Essay. June 2026.

Summary

Enterprise AI adoption is rising, but organisational impact remains uneven. The constraint is no longer model capability. It is whether an organisation has a faithful, machine-actionable representation of how it actually works.

As systems move from answering questions to taking actions, fluency is not enough. What matters is grounding: a shared model of entities, relationships, rules and permissions that defines what the business is, and what it is allowed to do.

Computer science has long had a name for this: ontology. What is changing now is not the idea itself, but its role. Ontologies are becoming the operational substrate that lets agents act safely, consistently and at scale inside enterprises.

This essay traces the gap between adoption and impact, examines where ontologies already run at scale, describes how an agent uses one at the moment of decision, and considers why an idea once dismissed as belonging to an earlier era of AI is re-emerging as the most defensible layer in the enterprise AI stack.

1. The gap between adoption and impact

Enterprise AI adoption is rising, but realised value is lagging behind expectations.

The United States Census Bureau tracks firm-level AI use every two weeks through its Business Trends and Outlook Survey, a nationally representative instrument covering roughly 200,000 businesses. Under its original, narrower question about using AI to produce goods and services, adoption sat in the low single digits through 2023 and 2024 (Bonney et al., 2024). After the question was broadened in late 2025 to cover any business function, the figure rose to around 18% of firms, or about 32% once weighted by employment (U.S. Census Bureau, 2026).

Figure 1. Adoption is real but firm-level integration lags individual use; the best peer-reviewed estimates of value remain modest and concentrated.

A Federal Reserve research note synthesising several government and academic surveys draws out the revealing contrast. While roughly 41% of individuals report using generative AI for work, only about 18% of firms have integrated it at the organisational level (Allen, 2026). Personal, ad-hoc use is running well ahead of deliberate enterprise integration.

The pattern is consistent across sectors. AI is being used inside organisations, but not yet embedded into how organisations operate.

What value has been measured is real but bounded. In the most rigorous economy-wide assessment to date, Acemoglu models AI’s effect through a task-based framework and concludes that the gain to total factor productivity over a decade is non-trivial but modest, on the order of half a percentage point and possibly less (Acemoglu, 2025). His reasoning matters more than his number. The early productivity evidence comes from easy-to-learn tasks with clear feedback, whereas the larger future gains must come from hard-to-learn, context-dependent decisions for which there is often no objective signal to learn from. Those context-dependent decisions are precisely what an enterprise ontology encodes.

The micro-evidence agrees. In a large field experiment with more than five thousand customer-support agents, Brynjolfsson, Li and Raymond found that access to a generative-AI assistant raised productivity by about 14% on average, but the gains went overwhelmingly to less-experienced workers, with little benefit to the most skilled, because the system largely captured and redistributed the tacit know-how of the best performers (Brynjolfsson, Li & Raymond, 2025). AI helped most where it could surface an organisation’s own latent expertise. Capability alone did not.

This gap is not primarily about model capability. It is about integration. And integration depends on something more fundamental than intelligence: representation.

2. From answering to acting

Most current AI systems are still used in an advisory mode. They generate text, summaries, and suggestions. In this setting, errors are tolerable. A human remains in the loop, and mistakes are filtered by judgement.

That changes when systems begin to act.

Figure 2. The same fluency carries very different stakes once the human leaves the loop and the system writes back to enterprise systems of record.

An agent that updates a database, issues a refund, modifies a policy or triggers a purchase order is no longer producing output for review. It is executing decisions inside enterprise systems of record. The cost of an error is no longer an inconvenience but a financial, operational or compliance event.

At that point, fluency stops being sufficient. Large language models generate text by predicting probable continuations rather than by consulting a verified record. In the most comprehensive recent survey of the phenomenon, Huang and colleagues distinguish two kinds of hallucination: factuality hallucination, where output diverges from real-world fact, and faithfulness hallucination, where output diverges from the user’s instruction, the supplied context, or its own internal logic (Huang et al., 2025). A grammatically flawless instruction issued against a misunderstood account schema is not a harmless slip. It is a confident, well-written and possibly costly mistake.

The model may be capable, but it is not grounded in the business it is operating inside. That gap is the source of most enterprise AI failure modes today.

3. The missing layer: representation

To understand the problem, it helps to separate three layers:

Most organisations have strong system layers and increasingly strong model layers. What they lack is a shared representation layer that connects the two.

In computer science, this is called an ontology. The canonical definition originates with Gruber, who characterised an ontology as an explicit specification of a conceptualisation: a formal description of the concepts and relationships that can exist for a community of agents (Gruber, 1993, 1995). Studer and colleagues refined this into the formulation now standard in the literature: a formal, explicit specification of a shared conceptualisation (Studer, Benjamins & Fensel, 1998). Each adjective does work. Formal means machine-readable. Explicit means the concepts and the constraints on their use are stated rather than assumed. Shared means the model captures consensus accepted by a group, not one person’s private view.

An ontology is not a taxonomy or a schema. A taxonomy is a hierarchy of categories. A data schema specifies how a particular database stores rows and columns. An ontology sits above both. It defines:

Crucially, it is shared. It is not the interpretation of one system or team. It is the agreed structure of the domain itself.

Without this layer, every system develops its own version of reality. “Customer” means one thing in finance, another in support, and another in product analytics. Humans reconcile this ambiguity informally. Agents cannot.

4. The semantic and kinetic layers

Two features of an ontology matter most for enterprise AI, and both are routinely underestimated.

The first is that an ontology is fundamentally about agreement. It is the single place where the data layer, the business logic and the AI converge on what the words mean. The Financial Industry Business Ontology, discussed below, exists in large part because the same concept, whether a counterparty, a beneficial owner or an interest rate swap, is defined differently in every internal system and every regulatory regime, and someone has to make those definitions agree (EDM Council & OMG, 2024).

The second feature distinguishes an ontology built for agents from a knowledge graph built for analytics. Analytics asks what is true. An agent additionally needs to know what may be done, by whom, under what conditions, and with what consequence. Research on formal ontology for information systems (Guarino, 1998) and on enterprise ontology as the implementation-independent essence of an organisation (Dietz, 2006) treats action and process as first-class notions that must be defined as precisely as objects and relations.

An operational ontology therefore pairs a semantic layer (objects, properties, links) with a kinetic one: the actions an agent may take, the rules that constrain them, and the permissions that govern who may act.

A data model tells you what exists. An ontology tells you what is true. An operational ontology goes further. It tells you what is permitted.

This is the difference between a map and a set of rules of the road. An enterprise needs both, but only the second keeps an autonomous system from driving into the river.

Figure 3. One shared model, used by people and agents alike, sitting between scattered data sources and the systems that act on them. The semantic layer follows Gruber (1993) and Studer et al. (1998); the kinetic layer follows the enterprise-ontology tradition of Guarino (1998) and Dietz (2006).

A concrete example makes the two layers tangible. Suppose an insurer wants an agent that can settle a straightforward motor claim. The semantic layer declares the relevant objects and links: a Claim is madeAgainst a Policy, which is heldBy a Customer, who owns a Vehicle; a Claim has a status drawn from a fixed set, and a reserveAmount. The kinetic layer declares the permitted moves and their guards. The action approvePayment may fire only when the claim’s status is Assessed, the amount falls below an authority threshold, and no open fraud flag is attached. It must record who authorised it, and it may not run twice on the same claim. The axioms enforce coherence, so that a claim cannot be Paid before it is Approved, and the permissions encode who, whether human or agent, is allowed to pull each lever.

None of this lives in the model’s weights. It is the explicit, shared, machine-checkable substance of how this insurer settles a claim.

Figure 4. The same insurance example drawn as a node graph. Entities (Customer, Policy, Claim, Vehicle) are joined by labelled, directed relationships and carry their own properties; the kinetic approvePayment action sits to the side with the guard that must hold before it can fire; and the axioms beneath constrain the whole.

5. This is not theoretical: ontologies already run at scale

A reasonable objection is that operational ontologies sound elegant on paper but unproven in practice. The record says otherwise. Some of the most demanding and heavily scrutinised information systems in the world are built on exactly this foundation.

In finance, the Financial Industry Business Ontology (FIBO) is an open, industry-standard ontology developed by the non-profit EDM Council together with the Object Management Group. It gives institutions a single authoritative definition for concepts such as legal entities, instruments and beneficial ownership, against which they map the disparate models held in their trading, risk and compliance systems. One motivation is regulatory: a common, machine-readable model lets a bank generate reports against requirements such as the Basel Committee’s BCBS 239 risk-data-aggregation principles from one consistent representation rather than from bespoke pipelines built for each regulator (EDM Council & OMG, 2024). FIBO has been developed over more than a decade by a large community precisely because reaching consensus on what financial terms mean is slow, contested work.

In the life sciences, the Gene Ontology provides a dynamic, controlled vocabulary describing gene and gene-product function that can be applied consistently across all eukaryotic organisms (Ashburner et al., 2000). It lets findings from a fly, a mouse and a human be annotated against the same formal terms and compared, and it is maintained by an international consortium as biological knowledge changes.

In medicine, SNOMED CT is a large clinical terminology built on description logic and used inside electronic health records across many national health systems. Its concepts are arranged by logical subsumption, so that Diabetes mellitus is formally a kind of Disorder of glucose metabolism, which lets software infer relationships and check the terminology for contradictions automatically.

A peer-reviewed account by Noy and colleagues, who are engineers and scientists from Google, Microsoft, IBM, Facebook and eBay, describes how each of these companies built large knowledge graphs to supply structured, factual grounding to their products (Noy et al., 2019). The authors are candid that building and, especially, maintaining these graphs at industrial scale is difficult. Entity resolution, keeping multiple stores in sync, and ingesting new knowledge without rebuilding everything are all recurring challenges. That candour is part of the value. It shows the approach is real, deployed and load-bearing, and also that it is genuinely hard.

6. How an agent uses an ontology at the moment of decision

An ontology changes what happens in the seconds before an agent acts. Without one, a language-model agent answers from its parameters and whatever text it happened to retrieve, and its fluency is the only thing standing between the organisation and a confident error. With one, the agent has a structured, authoritative substrate to consult and to be constrained by.

A grounded agent follows a constrained loop.

First, it resolves entities. It maps ambiguous references, such as a customer name or an account number, to the canonical objects in the ontology, so the agent is reasoning about the right Customer and not a near-namesake.

Second, it retrieves structured context by traversing explicit relationships rather than relying on unstructured similarity search. This is the logic behind retrieval-augmented generation, introduced by Lewis and colleagues to reduce non-factual output by grounding the model in retrieved evidence (Lewis et al., 2020), and behind its graph-based successors, which use the structure of a knowledge graph to gather context that scattered text lookup would miss (Edge et al., 2024; Pan et al., 2024).

Third, it validates every action against the kinetic layer before execution. Does the claim’s status permit payment? Is the amount within authority? Does this agent hold the permission? Reasoning and validation can catch a contradiction or a violated constraint before it becomes an executed mistake, not after.

Only then does it act, with the action and its justification recorded.

The ontology does not make the model honest. It makes dishonesty expensive to act on. A faithfulness error that survives the model’s own generation still has to pass the explicit constraints before it can move money. That is the practical meaning of grounding.

7. The return of structure

Structured knowledge representation has been out of fashion for two decades. The Semantic Web programme of the early 2000s, Berners-Lee, Hendler and Lassila’s vision of a web of machine-readable meaning, was ambitious but, in its most maximalist forms, commercially disappointing (Berners-Lee, Hendler & Lassila, 2001). Hand-authoring a formal model of a living business proved punishingly expensive, and such models tended to go stale the day they were finished. When large language models arrived, the implicit promise was that all this structure could be discarded: rather than model the world, simply train a model large enough to have absorbed it.

For years, the assumption was simple: if models are powerful enough, they will absorb the world implicitly. That assumption is breaking under operational pressure.

The more considered position emerging from recent research is that large models do not make ontology obsolete. They change its economics and raise its stakes at the same time. The field of neuro-symbolic AI, which combines the pattern-learning strengths of neural networks with the explicit, verifiable reasoning of symbolic systems, has grown rapidly since 2020, and a 2024 systematic review of the literature finds knowledge representation among its most active areas (Colelough & Regli, 2024). The motivation is well established: large models struggle with logical consistency and with reasoning reliably beyond their training distribution, and integrating explicit symbolic knowledge is a leading remedy (Wang, Yang & Wu, 2024).

Figure 5. A synthesis, not a winner: the neural component supplies fluency and can now help author the structure, while the symbolic ontology supplies ground truth and constraints. After Colelough and Regli (2024), Wang et al. (2024) and Pan et al. (2024).

Two developments pull in the same direction. First, models have made ontology tractable. The same language models that act on a knowledge base can now help build and maintain it, reading schemas, documents and code to propose entities and relationships, and flagging where the model and reality have drifted apart. This is an active research area, with work on language-model-assisted ontology learning and requirements engineering appearing at the European and Extended Semantic Web Conferences (Lippolis et al., 2025; Zhao et al., 2024). The bottleneck that once made knowledge engineering uneconomic is now being lowered by the very technology that seemed poised to render it obsolete.

Second, the empirical case for grounding has strengthened: retrieval-augmented and graph-based methods consistently use explicit structure to constrain a fluent model (Lewis et al., 2020; Edge et al., 2024; Pan et al., 2024).

The resulting picture is not a return to the old position but a synthesis. The neural component supplies fluency, perception and low-cost authoring. The symbolic ontology supplies ground truth, constraints and accountability. The model keeps the ontology from fossilising. The ontology keeps the model from acting on a misunderstanding. They co-evolve.

8. Why this is difficult to copy

A working ontology is not generic. It encodes how a specific organisation operates.

It contains definitions that differ across teams, exceptions embedded in workflows, constraints shaped by regulation and history, and tacit knowledge that is rarely written down. This makes it expensive to build and slow to replicate.

Unlike models, which are shared infrastructure, ontologies are proprietary structure. They are closer to institutional memory than software code.

The Brynjolfsson field experiment showed AI delivering value precisely by capturing and redistributing tacit expertise (Brynjolfsson, Li & Raymond, 2025). The act of building an ontology is the act of codifying that expertise deliberately rather than by accident. FIBO took a standards consortium more than a decade to negotiate, not because the technology was missing but because agreeing what the terms mean is the hard part, and that difficulty is exactly what makes the result hard to copy (EDM Council & OMG, 2024).

Each workflow modelled, each edge case resolved, each rule encoded lowers the cost of the next and raises the trustworthiness of the whole. The asset improves with use rather than decaying, so a lead tends to widen rather than erode. This also aligns with Acemoglu’s argument that the remaining productivity gains lie in hard-to-learn, context-dependent decisions with no objective signal to learn from (Acemoglu, 2025), since those are exactly the decisions an ontology makes explicit.

Figure 6. A wrapper holds only a prompt and an API call, which an incumbent can absorb; an ontology encodes the proprietary, tacit knowledge of the business, which resists copying.

The contrast is with the thin wrapper, a layer whose only contents are a prompt and an API call, which an incumbent can absorb at will. The defensible position belongs to the organisation whose hardest asset is the model of the business, not the model doing the talking.

9. The honest limitations

A claim of this kind should be tested against its weaknesses, of which there are several.

Ontologies are expensive and they rot. FIBO, the Gene Ontology and SNOMED CT are the products of large, sustained, well-funded communities, and the industrial knowledge-graph teams that Noy and colleagues describe spend much of their effort simply keeping the graph in sync with reality (Noy et al., 2019). Businesses reorganise, definitions drift, and a model no one maintains becomes worse than none, because people keep trusting it after it has begun to lie. The new ability to enlist models in keeping an ontology current is therefore not a convenience but a precondition. Without it, the economics may not close.

Ontologies are also not a universal solvent. The clinical-informatics literature documents cases where even a mature, carefully engineered terminology like SNOMED CT cannot cleanly express certain notions. Formal precision buys decidable reasoning at the cost of some things being awkward or impossible to say. An ontology raises the floor on reliability. It does not capture everything.

There is also the opposite failure mode, and it is the original sin of the Semantic Web: the temptation to model everything before shipping anything. Comprehensiveness is the enemy of deployment. The discipline is to model the minimum that lets one agent act safely on one workflow, prove it, and let the ontology grow along the paths that real actions take.

Beneath the optimism sit unresolved research problems, notably the reliable grounding of a neural model’s continuous representations in discrete symbols, which recent work treats as an open frontier rather than a solved matter (Colelough & Regli, 2024). Explicit structure reduces but does not eliminate hallucination. Faithfulness failures can persist even with correct context (Huang et al., 2025). Governance and human oversight remain necessary.

Finally, a note of caution about the central claim itself. That the ontology is the most defensible layer is a thesis consistent with the present evidence, not a proven law. The economic estimates are early and contested, and the strongest evidence for the agentic shift is still arriving. The argument should be held as the best current reading of the direction of travel, open to revision as harder data appears.

Conclusion

Enterprise AI is not limited by intelligence. It is limited by representation.

As organisations move from systems that answer to systems that act, they need a shared, machine-actionable model of how they actually work. Without it, agents operate on approximation. With it, they operate on structure.

This is not a new idea returning unchanged. It is an old idea whose economics have shifted. Models now make ontologies cheaper to build and more important to have. The synthesis of neural fluency and symbolic structure is where the field is heading, and the organisations that build a faithful representation of themselves will be the ones that can deploy AI that acts safely, consistently and at scale.

The models are becoming powerful enough to act. The question is whether the organisation they act on is defined clearly enough for them to do so safely.

References

Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121), 13–58.

Allen, J. S. (2026). Monitoring AI adoption in the U.S. economy (FEDS Notes). Board of Governors of the Federal Reserve System.

Ashburner, M., Ball, C. A., Blake, J. A., Botstein, D., Butler, H., Cherry, J. M., … Sherlock, G. (2000). Gene Ontology: Tool for the unification of biology. Nature Genetics, 25(1), 25–29.

Berners-Lee, T., Hendler, J., & Lassila, O. (2001). The Semantic Web. Scientific American, 284(5), 34–43.

Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L., Goldschlag, N., Haltiwanger, J., Kroff, Z., & Savage, K. (2024). Tracking firm use of AI in real time: A snapshot from the Business Trends and Outlook Survey (NBER Working Paper No. 32319; U.S. Census Bureau CES-WP-24-16). National Bureau of Economic Research.

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.

Colelough, B. C., & Regli, W. (2024). Neuro-symbolic AI in 2024: A systematic review. CEUR Workshop Proceedings, 3819. (Preprint arXiv:2501.05435.)

Dietz, J. L. G. (2006). Enterprise ontology: Theory and methodology. Springer.

Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., & Larson, J. (2024). From local to global: A Graph RAG approach to query-focused summarization. arXiv:2404.16130.

EDM Council & Object Management Group. (2024). Financial Industry Business Ontology (FIBO) [Standard specification and documentation]. Enterprise Data Management Council; Object Management Group.

Gruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition, 5(2), 199–220.

Gruber, T. R. (1995). Toward principles for the design of ontologies used for knowledge sharing. International Journal of Human-Computer Studies, 43(4–5), 907–928.

Guarino, N. (1998). Formal ontology and information systems. In N. Guarino (Ed.), Proceedings of FOIS’98 (pp. 3–15). IOS Press.

Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), Article 42.

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474.

Lippolis, A. S., Saeedizade, M. J., Keskisärkkä, R., Zuppiroli, S., Ceriani, M., Gangemi, A., Blomqvist, E., & Nuzzolese, A. G. (2025). Ontology generation using large language models. In The Semantic Web – ESWC 2025 (pp. 321–341). Springer.

Noy, N. F., Gao, Y., Jain, A., Narayanan, A., Patterson, A., & Taylor, J. (2019). Industry-scale knowledge graphs: Lessons and challenges. Communications of the ACM, 62(8), 36–43.

Pan, S., Luo, L., Wang, Y., Chen, C., Wang, J., & Wu, X. (2024). Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering, 36(7), 3580–3599.

Polanyi, M. (1966). The tacit dimension. University of Chicago Press.

Studer, R., Benjamins, V. R., & Fensel, D. (1998). Knowledge engineering: Principles and methods. Data & Knowledge Engineering, 25(1–2), 161–197.

U.S. Census Bureau. (2026). Business Trends and Outlook Survey: Artificial intelligence supplement. U.S. Department of Commerce.

Wang, W., Yang, Y., & Wu, F. (2024). Towards data- and knowledge-driven AI: A survey on neuro-symbolic computing. IEEE Transactions on Pattern Analysis and Machine Intelligence.

W3C. (2012). OWL 2 Web Ontology Language document overview (2nd ed.). World Wide Web Consortium.

W3C. (2017). Shapes Constraint Language (SHACL). World Wide Web Consortium.

Zhao, Y., Zhang, B., Hu, X., Ouyang, S., Kim, J., Jain, N., de Berardinis, J., Meroño-Peñuela, A., & Simperl, E. (2024). Improving ontology requirements engineering with OntoChat and participatory prompting. Proceedings of the AAAI Symposium Series, 4, 253–257.