would be defining values and setting rules for slots. These include: \n  - Income: average yearly income of customers over the age of 65 in the catchment\n  - Median house price: median house price in the area the customers are coming from. The median house price should consider completed house transactions only.\n  - Age pension ratios: the percentage of people over the age of 65 who receive the pension in the areas the customers are coming from\n  - Room size: room size refers to the total room size per person. For single ensuites, this would be the total room size, and for shared rooms the total size will be divided by the number of people sharing that room to determine room size per person.\n\nNow let’s look at the designed ontology, linking classes and slots that comply with the defined rules, and how the ontology becomes a knowledge base based on Noy & McGuinness (2001).\n\nTo better understand this process shown in the diagram above, let’s work on the same example but this time for an instance the model has not yet seen. For an aged care facility, to set a price for a new room (new instance) or the resale of existing rooms, the model, trained on the existing knowledge base, can predict prices based on classes and by filling in values for slots.\n\n## Ontology and agentic AI\n\nMany organisations and businesses may be tempted to develop and run AI agents to train a model on a knowledge base and make predictions. Such an approach could work if a business is comfortable using a black-box type prediction, any inaccuracies are not detrimental to the business and they have no obligation to be transparent about their methodology to make a prediction.\n\nHowever, organisations and businesses that operate in regulated environments have an obligation to be transparent to the regulators, and there may be some risks involved with such an approach.\n\nAs mentioned above, in Australia, any room price above A$789,686 in the aged care industry requires a government agency’s approval. Any predictions that are not based on a transparent and validated process may not be able to meet the regulatory requirement of providing logic and evidence to support the claim that their rooms are worth more than A$789,686. This could be due to some potential issues with consistency and validation. The predetermined ontology rules enable organisations to develop models that produce consistent and transparent results.\n\nAs most of the work is being done at the back end, decision makers may not be able to check exactly how an AI agent reached a certain prediction unless all of the statistical models are tested beforehand by data scientists and domain knowledge experts, and the AI agent is just running the same models, or the same experts work with an AI agent to cross-check the statistical models it is generating.\n\nThere is also another risk in exposing sensitive data to AI agents that are running on external LLMs and in external environments.\n\nSuppose an organisation has developed a novel way to establish an ontology and knowledge base, novel data cleaning approaches and novel statistical ensemble models, and the entire information is handed to an AI agent set up on a secure third-party platform such as AWS Bedrock. Is the sensitive information safe, and is there no risk of data leakage and compromised IP? Third-party platforms are not misleading about their data handling and training claims, but zero data retention (ZDR) still cannot be guaranteed due to situations that may be beyond third parties’ control, such as:\n\n- If servers are running in other countries, depending on the law, the countries can seek to access information on the national security risks basis\n- Highly sophisticated cyber attacks\n- Mandatory data sharing conditions imposed by AI companies themselves to access their models, such as Anthropic’s Fable 5.1\n\nThe cost element should not be ignored in all this. Many healthcare facilities, whether private, not-for-profit or government owned, have to be disciplined with their costs due to other significant opex they face, such as staff costs, equipment management and utility costs. If an AI agent is running on a token-based service, there is no certainty around how much exactly it will cost them to constantly train the models and make predictions. Boards in regulated environments generally do not respond positively to uncertain costs, as budgets have to be developed well in advance to stay financially sustainable.\n\nMany statistical analyses can be run locally on CPUs unless we are handling a substantial amount of data. A CPU was sufficient for us to train the model for over 9,000 rooms across Australia, despite our novel training model including a neural network training step. However, to run an AI agent, organisations may need to access third-party platforms with a ZDR option such as AWS Bedrock, OpenRouter and Fireworks AI. For maximum data privacy and minimum data leakage, running the AI agent locally on mini PCs powered by commercial-grade GPUs, such as the Nvidia DGX Spark or Apple Mac Studio, can be considered. Both these PCs are around A$9,500 (computeralliance.com.au and apple.com/au) and do not consume a substantial amount of power.\n\n## Where and how to use AI agents\n\nSo does that mean organisations operating in regulated environments should not explore and benefit from agentic AI capability? Absolutely not. The question is not whether to use AI agents, but where and how to use AI agents to keep the process transparent, safe and cost-effective.\n\nThe AI agent can be used to provide actionable business insights and to provide evidence and answers to a regulator. In our example of room prices, the regulator may ask ‘what logic did you use to come up with a number?’, and the AI agent could provide answers such as ‘we ran validated statistical and machine learning models to learn patterns and make predictions based on those patterns’. The AI agent could do a deeper dive on the impact of different variables on prices. For example, what impact room size and median house price may have on room prices, and how much changing room sizes and median house prices would move the price.\n\nAnother role of the AI agent would be to use this prediction to look at the big picture for an organisation. It could provide answers such as ‘adopting predicted prices would allow us to improve EBITDA by 15%’. However, to receive such insights, financial tools will have to be developed in the harness.\n\nI developed Prime following the same principles and structure. How Prime performs that orchestration is shown in the diagram below:\n\nHow does Prime work? Like other AI agents, Prime’s most integral components are the harness and the LLM. The harness is based on tools, rules and a guard. The tools return only facts from the knowledge base and validated predictions. The guard prevents any names (this could be expanded to include numbers) that do not come from the tools; in other words, Prime does not invent any names itself. While undertaking a test, Prime caught all 14 fake names in September 2026. The tools only call the predictions that have been generated locally by a robust statistical model and validated.\n\nOntology allows domain knowledge-based rules to be written outside of the code, avoiding the risk of code making errors while handling domain knowledge information. This, if not addressed, could lead to inaccurate results which may prove to be detrimental to organisations. For example, very early in the process of developing a statistical model to predict room prices for individuals, I observed inflated room prices for shared rooms. These prices were much higher than the normal ranges. Further investigation uncovered that the statistical model had processed the total size for shared rooms during training. For shared rooms, more than one person shares the room, so the total size must be divided by the number of people sharing the room to determine the room size per person.\n\nIf I had let Prime build the model, undertake the training and generate predictions, it is likely that Prime would have made the same error.\n\n## Conclusion\n\nSo where does this leave us? The journey I referred to earlier is my journey given the problem and situation. Ontology can take different forms for different problems in different organisations. The definitions of ‘Terms’, ‘Classes’ and ‘Slots’ can vary; an ontology can be designed differently, taking into consideration the internal factors.\n\nThe key takeaways from my experience include:\n\n- Test: If a board or a regulator asks ‘what logic or reasons did you use to support price decisions?’, the provided answer runs back through the chain: ontology, knowledge base, validated model, tools, guard and audit log, to provide a clear and transparent answer to support your decision.\n- Principle: LLMs can be swapped but the ontology and the validated model can’t. The main focus should be on designing the right ontology and transparent, robust statistical models. The AI agent is LLM-agnostic.\n- Order: The ontology is designed first and the AI agent developed later, not the other way around.\n\nThe renewed focus on ontology can be explained as it provides a structured framework for organisations to benefit from the evolving agentic AI capability while offering transparent, explainable and evidence-backed answers.\n\n## References\n\nGruber, T. R. (1993). A translation approach to portable ontology specifications. *Knowledge Acquisition*, 5(2), 199–220.\n\nNoy, N. F., & McGuinness, D. L. (2001). *Ontology development 101: A guide to creating your first ontology*. Stanford Knowledge Systems Laboratory, Technical Report KSL-01-05.\n\nComputer Alliance, computeralliance.com.au, Nvidia DGX Spark price, accessed November 2026.\n\nApple Australia, apple.com/au, Mac Studio price, accessed November 2026.", "url": "https://wpnews.pro/news/ontology-in-the-age-of-ai", "canonical_source": "https://safdarali2.substack.com/p/ontology-in-the-age-of-ai", "published_at": "2026-11-16 12:03:31+00:00", "updated_at": "2026-11-16 12:23:27.099738+00:00", "lang": "en", "topics": ["ai-agents", "artificial-intelligence", "ai-tools"], "entities": ["Thomas R. Gruber", "Noy & McGuinness", "Australia"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ontology-in-the-age-of-ai", "markdown": "https://wpnews.pro/news/ontology-in-the-age-of-ai.md", "text": "https://wpnews.pro/news/ontology-in-the-age-of-ai.txt", "jsonld": "https://wpnews.pro/news/ontology-in-the-age-of-ai.jsonld"}}