The Missing Enterprise Layer stalling your Agentic AI deployments
Why business architecture is becoming a precondition for agentic AI at scale
I recently watched frank coyle of UC Berkeley explain why agentic systems need ontologies at the AI Engineer World’s Fair, and it prompted me to reflect on a more fundamental question: can an organisation be ready for agentic AI if it has never made its own business architecture sufficiently explicit?
Coyle’s technical argument centres on the probabilistic character of large language models. An agent can interpret a request, formulate a plan and propose an action, but it requires a formal representation of the concepts, relationships and constraints that define the domain in which it is expected to act. He argues that an ontology addresses that need by providing a structured basis against which the coherence of the proposed action can be assessed.
Coyle is right that agentic systems need ontologies, and his argument captures a broader return to expert systems and symbolic controls. The extension I would make is that enterprise agentic systems need business architecture with ontology. Business architecture creates and governs the shared conception of purpose, value, capabilities, information, policy and authority; ontology formalises the elements that agents must interpret; and agents operate within the resulting boundaries.
This extension matters because an ontology can formalise a conception of the enterprise, but it cannot determine whether that conception is strategically coherent, accepted across the organisation or authoritative at the boundaries between functions. The missing element is the enterprise layer: the governed representation of the business that connects organisational intent to machine action. Business architecture provides the integrative discipline through which this layer can be created and maintained.
The stall occurs when a successful pilot must cross into live enterprise operation. Within a bounded experiment, business meaning, exceptions and decision rights can remain implicit because the scope is narrow and domain specialists are nearby. At enterprise scale, the agent encounters competing definitions, fragmented processes and responsibilities distributed across organisational boundaries. Without an enterprise layer, each deployment team must reconstruct this context for itself, and the programme accumulates bespoke integrations, controls and exceptions faster than it creates reusable capability.
Why pilots stall at the enterprise boundary
An AI pilot normally operates with selected data, known users, restricted actions and a delivery team close enough to interpret unexpected results. Much of its business context can consequently remain tacit, embedded within prompts, code, local documentation or the assumptions of subject-matter experts. A successful pilot may demonstrate that a model can perform a specified task, but it provides limited evidence that the organisation can support that task as part of a live, governed and cross-functional operation.
Scaling changes the nature of the problem because the agent must operate across systems and functions, distinguish unusual transactions from impermissible ones, recognise when local optimisation conflicts with an enterprise objective and know when authority should transfer to a human decision-maker. At that point, the question is no longer simply whether the agent can perform a task, but how that task contributes to an end-to-end flow of value: which outcome it serves, which capabilities and information it depends upon, and where accountability crosses functional boundaries. Requirements previously absorbed through human judgement consequently become explicit dependencies of the system.
The Capgemini Research Institute reports that only 2% of organisations have deployed AI agents at scale and that fewer than one in five possess mature supporting data and technology infrastructure. Limited readiness is accompanied by high attrition: RAND cites estimates that more than 80% of AI projects fail, while Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. Together, the figures describe organisations developing AI faster than they are establishing the conditions required to deploy it. Because they measure different stages of adoption, they should not be treated as direct comparators.
Taken together, Gartner’s concerns about escalating costs, unclear business value and inadequate risk controls, and RAND’s findings on poorly understood business problems, insufficient data, technology-led solution design and inadequate infrastructure, show that the barriers to AI deployment are organisational as well as technological. Viewed through a business architecture lens, these barriers form a connected pattern: unclear value reflects limited traceability between an AI intervention, the capability it changes and the outcome it serves; inadequate controls suggest that permitted actions, decision rights and accountable ownership remain underspecified; and poor data readiness often exposes unresolved differences in business meaning, information ownership and use across functions.
My prior research on business architecture and digital transformation reached a related conclusion, finding that stronger alignment between business and technology was associated with greater efficiency and better service delivery. In practical terms, organisations reported better outcomes where business and technology were integrated, ways of working were standardised and organisational links were stronger. Agentic AI makes this alignment more consequential because technology that initiates action requires organisational intent, information and authority to be expressed with greater precision than technology that merely informs decisions.
The organisational work hidden inside an ontology
Thomas Gruber’s 1993 formulation described an ontology as an “explicit specification of a conceptualization”. Studer, Benjamins and Fensel subsequently developed the widely used definition of an ontology as a “formal, explicit specification of a shared conceptualisation”. In the context of enterprise AI, the most consequential element of that definition is the requirement for the conceptualisation to be shared.
Although formalisation can improve precision, consistency and computational usability, sharedness remains an organisational achievement. A large enterprise rarely begins with a single uncontested definition of a customer, product, service, contract or transaction. Sales may define an active customer by reference to an open commercial relationship, finance by reference to a current account balance, and customer service by reference to an existing entitlement. Because each definition may be legitimate within its own context, an enterprise agent must know which one is authoritative for the action being considered, under which conditions it applies and who has the right to change it.
Coyle illustrates the consequences through agents that might attempt a second refund against an order already refunded, apply a status the business does not recognise or direct a payment to the wrong party. Preventing such outcomes requires an agreed definition of the relevant business state, the rules governing the proposed action and the authority to approve any exception. A technical system can apply those decisions, but it cannot determine what the organisation intended them to be.
Figure 1. Business architecture reconciles competing functional meanings before selected enterprise concepts are formalised for agentic use.
Get Dr Dennis O'Higgins’s stories in your inbox
Join Medium for free to get updates from this writer.
Business architecture builds the core of the enterprise layer by orchestrating relationships among strategy, capabilities, value streams, information, policy and accountability. It locates an agent within the enterprise and defines the purpose it serves, the boundaries within which it may act and the outcomes for which the organisation remains accountable. An ontology can then formalise selected elements of that layer for machine interpretation and use.
Operationalising the enterprise layer requires business architecture to establish strategic and organisational coherence, while ontology and knowledge engineering formalise the relevant semantics. Data architecture connects those semantics to authoritative information, security and policy mechanisms enforce authority, and AI engineering integrates the components into an operational system. This sequence matters because the enterprise must determine what its concepts mean and how they should govern action before those decisions can be encoded and executed.
Coyle also notes a return to the logic of expert systems, which is intriguing because the knowledge acquisition bottleneck limited their ability to scale: expert knowledge had to be elicited, encoded and continually updated as the domain changed. Large language models can reduce that burden by extracting and structuring knowledge, but they cannot decide which definition should prevail, who has authority to approve it or how it should change with strategy, policy and regulation. The return to symbolic methods therefore revives an organisational problem alongside the technical one. Business architecture with ontology provides the more complete response: business architecture establishes coherence, ownership and authority, while ontology makes selected business meaning machine-interpretable and enforceable.
The enterprise layer cannot be inferred from systems
The growing use of AI to generate semantic or context layers from existing schemas, metadata, dashboards and query histories may appear to reduce the need for prior architectural work. These techniques can accelerate discovery, identify recurring concepts and expose where systems apply different definitions to similar information. However, the resulting representation describes how existing systems encode the business. It does not determine which interpretation the enterprise has authorised to govern future action.
If three systems contain three definitions of an active customer, inference can discover and represent all three with considerable precision. It cannot decide which definition should govern credit control, service entitlement or regulatory reporting. That decision requires business ownership, policy judgement and an explicit mechanism for resolving disagreement. An inferred representation is valuable evidence, but it is not an authorised definition. Without governance, it risks becoming an accurate machine-readable representation of organisational fragmentation that agents subsequently treat as authoritative.
Every semantic or context-layer programme should therefore be able to answer three questions: who ratified its principal business definitions, who owns those definitions and their permitted uses, and how changes in strategy, policy or operating model will propagate into the representations consumed by agents. Where those questions remain unanswered, the programme has an unrecognised dependency on the enterprise layer.
In my experience, the strongest resistance to adopting business architecture as a strategic discipline is the perception that it delays delivery. When executive attention is fixed on the transformation currently holding priority, investing in a coherent view of strategy, value, capabilities, information and authority can appear indulgent. Organisations often reconsider only after delivery encounters cross-functional dependencies, initial attempts to remediate the resulting problems fail and the work previously deferred becomes unavoidable. At that point, the wider role of business architecture becomes visible: an end-to-end blueprint connecting strategy to execution that requires upfront investment while accelerating complex, boundary-spanning change.
Conversely, organisations making the strongest progress with enterprise AI, particularly agentic deployments, have typically invested in business architecture capabilities and treated them as an enterprise-wide strategic practice rather than as a subsidiary function of technology or transformation. This positioning matters because a practice subordinated to technology or a transformation programme lacks the enterprise-wide mandate needed to arbitrate between functions whose definitions, priorities and accountabilities are in dispute. Their advantage extends beyond possessing richer models because they have already clarified their capabilities, established common information definitions, organised change around end-to-end value rather than individual projects or systems, assigned ownership and developed governance mechanisms. Although the original justification may have been transformation, integration, regulatory compliance or operating-model redesign, those investments now provide a material source of AI readiness because much of the required business context is already explicit, stable and governable.
The value of control is easy to mismeasure
Conventional transformation cases typically assess value through cost, speed and service improvement, leaving the benefits of control largely outside the business case. For agentic AI, however, business architecture may create value precisely by enabling controls that add cost and processing time while preventing invalid actions, blocking unauthorised transactions, avoiding regulatory breaches and preserving decision evidence. A business case focused only on productive efficiency can therefore record the cost of control without recognising the failures it prevents, particularly in regulated and financially consequential environments where control effectiveness forms part of the value proposition.
The asymmetry is significant: the cost of control is immediate and measurable, while the value of an avoided failure remains counterfactual unless the organisation records blocked actions, overrides and escaped incidents. Without that evidence, a better-governed agentic system may appear less efficient precisely because the business case recognises the cost of prevention but not the losses it avoids.
From descriptive architecture to runtime governance
Business architecture has often been used descriptively to communicate strategy, represent the operating model and structure transformation portfolios. Agentic AI creates a more operational role in which capability boundaries determine an agent’s scope, information concepts shape its interpretation, value streams locate its actions, policies constrain its choices and decision rights trigger human intervention. Under these conditions, selected elements of the enterprise layer become part of the runtime governance environment.
This transition does not require every business architecture model to become executable; it does, however, require the relevant models to meet a higher standard of precision, internal coherence, ownership and traceability. The semantic and control structures used by agents must remain connected to changes in strategy, regulation, policy and operating model. Architecture governance can therefore no longer end when the initial transformation programme closes; the enterprise layer must be managed as a governed asset rather than a project artefact.
The enterprise layer provides a common business context for strong data foundations, secure technical architecture, model and agent governance, operational controls, domain expertise and organisational change, keeping them aligned with organisational purpose, value and authority. Each remains necessary in its own right.
I believe Coyle’s proposition can therefore be extended: agentic systems need ontologies; but enterprise agentic systems need business architecture governing that ontology. Business architecture creates and governs the shared enterprise conception, ontology formalises the elements that machines must interpret, and agents act within the resulting boundaries. Formalism without governed meaning recreates the brittleness associated with expert systems, while governed meaning without formalism cannot constrain machine action. Enterprise AI readiness ultimately depends upon whether an organisation can explain, consistently and authoritatively, what it does, how it creates value, what its information means and who is permitted to decide and act. An enterprise that cannot provide that account is not yet ready to delegate consequential action to a machine.
If you have ~20 minutes, here is the link to the talk that prompted this article: Why Agentic Systems Need Ontologies, Frank Coyle, UC Berkeley.
