How Much AI Governance is Enough?

Abstract

The rapid adoption of generative and agentic artificial intelligence is transforming AI from a specialised technological capability into an increasingly embedded component of organisational processes and decision-making. This transformation challenges conventional approaches to AI governance, which often emphasise principles, policies and model-level oversight without fully addressing the organisational consequences of increasingly autonomous and interconnected AI systems. This paper examines the question of how much AI governance is enough and argues that governance should be understood as a form of organisational control rather than as a fixed compliance layer. Drawing on recent research on AI governance, responsible AI, generative and agentic AI, organisational adoption, cybersecurity, resilience and decision-making, the paper develops a proportional view of governance in which control requirements increase with AI capability, autonomy, data sensitivity, decision consequence, workflow criticality and organisational dependency. The analysis shows that effective governance must extend beyond model oversight to encompass Shadow AI, human judgement, workflow integration, cybersecurity and resilience, third-party dependencies, procurement, identity and permissions, and the actions of autonomous AI agents. In particular, the emergence of agentic AI requires a shift from model governance to action governance, while increasing reliance on external AI providers requires a corresponding shift towards dependency governance. Governance should therefore be evaluated not by the number of policies or approval mechanisms an organisation maintains, but by its ability to demonstrate accountability, constrain AI according to risk, monitor behaviour, intervene when necessary and remain operational when AI systems or critical dependencies fail. The paper concludes that there is no universal level of sufficient AI governance. Instead, “enough” governance represents the level of organisational control necessary to preserve accountability, resilience and operational continuity as AI capability and organisational reliance increase.

Keywords: artificial intelligence governance; generative AI; agentic AI; organisational control; responsible AI; human–AI collaboration; cybersecurity; AI resilience; enterprise architecture

1. Introduction

Artificial intelligence (AI) has rapidly evolved from a specialist technology into an increasingly important organisational capability. Generative AI can produce text, code, analysis and recommendations at scale, while emerging agentic AI systems can coordinate tasks, interact with digital environments and enterprise applications, and increasingly execute actions with limited direct human intervention (Hosseini and Seilani, 2025; Kshetri, 2025). This development creates substantial opportunities for productivity improvement, decision support, process redesign and new forms of organisational value. At the same time, it fundamentally changes the nature of the governance challenge.

The central question is therefore no longer simply whether an organisation should permit employees to use AI. It is whether the organisation can understand where AI is being used, determine what AI systems are authorised to do, constrain their behaviour according to organisational and regulatory requirements, monitor their performance and retain accountability for decisions and actions in which AI is involved. As AI becomes increasingly embedded in organisational workflows, governance must extend beyond individual models and applications to encompass the processes, data, identities, technologies and organisational decisions through which AI creates effects (Papagiannidis, Mikalef and Conboy, 2025; Schulte and Kanbach, 2026).

This challenge is intensified by the speed at which AI adoption can develop. Employees may adopt publicly available or externally provided AI services before organisations have established formal approval processes, security controls or clearly defined policies. Research on Shadow AI demonstrates that such unsanctioned or insufficiently governed use can create risks involving sensitive data, privacy, unreliable outputs, bias and governance gaps, while also demonstrating that employee demand for AI capabilities can develop faster than organisational governance mechanisms (Silic, Silic and Kind-Trüller, 2025). Shadow AI should therefore not be understood solely as a compliance problem. It can also indicate that formal governance and approved technology provision have failed to keep pace with organisational demand.

The governance problem becomes more significant as AI moves from producing information towards influencing or executing decisions and actions. Research on organisational decision-making shows that generative AI can participate in multiple stages of decision processes, including attention, information processing, option development, choice, implementation and feedback (Schulte and Kanbach, 2026). Experimental research further indicates that the value of AI in team decision-making depends on how knowledge is distributed between humans and AI systems and how effectively their respective contributions are integrated (Zercher et al., 2025). Consequently, human involvement cannot be treated as a simple procedural requirement. Effective governance must ensure that human judgement remains meaningful, that responsibilities are clear and that decision-makers can understand when AI outputs should be trusted, challenged or rejected.

The emergence of agentic AI extends this problem further. Agentic systems can move beyond generating recommendations to interacting with software, invoking tools, accessing information and executing actions. This creates governance requirements around identity, permissions, authority, monitoring and intervention that are qualitatively different from those associated with conventional decision-support applications (Hosseini and Seilani, 2025; Kshetri, 2025). An organisation must therefore govern not only what an AI system produces, but also what it can access, what it is permitted to change, which systems it can interact with and who remains accountable for its actions.

At the same time, AI governance cannot be reduced to technical controls. Research on AI governance identifies governance as a multidimensional organisational phenomenon involving structures, responsibilities, processes and relationships across different organisational and institutional levels (Batool, Zowghi and Bano, 2025). Responsible AI governance similarly emphasises the need to translate principles into operational practices across the design, deployment, monitoring and evaluation of AI systems (Papagiannidis, Mikalef and Conboy, 2025). Governance therefore has to connect technological controls with organisational authority, business processes and managerial accountability.

This creates an important tension. Organisations need sufficient governance to prevent unacceptable uses, protect sensitive information, manage material risks and preserve accountability. However, governance that is excessively rigid, slow or disconnected from operational realities may inhibit experimentation, delay legitimate innovation or encourage employees to circumvent formal processes. AI governance must therefore be proportionate to the risks and consequences associated with particular uses rather than applying identical controls to every AI application. Gartner (2026) captures this principle by presenting AI governance as a maturity journey rather than a universal governance model. As organisations progress from experimentation towards wider deployment and more autonomous forms of AI, governance becomes increasingly formalised, systematic and embedded in organisational processes.

This maturity perspective is particularly important because AI adoption is not simply a technology implementation exercise. Research on organisational adoption shows that the ability to derive value from generative AI depends on organisational conditions such as technological complexity, organisational capabilities, skills and the capacity to manage change (Hughes et al., 2026). Similarly, Sarkar (2026) argues that the value of generative AI depends increasingly on its integration into organisational workflows rather than on isolated experimentation with individual tools. AI governance must therefore evolve alongside the way AI is incorporated into business processes. Governance that is adequate for experimentation may be insufficient once AI becomes embedded in critical workflows or capable of taking autonomous actions.

This also connects AI governance with cybersecurity and organisational resilience. The increasing use of AI creates new data flows, attack surfaces, dependencies and opportunities for unintended disclosure or manipulation (Radanliev, Santos and Ani, 2025). As AI systems become more deeply integrated into organisational processes, failures can have consequences beyond the technical system itself, potentially affecting business continuity, decision quality and organisational performance. Cyber resilience research consequently emphasises the importance of organisational capabilities that extend beyond prevention to include the ability to withstand disruption, recover from incidents and adapt to changing conditions (Neri, Niccolini and Virili, 2026). AI governance must therefore consider not only whether an AI system operates as intended under normal conditions, but also whether the organisation can detect, contain and recover from failures or unexpected behaviour.

The governance challenge can thus be understood as a transition from governing AI as a collection of technologies to governing AI as part of the organisation's operating system. Accountability must extend across AI applications, data, workflows, identities, human decision-makers and external technology dependencies. Governance must also remain adaptive because AI capabilities, deployment patterns and associated risks continue to change (Batool, Zowghi and Bano, 2025; Gartner, 2026).

This paper develops this argument by treating AI governance primarily as a question of organisational control. Organisational control refers to the ability to understand where AI is used, establish legitimate authority over its use, define what systems and users are permitted to do, constrain actions according to risk, monitor behaviour and outcomes, maintain meaningful human accountability, and respond effectively when AI systems behave unexpectedly. Such control must be embedded not only in policies but also in architecture, identity and access mechanisms, data-management practices, workflow design, monitoring, procurement and organisational decision-making.

The central argument is therefore that the appropriate level of AI governance is the level of organisational control required to preserve accountability, security and resilience in relation to an organisation's current AI capabilities, autonomy and exposure. There is no universally sufficient governance model. As AI moves from experimentation towards widespread workflow integration and increasingly autonomous action, governance must become more continuous, contextual and embedded in the organisation's technological and managerial infrastructure. The ultimate test of governance maturity is consequently not whether an organisation has documented AI principles, but whether it can demonstrate effective control over how AI is used, what it can do, how its effects are monitored and who remains accountable when outcomes matter.

2. From AI Principles to Organisational Control

The first generation of AI governance was largely principle-oriented. Organisations developed statements concerning fairness, transparency, privacy, accountability and responsible use, often in response to emerging ethical concerns and regulatory expectations. These principles remain an important foundation for responsible AI. However, the existence of principles or policies does not, in itself, demonstrate that AI is effectively governed. The central challenge is to translate normative commitments into organisational practices that influence how AI is selected, developed, deployed, monitored and used.

Batool, Zowghi and Bano's (2025) systematic literature review demonstrates the breadth of contemporary AI governance and highlights fundamental questions concerning who is accountable, what is governed, when governance should occur across the AI lifecycle and how governance is implemented through policies, frameworks and tools. Their findings also demonstrate that AI governance operates across multiple organisational and institutional levels. It therefore cannot be reduced to a single policy, compliance function or technical control. Effective governance requires coordination between organisational responsibilities, technological mechanisms and the processes through which AI is embedded in practice.

This distinction can be understood as the difference between governance as intent and governance as operational control. A policy may prohibit employees from entering confidential information into publicly available AI services, but the policy alone does not establish whether the organisation can identify which services employees are using, determine what information is being transferred, prevent inappropriate disclosure or investigate violations. Similarly, a requirement for human oversight provides limited protection if the organisation has not defined which decisions require human approval, who is accountable for those decisions, what information must be available to the reviewer and whether the reviewer has sufficient expertise and authority to challenge an AI-generated recommendation.

The literature on responsible AI increasingly emphasises this need for operationalisation. Papagiannidis, Mikalef and Conboy (2025) conceptualise responsible AI governance as involving structural, relational and procedural practices through which responsible AI principles are translated into organisational activities. This perspective is important because responsibility is not established merely by defining desirable outcomes. It depends on the organisational structures, relationships and processes that determine how AI is designed, deployed, monitored and evaluated. Batool, Zowghi and Bano (2025) similarly show that AI governance involves multiple actors and levels of responsibility, reinforcing the need to connect formal governance arrangements with actual organisational practices.

The practical implication is that AI governance must increasingly become part of the organisation's operating environment. Governance should influence which AI systems can be procured, what data they may access, which users can invoke them, which decisions they may support or execute, how their outputs are monitored and what happens when their behaviour deviates from expectations. This makes governance inseparable from areas such as data management, cybersecurity, identity and access management, procurement, risk management and business-process design. Governance is consequently not simply an additional layer placed above AI systems; it becomes part of the organisational infrastructure through which AI is controlled.

This shift is particularly important because the object of governance is changing. When AI is used primarily to generate text, images, code or recommendations, governance can focus heavily on output quality, data protection, acceptable use and human review. As AI becomes integrated into workflows, however, the consequences of AI behaviour increasingly depend on the actions that follow from its outputs. Agentic AI can interact with applications, retrieve information, invoke tools and execute tasks with varying degrees of autonomy (Hosseini and Seilani, 2025; Kshetri, 2025). The governance question therefore expands from whether an AI-generated output is acceptable to whether the system is authorised to perform the action associated with that output.

This creates a fundamental distinction between output governance and action governance. An organisation may be able to review an AI-generated recommendation before it is used, but it faces a different risk when an AI system can independently modify records, initiate transactions, send communications, change configurations or access sensitive information. In such circumstances, governance requires explicit boundaries around identity, authority, permissions and intervention. Kshetri (2025) highlights the importance of security, identity and oversight as agentic AI systems become more autonomous, while Hosseini and Seilani (2025) demonstrate through their systematic review that agentic AI introduces broader questions concerning autonomy, interaction and organisational integration.

The increasing autonomy of AI therefore makes organisational control a more useful concept than policy compliance alone. Organisational control concerns whether the organisation can establish and enforce the conditions under which AI operates. This includes knowing where AI is deployed, understanding what data and systems it can access, determining which actions it is authorised to perform, monitoring its behaviour and outcomes, and intervening when necessary. It also requires clearly defined human and organisational accountability so that responsibility does not become obscured when AI participates in a decision or executes an action.

The distinction between governance and control should not, however, be interpreted as a separation between the two. Governance establishes the authority, rules, responsibilities and decision rights through which AI is managed, while organisational control provides the mechanisms through which those arrangements become effective in practice. Governance without control risks becoming declarative; control without governance risks becoming fragmented, inconsistent or disconnected from organisational objectives. Effective AI governance therefore requires both.

This perspective also changes how governance maturity should be assessed. A mature organisation should not be judged primarily by the existence of an AI policy, ethics framework or approval committee. Instead, maturity should be reflected in whether the organisation can demonstrate that its governance arrangements operate under real conditions. This means being able to identify AI use, assign accountability, enforce appropriate boundaries, monitor material risks, investigate unexpected behaviour and intervene when necessary. Gartner (2026) similarly frames AI governance as a progressive maturity journey in which governance becomes increasingly formalised, systematic and embedded as AI adoption expands.

The emphasis on demonstrated control is particularly relevant because organisational AI adoption frequently develops faster than formal governance. Shadow AI research shows that employees can adopt AI capabilities outside approved organisational frameworks, creating new data, security and accountability risks (Silic, Silic and Kind-Trüller, 2025). The appropriate response is therefore not necessarily to prohibit AI use, but to establish governance mechanisms that make legitimate experimentation visible and manageable while distinguishing it from uses that create unacceptable risk. Governance becomes effective when the organisation can bring AI activity within a defined sphere of accountability without unnecessarily constraining productive use.

Organisational control must also remain dynamic. AI systems, underlying models, external providers and organisational use cases can change rapidly. A governance decision that is appropriate during experimentation may become inadequate once an AI application is integrated into a critical business process or receives access to sensitive organisational data. Gartner (2026) therefore presents governance as an evolving journey from planning and experimentation towards stabilisation, scaling and increasingly systematic governance. This progression reflects a broader principle: governance intensity should develop alongside the capability, autonomy, reach and consequences of AI use.

The transition from AI principles to organisational control consequently represents more than a change in terminology. It reflects a shift in the fundamental purpose of AI governance. Principles establish what an organisation considers acceptable; governance establishes who has authority and responsibility; and organisational control determines whether those expectations can actually be enforced and demonstrated in practice. As AI becomes more autonomous and deeply integrated into organisational processes, the ability to exercise such control becomes increasingly important. The question is no longer simply whether an organisation has established responsible AI principles, but whether it can demonstrate that those principles meaningfully constrain, guide and govern AI behaviour.

3. The Governance Challenge Changes as AI Becomes More Autonomous

Generative AI initially entered organisations primarily as an assistive technology. Employees used it to draft documents, summarise information, generate code, analyse material and support routine knowledge work. Although these systems could substantially influence human judgement, the human user generally remained the visible decision-maker and retained direct control over whether and how an AI-generated output was used. This configuration placed much of the governance emphasis on acceptable use, data protection, output reliability, human review and employee awareness (Silic, Silic and Kind-Trüller, 2025; Radanliev, Santos and Ani, 2025).

The emergence of increasingly agentic AI changes this relationship. Agentic systems can plan sequences of tasks, retrieve information, interact with software, invoke external tools and execute actions with varying degrees of human intervention (Hosseini and Seilani, 2025). The governance problem therefore extends beyond the quality or appropriateness of an AI-generated output. The critical question increasingly becomes: what is the AI system authorised to do?

This distinction is strategically important because an agent does not necessarily operate as a passive recipient of human instructions. It can maintain context, make intermediate decisions and initiate a sequence of actions across connected applications and information sources. Consequently, the effective risk associated with an agent depends not only on the underlying model but also on the data, identities, tools, permissions, applications and workflows to which it has access (Hosseini and Seilani, 2025; Kshetri, 2025). A relatively capable model operating within a tightly constrained environment may create less organisational risk than a less capable system that has extensive access to sensitive information or business-critical processes.

This makes authority a central concept in AI governance. Traditional application security often focuses on whether a user is authorised to access a particular system or resource. Agentic AI introduces an additional question: what is the AI itself authorised to access, decide and execute on behalf of the organisation? Kshetri (2025) highlights the importance of identity, security and oversight as autonomous intelligent systems become more capable of acting within organisational environments. The governance challenge is therefore not simply to identify the human user behind an AI interaction, but to establish and maintain a clear chain of authority between the human, the AI system and the actions the system is permitted to perform.

The practical implications become apparent when comparing different AI use cases. An agent that can access only publicly available information and generate internal recommendations represents a fundamentally different risk from an agent that can retrieve confidential customer records, modify enterprise data, initiate financial transactions or communicate externally without prior approval. Treating these systems as equivalent because they are both classified as “AI” would obscure the differences in potential impact. Governance must instead reflect the consequences of what the system can actually do.

This suggests that governance intensity should be proportionate to the combination of AI capability, autonomy, data sensitivity and potential organisational consequence. Low-risk experimentation may require relatively lightweight controls where systems have limited access, cannot execute consequential actions and operate within clearly defined boundaries. As AI gains access to sensitive information, becomes embedded in critical workflows or acquires the ability to initiate actions, stronger controls become necessary. These can include more restrictive permissions, explicit action boundaries, continuous monitoring, logging, human approval for consequential decisions, rapid intervention mechanisms and tested recovery procedures (Kshetri, 2025; Radanliev, Santos and Ani, 2025).

Such proportionality also addresses an important practical problem in AI governance. Applying the strongest controls to every AI use case would create unnecessary organisational friction and could discourage legitimate experimentation. Conversely, applying lightweight governance to systems capable of materially affecting customers, financial resources, operational processes or sensitive information would expose the organisation to disproportionate risk. The objective is therefore not maximum control in every circumstance, but sufficient control relative to the potential consequences of AI behaviour.

The distinction between recommendation and action is particularly important. An AI system that recommends that an invoice be paid still leaves the organisation with an opportunity for human verification. An agent that can independently initiate the payment changes the control environment because the consequences of an erroneous or manipulated decision can occur before a human has intervened. Similarly, an AI system that recommends a configuration change presents a different risk from an agent that can directly modify production infrastructure. As autonomy increases, governance must move progressively from reviewing outputs towards controlling permissions, actions and system interactions.

This development also changes the meaning of human oversight. Human involvement cannot simply be treated as a formal approval step that automatically makes an AI system adequately governed. Research on human–AI decision-making shows that the effectiveness of AI participation depends on how humans and AI systems exchange information and how their respective contributions affect decision quality (Zercher et al., 2025). Meaningful oversight therefore requires that humans have sufficient information, competence, time and authority to understand and challenge AI-generated recommendations or actions. Where a human is nominally responsible but lacks the practical ability to review the AI's reasoning, evidence or proposed action, accountability may exist formally while control remains largely with the system.

The increasing autonomy of AI also creates a need for explicit boundaries around identity and permissions. AI agents should not automatically inherit broad privileges simply because the human who initiated their use possesses them. Instead, the authority granted to an agent should be determined by the task it is expected to perform, the information it requires and the potential consequences of its actions. This is consistent with the broader security principle of least privilege, but applied to AI-enabled actors as well as conventional users and applications. The resulting governance environment must make it possible to identify which AI system performed an action, under whose authority it operated, what information and tools it used, and what controls were in place at the time.

These requirements become more complex when agents interact with multiple external services. AI systems may depend on foundation-model providers, APIs, cloud platforms, specialised applications and third-party data sources. Such dependencies introduce additional risks concerning data handling, service availability, security and accountability (Radanliev, Santos and Ani, 2025). Governance must consequently extend beyond the internal configuration of the agent to include the external technological ecosystem on which its behaviour depends.

The transition towards agentic AI therefore changes governance from primarily a question of what AI may generate to a broader question of what AI may access, decide and do. This represents a significant expansion of organisational control. The organisation must understand not only the model and its outputs, but also the agent's identity, authority, permissions, tools, data access, workflow position and ability to affect the external environment.

This development is consistent with Gartner's (2026) view of AI governance as a maturity journey. Governance mechanisms that may be adequate during early experimentation become insufficient when AI systems are stabilised, scaled and integrated into important organisational processes. As adoption increases, governance needs to become more systematic, repeatable and embedded in organisational infrastructure rather than relying primarily on individual judgement or manual approval.

The resulting principle is straightforward: the greater the autonomy and consequence of an AI system, the stronger and more continuous the organisational controls surrounding it must become. Governance should therefore evolve with AI capability. The challenge is not to eliminate autonomy, but to ensure that autonomy remains bounded by clearly defined authority, proportionate permissions, effective monitoring, meaningful human accountability and the ability to intervene when necessary. In this sense, the governance of agentic AI is ultimately the governance of organisational action performed through AI.

4. AI Governance as a Maturity Journey

A central implication of contemporary AI governance research is that there is no single governance arrangement that is appropriate for every organisation or every AI application. Gartner (2026) explicitly frames AI governance as a maturity journey rather than a universal control model. Organisations initially need to establish ownership, define boundaries and identify unacceptable uses. As experimentation expands, governance needs to become more visible, repeatable and responsive. When AI becomes operationally significant, governance must increasingly be formalised, scaled and integrated into ordinary organisational management. At more advanced stages, governance should become sufficiently embedded that appropriate controls operate as part of normal organisational processes rather than as exceptional approval activities.

This maturity perspective is consistent with the broader academic literature. Batool, Zowghi and Bano (2025) demonstrate that AI governance encompasses different organisational and institutional levels, actors and lifecycle stages, while Papagiannidis, Mikalef and Conboy (2025) emphasise the need to translate responsible AI principles into structural, relational and procedural practices. These findings suggest that governance capability develops as organisations acquire greater experience with AI and as the technology becomes more deeply integrated into organisational activities. Governance is therefore not a static state that an organisation reaches once and then maintains unchanged.

The need for a maturity perspective is particularly evident during the early stages of AI adoption. Excessive governance can create unnecessary friction when organisations are still experimenting with relatively low-risk applications. If every use of generative AI requires extensive approval, documentation and review regardless of its potential consequences, employees may perceive formal governance as an obstacle rather than an enabler. They may consequently seek alternative tools or informal workarounds, increasing the risk of Shadow AI and reducing organisational visibility into AI usage (Silic, Silic and Kind-Trüller, 2025). Effective governance must therefore distinguish between experimentation that can be safely encouraged and applications that require stronger organisational control.

The opposite problem is equally significant. Organisations may establish a governance approach that is appropriate for experimentation but fail to strengthen it when AI becomes embedded in critical business processes. Research on organisational adoption shows that generative AI implementation is influenced not only by the characteristics of the technology but also by organisational capabilities, skills, complexity and the capacity to manage organisational change (Hughes et al., 2026). As AI moves from isolated experimentation into core workflows, governance must evolve accordingly. A lightweight experimental policy may be inadequate once an AI system processes sensitive information, influences consequential decisions or interacts directly with operational systems.

The resulting principle is therefore proportionality over uniformity. Governance intensity should reflect the characteristics and consequences of a particular AI use rather than the mere fact that AI is involved. An application that processes non-sensitive information, provides optional assistance and has no ability to execute actions presents a different governance requirement from an autonomous system that processes personal or commercially sensitive data, influences high-impact decisions or interacts directly with critical enterprise systems.

Several characteristics are particularly important when determining the appropriate level of control. The sensitivity of the data processed by an AI system determines the potential consequences of misuse or disclosure. The consequences of its decisions determine the organisational impact of erroneous or manipulated outputs. The degree of autonomy determines how much activity can occur without direct human intervention. The extent to which the system is integrated into critical workflows determines how widely a failure can propagate. Finally, the organisation's ability to detect, contain and recover from failures determines how much residual risk can be tolerated. These characteristics should be considered together rather than independently because the combination of high data sensitivity, significant autonomy and critical workflow integration can produce risks that are substantially greater than those associated with any individual characteristic.

The importance of proportionality becomes particularly clear with agentic AI. As agents acquire the ability to plan and execute sequences of actions, governance must increasingly address identity, permissions, authority, monitoring and intervention (Hosseini and Seilani, 2025; Kshetri, 2025). An agent operating within a tightly constrained environment may require relatively limited oversight, whereas an agent with access to sensitive information or authority to modify business-critical systems requires substantially stronger controls. Governance maturity therefore needs to reflect not only the sophistication of the underlying model but also the operational authority granted to the system.

Maturity should also be understood as an organisational capability rather than simply an accumulation of policies and controls. An organisation may possess extensive documentation while having limited visibility into actual AI usage or limited ability to intervene when an AI system behaves unexpectedly. Conversely, an organisation with relatively simple policies may demonstrate stronger governance if it can reliably identify AI deployments, assign accountability, enforce permissions, monitor material risks and intervene effectively. Batool, Zowghi and Bano (2025) and Papagiannidis, Mikalef and Conboy (2025) support this broader understanding by positioning governance as a combination of structures, processes, responsibilities and organisational practices rather than as a single formal artefact.

This also means that governance maturity cannot be assessed solely through periodic certification or policy reviews. AI environments are dynamic. Models are updated, providers modify their services, prompts and workflows change, new data sources are introduced and applications become connected to additional enterprise systems. Organisational use can therefore change even when the formal AI policy remains unchanged. Radanliev, Santos and Ani (2025) emphasise the evolving cybersecurity and resilience implications of generative AI, while Hughes et al. (2026) demonstrate that organisational adoption is shaped by changing technological and organisational conditions. Governance must consequently be capable of adapting as the systems and environments it governs change.

Continuous governance does not mean continuous manual approval. Mature governance should increasingly rely on mechanisms that make appropriate behaviour systematic and observable. Depending on the risk, these mechanisms may include automated policy enforcement, identity and access controls, monitoring, logging, data-protection controls, workflow restrictions and escalation mechanisms. The objective is to move governance from an episodic activity towards an embedded organisational capability. Gartner (2026) describes this progression as a movement towards increasingly systematic and integrated governance as AI adoption scales.

This perspective also changes how governance effectiveness should be evaluated. The relevant question is not simply whether an organisation has increased the number of AI policies, committees or approval processes. It is whether those mechanisms remain appropriate to the organisation's actual AI exposure. A governance process that is highly sophisticated but does not cover significant AI deployments provides less protection than a proportionate process that consistently governs the systems that matter most. Governance maturity should therefore be assessed in relation to organisational exposure, rather than by the volume of governance activity alone.

The relationship between AI capability and organisational control is consequently dynamic. When AI capability increases, organisational control must either increase with it or the organisation's governance gap will widen. This is particularly important as AI becomes integrated into decision-making and autonomous workflows. Schulte and Kanbach (2026) demonstrate the expanding range of organisational tasks in which generative AI can participate, while Zercher et al. (2025) show that the effects of AI participation on decision-making depend on how human and AI contributions are structured. As the organisational consequences of AI participation increase, governance must similarly become more deliberate about accountability, intervention and decision rights.

Maturity should therefore be understood as the ability to maintain an appropriate relationship between AI capability and organisational control over time. Early governance establishes boundaries and accountability. Developing governance makes AI use visible and repeatable. More mature governance embeds appropriate controls into workflows, architecture and management processes. Advanced governance becomes increasingly adaptive, allowing organisations to accommodate new AI capabilities without losing visibility, accountability or control.

The objective is consequently not to maximise governance. It is to establish sufficient and proportionate governance for the level of AI capability, autonomy and organisational consequence involved. Governance that is too weak creates uncontrolled exposure; governance that is unnecessarily rigid can drive AI use outside formal channels. The mature organisation is therefore not the one with the most governance, but the one that can continuously adjust its governance to the risks and opportunities created by changing AI capabilities while maintaining effective organisational control (Gartner, 2026; Batool, Zowghi and Bano, 2025).

5. From Shadow AI to Governed AI Adoption

One of the clearest indications that AI adoption can develop faster than organisational governance is the emergence of Shadow AI. Employees may use publicly available or externally provided AI systems outside formally approved organisational frameworks, often because these tools are easily accessible and can provide immediate benefits for productivity, analysis, communication or software development. Research by Silic, Silic and Kind-Trüller (2025) shows that such use can create risks involving data privacy, unreliable outputs, bias and governance drift. However, the significance of Shadow AI extends beyond employee compliance. It can also reveal a structural gap between the AI capabilities employees need and the governance mechanisms organisations have established to provide them.

This distinction is important because Shadow AI should not automatically be interpreted as evidence of irresponsible employee behaviour. Employees may adopt unsanctioned tools because formal processes are too slow, approved alternatives are unavailable, or existing governance does not accommodate legitimate business needs. The phenomenon can therefore provide organisations with information about where technological adoption is occurring faster than organisational decision-making. Treating every instance exclusively as a policy violation may remove the visibility needed to understand why employees are turning to external AI services in the first place (Silic, Silic and Kind-Trüller, 2025).

A purely restrictive response can consequently be counterproductive. Where AI tools provide significant productivity benefits and can be accessed with minimal friction, prohibition does not necessarily eliminate demand. It may instead shift AI use further outside organisational visibility. This creates a paradox in which stronger formal restrictions can coexist with weaker practical control. Effective governance should therefore seek to bring legitimate AI use into controlled organisational environments rather than assuming that all unapproved use can simply be eliminated.

This requires organisations to distinguish between uses that are unacceptable, uses that can be permitted under defined conditions and uses that should be actively supported through approved organisational services. Low-risk experimentation may be accommodated within relatively simple boundaries, while applications involving sensitive information, consequential decisions or critical business processes require stronger controls. Such differentiation is consistent with the broader principle of proportionality in AI governance and with Gartner's (2026) view that governance should evolve according to the maturity and scale of organisational AI adoption.

The first requirement for such an approach is visibility. Organisations need to understand which AI services are being used, by whom, for what purposes and in which business processes. They also need to understand what information is being processed and whether AI services are connected to internal systems or external data sources. Without this visibility, governance remains partly theoretical because the organisation cannot reliably determine the scope of its AI exposure. Batool, Zowghi and Bano (2025) emphasise the importance of understanding who is responsible for AI governance and what is subject to governance across the AI lifecycle. Shadow AI makes these questions particularly difficult because the organisation may not even have complete knowledge of where AI is being used.

Visibility also creates a direct connection between AI governance and data governance. The relevant question is not simply whether a particular AI service has been approved. It is whether a particular type of information may appropriately be processed by that service for a particular purpose. An AI application may be acceptable for analysing publicly available information while being inappropriate for processing confidential business information, personal data, intellectual property or commercially sensitive material. Governance therefore needs to consider the relationship between the characteristics of the information, the capabilities and terms of the AI service, and the purpose for which the information is being processed.

Reliable data classification is consequently an important foundation for governed AI adoption. If organisations cannot distinguish between public, internal, confidential and highly sensitive information, they cannot consistently determine which data may be transferred to external AI services. Data governance must therefore provide sufficient information about the sensitivity, ownership and permitted use of data for AI-related decisions to be meaningful. This reinforces the broader argument that AI governance cannot be isolated from established data-management capabilities.

Identity and access management create a second essential connection. The governance question is not only what information an AI system may process but also who or what is authorised to use it. This becomes increasingly important as AI moves from individual employee interactions towards integrated enterprise applications and agentic systems. Where AI can access internal resources, governance must establish the identity under which it operates, the permissions it receives and the actions it can perform. Kshetri (2025) highlights the growing importance of identity and oversight as AI systems become more autonomous, while Hosseini and Seilani (2025) demonstrate that agentic AI introduces new questions around system autonomy and interaction with organisational environments.

Monitoring provides a third component of effective control. Policies and access restrictions establish intended behaviour, but organisations also need mechanisms to determine whether actual behaviour conforms to those expectations. Monitoring can provide visibility into AI usage, unusual activity, inappropriate data transfers, policy violations and changes in how AI systems are being used. Radanliev, Santos and Ani (2025) emphasise the evolving cybersecurity and resilience implications of generative AI, reinforcing the need to treat AI usage as part of the organisation's broader security and resilience environment rather than as an isolated productivity technology.

These requirements demonstrate why AI governance increasingly depends on capabilities that have traditionally been associated with cybersecurity, privacy, data governance and enterprise architecture. Cybersecurity contributes identity management, access control, monitoring and threat detection. Data governance provides classification, ownership and rules concerning the use and disclosure of information. Privacy governance establishes requirements for the processing of personal information. Enterprise architecture determines how AI services connect to applications, data and business processes. Procurement and vendor management become relevant where external AI providers process organisational information or provide critical AI capabilities.

The organisational implication is that AI governance should not become another isolated control function. Instead, it should connect existing capabilities around a common question: can the organisation maintain appropriate control over how AI accesses information, interacts with systems and influences organisational activity? This requires collaboration between technology, security, privacy, data, procurement, legal and business functions rather than assigning AI governance exclusively to an AI ethics committee or technical team (Batool, Zowghi and Bano, 2025; Papagiannidis, Mikalef and Conboy, 2025).

This integrated approach also changes how Shadow AI should be managed. The objective is not simply to discover unauthorised tools and remove them. Organisations should use evidence of Shadow AI to identify unmet business requirements, recurring use cases and weaknesses in the approved AI environment. Where employees repeatedly adopt a particular external capability, this may indicate that the organisation should provide an approved alternative, establish appropriate data boundaries or redesign the relevant workflow. Governance can therefore turn Shadow AI from a purely defensive problem into a source of organisational intelligence about emerging patterns of AI adoption.

The transition from Shadow AI to governed AI adoption consequently depends on making AI use both visible and governable. Visibility without appropriate controls merely reveals risk, while controls without visibility cannot reliably be applied to the full AI environment. Effective governance requires both: knowledge of where AI is being used and the organisational, technological and procedural mechanisms necessary to influence how it is used.

This reinforces the broader argument that AI governance is fundamentally a question of organisational control. The organisation does not need to control every AI interaction to the same degree. It needs to understand where material risks arise and apply proportionate controls to the data, identities, systems and workflows involved. As AI adoption becomes more widespread and autonomous, the ability to integrate these capabilities will become increasingly important to maintaining both innovation and accountability.

6. Human Judgement in the Age of AI

As AI systems become more capable, the role of human judgement does not necessarily diminish. In high-consequence contexts, it can become more important because AI systems increasingly participate in decisions whose consequences extend beyond the technical system itself. The objective of governance should therefore not be to retain humans merely as symbolic approvers, but to ensure that human judgement is exercised where contextual knowledge, accountability, ethical considerations and organisational responsibility provide genuine value.

This distinction is particularly important because AI can influence decision-making without formally making the final decision. An AI system may determine which information a decision-maker sees, summarise evidence, identify potential options, rank alternatives or recommend a preferred course of action. Even when a human retains formal authority, these forms of AI participation can substantially shape the decision process. Schulte and Kanbach (2026) demonstrate the increasingly diverse roles that generative AI can perform across organisational decision-making, extending from attention and information processing through design and choice to implementation and feedback. AI therefore has the potential to influence the decision cycle at multiple stages rather than simply providing information to a human decision-maker.

Recent empirical research illustrates both the opportunities and limitations of such human–AI collaboration. Zercher et al. (2025) examined teams working with AI under different knowledge configurations and found that AI could improve decision performance where it helped address information asymmetries, but that AI could also introduce new asymmetries and collaboration difficulties. These findings are significant for governance because they demonstrate that simply adding AI to a decision process does not guarantee better decisions. The effects depend on how knowledge is distributed between humans and AI, how the interaction is structured and whether participants can effectively evaluate and integrate AI contributions.

This challenges a simplistic interpretation of human oversight. Human oversight should not be understood as a binary condition in which a person either approves or rejects an AI-generated recommendation. The presence of a human at the end of a process does not necessarily mean that the human retains meaningful control. If the AI system determines the relevant evidence, frames the available options and presents a highly authoritative recommendation, the human may effectively become a confirmation mechanism rather than an independent decision-maker.

Meaningful human oversight therefore requires more than formal involvement. The human decision-maker needs sufficient information to understand the basis of the AI recommendation, sufficient expertise to identify potential weaknesses, sufficient time to assess the implications and sufficient authority to reject or modify the recommendation. The quality of oversight consequently depends on the design of the organisational decision process as much as on the technical characteristics of the AI system.

This is particularly important in high-consequence decisions. Governance should establish what evidence supports an AI recommendation, what information the system relied upon, what limitations or uncertainties are known, whether the decision can be reversed and what consequences would arise if the AI system were wrong. It should also establish who has the authority to override the system and under what circumstances intervention is mandatory. These questions transform human oversight from a procedural requirement into an organisational control mechanism.

The reversibility of decisions is especially important when determining the appropriate level of human involvement. An AI system recommending a low-impact, easily reversible action presents a different governance challenge from one making recommendations that could materially affect customers, employees, financial resources or legal obligations. Where consequences are difficult to reverse, governance should generally require stronger human scrutiny and clearer decision rights. This is consistent with the broader principle that AI governance should be proportionate to the consequences and autonomy associated with a particular use case (Gartner, 2026).

The issue becomes even more significant as AI systems move from recommendation towards execution. An AI assistant that proposes an action gives a human an opportunity to intervene before implementation. An agent that can independently execute the action changes the nature of the control environment. In such circumstances, human oversight may need to occur through predefined authority boundaries, approval thresholds, monitoring and intervention mechanisms rather than through manual review of every individual action (Hosseini and Seilani, 2025; Kshetri, 2025). Human control can therefore be designed into the architecture of an AI-enabled process rather than relying exclusively on continuous manual supervision.

This also suggests that human judgement should be understood as a scarce organisational resource rather than a universal control applied to every AI interaction. Requiring human review for every low-risk AI-generated output would create unnecessary organisational friction and potentially reduce the value of automation. Conversely, removing meaningful human judgement from decisions with significant consequences could transfer excessive authority to systems whose limitations and uncertainties may not be fully understood. Effective governance should therefore determine where human judgement adds material value and ensure that it is concentrated at those points.

The challenge is compounded by the fact that AI-generated outputs can appear authoritative even when their underlying reasoning or information is incomplete. Radanliev, Santos and Ani (2025) highlight the broader cybersecurity and resilience implications associated with generative AI, including risks arising from system behaviour, data and organisational dependencies. Governance should therefore avoid assuming that human reviewers will automatically identify AI failures. Review processes need to provide humans with sufficient contextual information and opportunities for challenge rather than simply placing responsibility on them after the AI has produced an outcome.

Human accountability must consequently remain explicit even where AI performs substantial parts of a decision process. The organisation should be able to identify who is responsible for defining the decision criteria, who authorised the AI system, who is accountable for its deployment and who has authority to intervene when its behaviour becomes problematic. Batool, Zowghi and Bano (2025) emphasise the importance of clearly understanding who is responsible for AI governance, while Papagiannidis, Mikalef and Conboy (2025) underline the importance of embedding responsible AI practices into organisational structures and processes. AI participation should therefore not create ambiguity about where organisational accountability resides.

The governance implication is that organisations should avoid treating “human in the loop” as sufficient evidence of responsible AI. The relevant question is not whether a human is formally present, but whether the human retains meaningful authority and whether the organisational process enables informed intervention. Meaningful human control requires an appropriate combination of information, competence, decision rights, time, reversibility and escalation mechanisms.

As AI becomes increasingly integrated into organisational decision-making, this distinction will become more important. The objective of responsible AI governance is not to preserve human involvement for its own sake, nor to maximise automation. It is to determine where human judgement is necessary, where AI can appropriately augment that judgement and where organisational controls must prevent AI from exercising authority beyond its intended boundaries. In this sense, effective AI governance does not place humans merely in the loop; it ensures that humans remain meaningfully in control where organisational accountability and consequence require it.

7. AI Architecture, Data and Workflow Integration

AI governance cannot be separated from the architecture through which AI operates. An AI application is not simply a model producing an output. It is a socio-technical system in which models interact with data, prompts, interfaces, APIs, identity mechanisms, tools, users, workflows and external service providers. The governance implications therefore arise not only from the characteristics of the underlying model, but also from the environment in which the model is deployed and the actions that environment enables.

This distinction becomes particularly important with agentic AI. An agent may rely on a foundation model supplied by one provider, retrieve organisational information from another system, invoke tools through APIs, interact with enterprise applications and ultimately influence or execute actions within a business process. Hosseini and Seilani (2025) demonstrate that agentic AI increasingly involves autonomous interaction with complex technological environments, while Kshetri (2025) highlights the associated importance of identity, security and oversight. Governance of the underlying model alone therefore provides incomplete assurance because the consequences of AI behaviour depend substantially on the systems, data and permissions surrounding it.

The relevant unit of governance increasingly becomes the AI-enabled workflow rather than the model in isolation. An AI system embedded within a customer-service process, financial operation, software-development pipeline or decision-making process should be governed according to how it participates in that workflow and what consequences can result from its actions. This requires understanding not only what the AI model is capable of producing, but also where its outputs go, which systems consume them, what decisions they influence and whether subsequent actions can occur automatically.

This perspective is consistent with recent research on organisational decision-making. Schulte and Kanbach (2026) show that generative AI can participate across multiple stages of decision processes, including attention, intelligence, design, choice, implementation and feedback. AI therefore increasingly becomes part of the process through which organisational outcomes are produced rather than simply an external tool providing information to employees. Governance must consequently address the complete chain through which AI contributes to an organisational outcome.

The relationship between AI architecture and organisational value is similarly important. Sarkar (2026) argues that the strategic value of generative AI depends increasingly on the integration of AI into organisational workflows rather than on isolated experimentation with individual tools. Hughes et al. (2026) likewise demonstrate that organisational adoption depends on technological, organisational and environmental conditions, including organisational capabilities and the ability to manage change. AI pilots may therefore demonstrate technical feasibility without producing substantial organisational value if they remain disconnected from enterprise data, established processes and decision structures.

Workflow integration, however, creates a governance paradox. The more deeply AI is embedded into organisational processes, the greater the potential value but also the greater the consequences of failure. An isolated AI assistant may produce an incorrect recommendation that a user simply ignores. An AI system embedded within a critical operational workflow may propagate the same error across multiple transactions or decisions before the problem is detected. An agent capable of executing actions can create consequences even more rapidly. AI governance must therefore develop alongside the degree of workflow integration rather than treating integration as a purely technical or productivity consideration.

This creates a direct relationship between integration, consequence and control. As AI becomes more deeply integrated into organisational processes, organisations need greater visibility into the data it accesses, the decisions it influences and the actions it can initiate. They also need stronger mechanisms for restricting inappropriate behaviour and detecting deviations from expected behaviour. The governance requirements of an AI-enabled workflow should therefore be determined partly by its operational significance and potential consequences, rather than solely by the sophistication of the AI technology involved.

This has an important architectural implication: control should be embedded where AI operates. Governance should not depend exclusively on policies or manual approval processes positioned outside the technical and operational environment. Where sensitive data is involved, appropriate access restrictions should be enforced through data and identity mechanisms. Where AI agents can invoke tools or applications, their permissions should be constrained according to the tasks they are authorised to perform. Where actions can have significant consequences, appropriate approval thresholds, verification mechanisms or intervention points should be incorporated into the workflow itself. Where AI contributes to consequential decisions, the relevant inputs, outputs, interventions and decisions should be sufficiently traceable to support accountability and subsequent review.

This principle also applies to monitoring. AI systems operate within environments that can change continuously. Models may be updated, data sources may change, applications may be reconfigured and workflows may evolve. Periodic governance reviews cannot therefore provide complete assurance. Radanliev, Santos and Ani (2025) emphasise the changing cybersecurity and resilience implications of generative AI, reinforcing the importance of ongoing monitoring and the ability to respond when AI behaviour or its surrounding environment changes.

Embedding governance into architecture does not mean that every AI interaction should be subjected to intrusive technical controls. Rather, architectural governance should make appropriate controls proportional to the risks associated with a particular workflow. Low-risk applications may require basic visibility, authentication and monitoring, whereas systems that process sensitive information or execute consequential actions may require stronger identity controls, restricted permissions, detailed logging, continuous monitoring and explicit human intervention. This approach is consistent with Gartner's (2026) broader maturity perspective, in which governance becomes increasingly systematic and embedded as AI adoption scales.

The architectural perspective also changes the role of data governance. Data should not be treated merely as an input to an AI model. Its classification, ownership, provenance, permitted use and sensitivity determine what an AI system should be allowed to do with it. If an AI agent can retrieve information from multiple organisational repositories, governance must establish whether those sources can legitimately be combined and whether the resulting information can be used for the intended purpose. This makes data governance a direct component of AI control rather than a separate administrative function.

External providers introduce another architectural dimension. AI-enabled workflows may depend on foundation-model providers, cloud platforms, APIs, specialised applications and third-party data services. Radanliev, Santos and Ani (2025) emphasise the broader cybersecurity and resilience implications of generative AI, including dependencies and emerging attack surfaces. Organisations therefore need to understand not only their internal AI architecture but also the external services on which critical AI-enabled workflows depend. Where such dependencies become operationally significant, supplier governance, security assurance and continuity considerations become part of AI governance.

The resulting perspective represents a significant shift from governance as documentation towards governance as architecture. Documentation remains necessary because policies establish organisational intent, responsibilities and boundaries. However, governance becomes materially stronger when those expectations are translated into the technical and operational mechanisms through which AI actually operates. A policy can state that an AI agent should not access confidential information, but architectural controls are required to enforce that boundary. A policy can require human approval for a high-impact decision, but the workflow must provide an effective intervention point. A policy can require accountability, but systems must generate sufficient records to establish what happened and who authorised it.

AI governance therefore increasingly becomes part of enterprise architecture and process design rather than an approval layer applied after technology has been selected or implemented. Papagiannidis, Mikalef and Conboy (2025) emphasise the importance of embedding responsible AI governance into organisational structures and processes, while Batool, Zowghi and Bano (2025) demonstrate that effective governance spans multiple actors and stages of the AI lifecycle. These findings support the view that governance should be incorporated from the beginning of AI-enabled system design rather than treated as a subsequent compliance activity.

The strategic consequence is that organisations should design AI-enabled workflows with governance requirements already embedded in their architecture. Data access, identity, permissions, decision rights, monitoring, intervention and auditability should be considered part of the system itself. As AI becomes more integrated and autonomous, the distinction between the technology being governed and the mechanisms through which it is governed becomes increasingly blurred.

The result is a fundamental change in the nature of AI governance. Governance should not merely describe how an organisation intends AI to behave; it should help determine what AI is technically and organisationally able to do. When governance is embedded in architecture, organisational principles become enforceable boundaries, accountability becomes traceable, and risk management becomes part of the operation of AI-enabled workflows rather than a separate activity performed around them.

8. Cybersecurity and AI Resilience

AI introduces cybersecurity risks of its own while simultaneously changing the broader cyber-risk environment in which organisations operate. Generative AI can create new attack surfaces, facilitate inappropriate data disclosure and introduce vulnerabilities associated with model behaviour, external services, integrations and uncontrolled use. As AI becomes more deeply embedded in organisational workflows, these risks can no longer be treated as separate from conventional cybersecurity. Radanliev, Santos and Ani (2025) argue that responsible deployment of generative AI requires cybersecurity and resilience to be considered throughout the AI lifecycle, encompassing design, implementation, monitoring, compliance and feedback. This perspective reinforces the need to treat AI governance as an ongoing security and resilience capability rather than as a one-time assessment of an AI system.

The cybersecurity implications of AI extend beyond attacks against the underlying model. Risk can arise from the data supplied to an AI system, the interfaces through which it connects to other systems, the permissions assigned to it, the external services on which it depends and the ways in which employees use it. Shadow AI can create unmanaged data flows, while agentic systems can create additional exposure because they may have the authority to interact with enterprise applications and external services (Silic, Silic and Kind-Trüller, 2025; Hosseini and Seilani, 2025). AI governance must therefore consider the entire environment surrounding an AI system rather than treating the model as the sole security boundary.

This reinforces a broader principle: AI governance should not focus exclusively on preventing undesirable behaviour. It must also prepare the organisation for governance failure. No AI system can be assumed to operate correctly under all circumstances. Models can produce incorrect or misleading outputs, data can be disclosed inappropriately, external services can become unavailable, integrations can fail and system behaviour can change following updates. AI systems can also be manipulated through malicious inputs or compromised surrounding infrastructure. Radanliev, Santos and Ani (2025) emphasise the importance of considering cybersecurity and resilience throughout the AI lifecycle precisely because responsible AI deployment must account for changing and potentially adverse conditions.

The possibility of failure becomes particularly significant when AI participates in consequential organisational processes. An incorrect recommendation in an informal productivity task may have limited consequences. The same failure in a financial, operational, customer-facing or security-critical workflow can propagate rapidly and affect multiple downstream activities. The governance question must therefore extend beyond whether an AI system performs acceptably under normal conditions to whether the organisation can contain its effects when it does not.

This introduces the concept of bounded failure. Organisations cannot necessarily guarantee that an AI system will never produce an incorrect result or behave unexpectedly. They can, however, design the surrounding environment so that errors are detected, constrained and recoverable. Appropriate boundaries may include restricted data access, limited permissions, transaction thresholds, approval requirements, monitoring, anomaly detection and mechanisms for suspending or isolating an AI capability. The objective is to prevent an individual AI failure from becoming an uncontrolled organisational failure.

Monitoring and logging are therefore central components of resilient AI governance. Organisations need sufficient visibility to determine what an AI system did, which information it accessed, which tools or services it invoked and what decisions or actions followed. This is particularly important when AI systems participate in complex workflows because an incident may otherwise be difficult to reconstruct. Auditability is consequently not merely a compliance requirement; it is also a prerequisite for effective incident response, investigation and organisational learning.

Access control provides another essential resilience mechanism. The authority granted to an AI system should be limited to what is necessary for its intended function. This principle becomes increasingly important as agents acquire the ability to access multiple systems or execute actions on behalf of users (Kshetri, 2025). Restricting permissions limits the potential impact of compromised, manipulated or malfunctioning AI systems. In this respect, identity and access controls become mechanisms for containing AI failure rather than merely protecting systems against unauthorised access.

Resilience also requires fallback mechanisms. Organisations should determine what happens when an AI service becomes unavailable, produces unreliable outputs or must be suspended following a security incident. Critical workflows may require alternative processes, manual operating procedures or substitute technologies. This is particularly relevant where organisations depend on external AI providers because service outages, changes in provider behaviour or contractual disruptions can affect the availability of AI-enabled business processes. Radanliev, Santos and Ani (2025) emphasise the importance of resilience in the context of evolving AI-related dependencies and cybersecurity risks.

The ability to stop an AI system safely therefore becomes a governance capability in its own right. An organisation should know under what circumstances an AI application, model or agent can be restricted, suspended or disconnected from the systems with which it interacts. For highly autonomous systems, this capability should not depend exclusively on an ad hoc managerial decision during a crisis. Appropriate intervention mechanisms should be designed into the system and organisational process from the outset.

Safe shutdown is particularly important for agentic AI because autonomy can allow an error to propagate across multiple actions. An agent that is generating an incorrect recommendation can be ignored by a user; an agent that is repeatedly executing incorrect actions may amplify the consequences before a human recognises the problem. Governance must therefore establish intervention thresholds and mechanisms that allow potentially harmful activity to be contained rapidly (Hosseini and Seilani, 2025; Kshetri, 2025).

Human intervention nevertheless remains only one element of resilient governance. Organisations also need clearly defined incident-response responsibilities, escalation procedures and recovery processes. When an AI-related incident occurs, responsibility may cross organisational boundaries involving security, data protection, legal, technology, business operations, procurement and senior management. Governance must therefore establish not only how an AI system can be technically restricted, but also who has authority to initiate intervention and how the organisation will coordinate its response.

This connects AI resilience with the broader concept of organisational cyber resilience. Neri, Niccolini and Virili (2026) conceptualise cyber resilience as an organisational capability rather than a purely technical property. From this perspective, resilience includes the ability to anticipate and absorb disruption, respond effectively, recover operations and adapt based on experience. Applied to AI, this means that governance should address not only the prevention of AI-related incidents but also the organisation's ability to continue operating when AI capabilities become unreliable or unavailable.

Resilient AI governance should therefore include mechanisms for learning after failure. Incidents, near misses and unexpected AI behaviour can reveal weaknesses in prompts, data, permissions, workflows, monitoring or organisational decision-making. Governance should provide feedback mechanisms through which these lessons can influence system configuration, policies, training and future AI deployments. This creates a continuous relationship between monitoring, incident response and governance improvement rather than treating each incident as an isolated technical problem.

The resulting objective is not to build AI systems that never fail. Such an objective would be unrealistic given the complexity and changing nature of AI technologies and their surrounding environments. The objective is to ensure that failures are detectable, bounded, interruptible and recoverable. Organisations should know when AI behaviour has departed from expectations, have mechanisms to limit the resulting consequences, retain the ability to intervene and maintain alternative ways of operating when necessary.

This perspective also changes how AI governance maturity should be assessed. A mature organisation should not merely demonstrate that an AI system performs adequately under normal conditions. It should be able to demonstrate what happens when the system produces an unreliable output, loses access to a dependency, exposes information, becomes compromised or must be taken offline. The ability to answer these questions provides a stronger indication of organisational control than the existence of policies or model-validation documentation alone.

AI governance is therefore inseparable from cybersecurity and resilience. As AI becomes more autonomous and deeply integrated into organisational workflows, the consequences of failure increase and the need for containment, intervention and recovery becomes more important. The central governance objective is consequently not perfect prevention, but controlled and recoverable operation under adverse conditions. In a mature AI-enabled organisation, resilience is demonstrated by the ability to continue exercising organisational control even when the AI system itself cannot be fully trusted.

9. From Model Governance to Dependency Governance

Another consequence of enterprise AI adoption is the expansion of organisational dependencies. AI systems increasingly rely on external foundation-model providers, cloud infrastructure, software components, data providers, APIs and specialised AI services. An organisation may therefore control the application through which employees interact with AI while having limited control over the underlying technologies and services on which that application depends. As AI becomes embedded in critical workflows, these external dependencies become an increasingly important component of organisational risk.

This changes the scope of AI governance. Traditional model governance focuses primarily on the lifecycle of an AI system within the organisation, including its design, development, validation, deployment and monitoring. However, an enterprise AI application may incorporate models and services that are developed, operated or modified by third parties. Governance that considers only the internal application therefore provides incomplete assurance. Organisations also need to understand the external technologies and providers on which AI-enabled processes depend.

The issue resembles broader cybersecurity challenges associated with cloud computing and software supply chains. Modern organisations rarely operate technology entirely within their own boundaries. Instead, they depend on interconnected providers whose services can affect the availability, security and behaviour of internal systems. Radanliev, Santos and Ani (2025) highlight the evolving cybersecurity and resilience implications associated with generative AI, including the broader ecosystem of technologies and services through which AI is deployed. As these dependencies become more critical, AI governance must consequently extend beyond the model lifecycle towards dependency governance.

Dependency governance requires organisations to understand the external components that are necessary for an AI-enabled process to function. This includes the foundation model, hosting environment, APIs, data sources, supporting software and other services involved in delivering the capability. The objective is not necessarily to eliminate external dependencies. In many cases, using specialised external providers is economically and technically rational. The governance challenge is to understand the dependency sufficiently to determine its importance, associated risks and available alternatives.

This is particularly important because changes by external providers can alter an AI system without any corresponding change within the organisation's own application. A provider may update a foundation model, modify service characteristics, change pricing or usage conditions, alter data-processing practices, introduce new functionality or discontinue a service. Such changes can affect system behaviour, performance, security, cost or compliance. AI governance must therefore recognise that the effective lifecycle of an enterprise AI system may extend beyond the organisation's direct control.

Procurement consequently becomes an important component of AI governance. Contracts with AI providers can determine how organisational data is processed and retained, whether data may be used for provider-side model improvement, how security incidents are communicated, what audit or assurance rights are available, what service levels apply and what happens when the relationship ends. These contractual conditions can materially affect organisational risk. Papagiannidis, Mikalef and Conboy (2025) emphasise the importance of organisational structures and processes in responsible AI governance, while Batool, Zowghi and Bano (2025) demonstrate that AI governance involves multiple organisational actors and governance levels. Procurement and vendor management should therefore not be treated as peripheral administrative functions when AI services become strategically important.

Architecture creates a further dependency consideration. Organisations that build critical workflows around proprietary interfaces or provider-specific capabilities may create substantial switching costs. Data, prompts, workflow logic, integrations and application behaviour may become tightly coupled to a particular provider. The resulting dependency can reduce strategic flexibility and make it difficult to change providers even when security, cost, performance or regulatory conditions change. Governance must therefore consider not only whether a provider is acceptable today, but also the degree of strategic dependency created by the architectural decision.

This introduces the concept of optionality into AI governance. Optionality refers to the organisation's ability to change, replace, restrict or suspend an AI dependency without unacceptable disruption to critical operations. It does not necessarily require maintaining multiple providers for every AI capability. Rather, it requires organisations to understand where dependency has become critical and whether credible alternatives or fallback processes exist. For low-consequence applications, significant provider dependency may be entirely acceptable. For business-critical workflows, the inability to replace or operate without a particular AI service may represent a material resilience risk.

The importance of optionality becomes particularly clear when considering service disruption. If an AI provider becomes unavailable, an organisation needs to know which processes will be affected, whether alternative services exist, how quickly they can be activated and whether operations can continue without AI. Neri, Niccolini and Virili (2026) emphasise that organisational cyber resilience involves the capacity to withstand disruption, recover operations and adapt. Applied to AI dependencies, this means that resilience requires organisations to understand not only how AI contributes to normal operations but also how those operations can continue when the AI dependency fails.

Dependency governance also needs to consider concentration risk. If multiple business processes depend on the same foundation-model provider, cloud environment or AI platform, a single provider incident may affect multiple organisational capabilities simultaneously. What appears to be a collection of independent AI applications may therefore represent a common underlying dependency. The organisation's exposure is determined not only by the individual risk of each application but also by the extent to which critical processes share common providers and infrastructure.

This is particularly relevant for agentic AI. Agents may depend on multiple services simultaneously, including foundation models, retrieval systems, APIs, enterprise applications and external tools. A failure in any one component can affect the agent's ability to complete its task or can change the way it behaves. Hosseini and Seilani (2025) highlight the increasing complexity of agentic systems, while Kshetri (2025) emphasises the importance of security and oversight as autonomous systems interact with broader technological environments. Dependency governance must therefore account for the complete service chain through which an agent operates.

The governance question should consequently extend beyond whether an AI provider is formally compliant. Compliance and assurance remain important, but they provide only one dimension of the relationship. An organisation may work with a provider that satisfies its stated security and regulatory requirements while still creating an excessive operational dependency. Conversely, a provider with a strong technical offering may be unsuitable for a critical process if the organisation has no viable recovery or exit strategy.

This suggests a broader governance question: what organisational dependency are we creating, and what happens if that dependency becomes unavailable, compromised or materially changes? Answering this question requires organisations to understand the criticality of AI-enabled workflows, the providers supporting them, the data and systems involved, the availability of alternatives and the consequences of interruption.

AI governance should therefore incorporate dependency assessment alongside model assessment. Before deploying AI into a critical workflow, organisations should understand not only whether the technology performs adequately, but also how deeply the workflow will depend on the provider, how the relationship can be monitored, what contractual protections exist and how the organisation could recover if the service were disrupted. The greater the operational importance of the AI capability, the more important these considerations become.

This perspective also reinforces the relationship between AI governance and organisational resilience. A resilient organisation does not necessarily avoid external dependencies. It understands them, prioritises them and prepares for their failure. Radanliev, Santos and Ani (2025) and Neri, Niccolini and Virili (2026) support this broader understanding of resilience as the ability to manage changing technological and operational conditions rather than simply preventing individual incidents.

The transition from model governance to dependency governance therefore represents another expansion of organisational control. The organisation must govern not only what its AI systems do, but also the external technological relationships that make those systems possible. Procurement, architecture, vendor management, security, data governance and business continuity consequently become interconnected components of AI governance.

The fundamental question is therefore no longer simply “Is this AI provider compliant?” It is “What dependency are we creating, how critical is it, and can we remain operational if that dependency fails or changes?” Organisations that can answer these questions are better positioned to scale AI without allowing technological convenience to become uncontrolled strategic dependency.

10. AI Governance and Organisational Value

AI governance is often presented primarily as a constraint on innovation. Such a view, however, treats governance as a mechanism for preventing undesirable outcomes rather than as an organisational capability that can enable responsible adoption. Effective governance can create predictable boundaries within which employees and business units can experiment, evaluate use cases and scale successful applications. The objective is therefore not to maximise control at the expense of innovation, but to establish sufficient control to allow AI to be adopted with confidence.

This distinction becomes increasingly important as organisations move from isolated experimentation towards the systematic integration of AI into business processes. Generative AI adoption is influenced not only by technological capability but also by organisational conditions, including skills, resources, organisational readiness and the capacity to manage change (Hughes et al., 2026). Sarkar (2026) similarly argues that the organisational value of generative AI depends on how effectively it is integrated into actual workflows and functions rather than simply made available as a general-purpose technology. AI governance therefore has an important role in creating the organisational conditions under which adoption can translate into measurable value.

Organisations need clarity about which AI uses are permitted, what types of data may be processed, which providers can be used, when human review is required and what evidence is necessary before an AI application can move into production. Without such boundaries, employees and business units may face uncertainty about what they are authorised to do. This uncertainty can encourage inconsistent practices, duplicated risk assessments or the use of unapproved systems. As Silic, Silic and Kind-Trüller (2025) demonstrate in their analysis of Shadow AI, employees may adopt AI tools outside formal organisational processes when governance does not adequately accommodate emerging technological demand. Governance that provides clear and usable pathways for legitimate adoption can therefore reduce rather than increase organisational friction.

This also changes how governance should be understood in relation to organisational value. Governance should not simply ask whether an AI application is technically compliant. It should also consider whether the governance process is proportionate to the value and risk of the proposed use. A low-risk productivity application that processes non-sensitive information may require relatively limited scrutiny, whereas an AI system embedded in a consequential business decision may require extensive validation, monitoring and human oversight. Batool, Zowghi and Bano (2025) emphasise the multidimensional nature of AI governance, while Gartner (2026) similarly presents governance as a maturity journey in which organisational controls develop alongside AI adoption and capability.

The principle of proportionality is therefore central. Excessive governance can become a barrier to legitimate experimentation if every AI application is subjected to the same approval requirements regardless of its risk. At the other extreme, insufficient governance can allow AI systems to become embedded in critical processes without adequate evidence of reliability, accountability or resilience. The objective is not maximum governance but appropriate governance.

This is particularly important because AI adoption does not automatically generate organisational value. Generative AI may increase productivity, support decision-making or improve access to organisational knowledge, but these benefits depend on how AI is incorporated into existing work. Sarkar (2026) emphasises the importance of moving from AI experimentation towards workflow-level integration, while Hughes et al. (2026) show that successful organisational adoption depends on technological, organisational and environmental conditions. AI governance must therefore operate alongside change management, skills development, workflow redesign and organisational learning rather than being treated as a separate compliance activity.

The relationship between governance and value is also evident in decision-making. Generative AI can influence how information is gathered, interpreted and evaluated, potentially changing the distribution of cognitive work between employees and AI systems. Schulte and Kanbach (2026) identify emerging roles for generative AI across organisational decision-making processes, while Zercher et al. (2025) show that collaboration between teams and generative AI can affect both decision processes and decision quality. Governance therefore has to consider not only whether an AI system produces an acceptable output, but also how its use changes the process through which organisational decisions are reached.

This reinforces the importance of defining the purpose and boundaries of AI use. Governance should clarify where AI can accelerate work, where human judgement remains necessary and where AI involvement requires additional evidence or approval. Such boundaries can give employees greater confidence in using AI appropriately because they reduce ambiguity about acceptable behaviour. In this sense, governance can function as an infrastructure for responsible experimentation rather than simply as a mechanism for restricting it.

The same principle applies to procurement and technology selection. Organisations that establish approved providers, security requirements, data-handling expectations and minimum assurance criteria can reduce the uncertainty associated with evaluating every AI service from first principles. Dependency governance therefore complements value-oriented AI governance: organisations can move more quickly when they have already established trusted mechanisms for assessing providers and managing common risks. This supports the broader argument that governance becomes more valuable when it is embedded into organisational processes rather than implemented as a separate approval layer.

The appropriate objective can therefore be described as risk-adjusted speed. Low-risk applications should be capable of moving rapidly through proportionate governance processes, allowing experimentation and learning without unnecessary administrative delay. Higher-risk applications should encounter stronger requirements for testing, validation, human oversight, monitoring and resilience before they become operationally significant. The speed of adoption should consequently reflect the potential consequences of failure.

For example, an AI tool used to summarise non-sensitive internal documents may require relatively limited controls. By contrast, an AI system involved in financial transactions, employee decisions, sanctions screening or other high-consequence processes should require substantially greater scrutiny. The difference is not that one application requires governance and the other does not. Both require governance, but the intensity and nature of control should reflect the consequences associated with their use.

This approach also provides a more useful way of evaluating governance effectiveness. A governance process should not be judged solely by the number of policies it produces or the number of applications it prevents from being deployed. It should also be assessed by whether it enables legitimate AI use to proceed safely, whether employees understand the boundaries of acceptable use, whether risks are identified early and whether successful applications can move from experimentation into sustainable operational use. Gartner (2026) frames AI governance as an evolving organisational journey, reinforcing the idea that governance should develop as organisational capability and AI adoption mature.

Ultimately, organisational value depends on achieving a balance between control and adaptability. Governance that is too weak may expose the organisation to uncontrolled risk, while governance that is excessively rigid may prevent the organisation from realising legitimate benefits from AI. Responsible AI governance therefore requires a shift from viewing control and innovation as competing objectives towards recognising that appropriate control can make innovation more scalable.

The goal of AI governance should consequently be not to slow AI adoption, but to make the speed of adoption proportionate to risk. By establishing clear boundaries, embedding controls into workflows and differentiating oversight according to consequence, organisations can create conditions in which low-risk experimentation is accelerated while high-consequence applications receive the scrutiny they require. In this sense, effective governance becomes part of the organisation's capacity to create value from AI rather than simply a cost of adopting it.

11. The Agentic Enterprise and the Expansion of Organisational Control

The emergence of agentic AI changes not only the capabilities of enterprise technology but also the scope of organisational governance. Traditional enterprise systems generally execute predefined rules or workflows, with their behaviour largely determined by explicitly programmed logic. Agentic AI introduces a different operational possibility. AI systems can interpret objectives, reason about possible actions, select tools, interact with external systems and adjust their behaviour in response to changing circumstances (Hosseini and Seilani, 2025). The governance question therefore shifts from what an AI system generates to what it is authorised to do.

This represents a significant expansion of organisational control. A conventional AI application may provide a recommendation that a human subsequently evaluates and acts upon. An agentic system can increasingly participate in the action itself by retrieving information, invoking software tools, initiating processes or communicating with other systems. Kshetri (2025) highlights the resulting importance of security, identity and oversight as intelligent systems become increasingly autonomous. The critical governance issue is consequently not autonomy in isolation, but the relationship between autonomy, authority and organisational consequence.

Agentic systems should therefore not be governed simply as another category of software application. They may not possess legal or institutional responsibility, but they can perform activities that have real organisational consequences. An AI agent that creates a purchase order, modifies a customer record, initiates a transaction or changes an operational configuration is participating in an organisational process even if responsibility ultimately remains with human decision-makers. Governance must therefore establish who owns the agent, what it is authorised to do, under which circumstances it may act and how its actions can be reviewed.

This requires explicit control over identity, authority and permissions. An agent should have a distinguishable identity rather than operating as an indistinguishable extension of the human user who initiated it. Its permissions should correspond to the task it has been authorised to perform, rather than automatically inheriting all privileges available to that user. This distinction becomes particularly important where employees have broad access to organisational systems. Giving an AI agent unrestricted access simply because its initiating user possesses that access can transform a legitimate user privilege into an excessive machine authority.

The principle is particularly important because agentic systems can operate at machine speed and across multiple interconnected services. Hosseini and Seilani (2025) describe agentic AI as involving increasingly autonomous interaction with tools and environments, while Kshetri (2025) emphasises the security implications of granting autonomous systems access to organisational resources. A small configuration error in an agent's permissions can therefore have consequences beyond the original interaction. Governance must consequently establish boundaries before an agent is allowed to act rather than relying exclusively on retrospective review.

This also requires a distinction between capability and authority. An agent may technically be capable of performing an action without being authorised to perform it. Governance should therefore determine which capabilities are necessary for a particular task and which actions require additional approval. For example, an agent may be permitted to retrieve information and prepare a transaction but not execute the transaction without human confirmation. Similarly, an agent may be allowed to draft a change to a production system while requiring explicit authorisation before implementing it.

Such controls extend the logic of Zero Trust into the domain of machine identity. Trust should not be treated as a permanent property granted to an AI system because it has passed an initial approval process. Instead, authority should remain conditional on identity, context, task, environment and risk. This is particularly important because an agent's behaviour can change depending on the information it receives, the tools available to it and the instructions or objectives governing its operation. Kshetri (2025) consequently places security, identity and oversight at the centre of governing autonomous intelligent systems.

The principle of least privilege is therefore especially relevant to agentic AI. Agents should receive only the permissions necessary to perform their assigned functions, with access restricted where possible by task and context. High-consequence actions should require stronger controls than routine information retrieval or low-risk administrative activity. This creates a direct relationship between the degree of autonomy granted to an agent and the strength of the controls surrounding that autonomy.

Human oversight remains important, but agentic systems also expose the limitations of treating human involvement as a simple approval step. If an agent can perform a large number of actions autonomously, requiring a human to approve every individual action may be impractical and may encourage superficial approval behaviour. Effective oversight should instead focus human attention on actions where organisational consequence or uncertainty is sufficiently high. Zercher et al. (2025) demonstrate that AI participation can alter team decision-making processes and decision quality, reinforcing the need to consider how human and AI contributions interact rather than treating AI as an isolated technical component.

This means that meaningful human control may require intervention thresholds, escalation mechanisms and the ability to suspend or reverse agent activity. The objective is not to ensure that a human observes every action, but to ensure that the organisation retains the ability to understand, challenge and, where necessary, stop consequential actions. Such control becomes increasingly important as agentic systems move from assisting employees towards executing parts of organisational workflows.

Auditability is another essential component of action governance. Organisations need to know not only which model generated an output but also which agent acted, what authority it possessed, which tools it accessed, what information informed the action and what action was ultimately taken. This creates an evidential record through which organisational decisions and machine actions can be reviewed. Radanliev, Santos and Ani (2025) similarly emphasise the importance of cybersecurity and resilience across the generative AI lifecycle, reinforcing the need for monitoring and control mechanisms that extend beyond the model itself.

The implications extend across the enterprise because agents can connect previously separate systems and processes. An agent may retrieve information from one application, use it to make a decision and then initiate an action in another. Governance must therefore follow the agent across the workflow rather than evaluating each system independently. Sarkar (2026) highlights the increasing importance of integrating generative AI into organisational workflows, while Schulte and Kanbach (2026) identify the expanding role of generative AI across organisational decision-making activities. As AI becomes embedded across these processes, governance must account for the cumulative consequences of interconnected AI-enabled actions.

This also changes the meaning of accountability. The AI agent cannot be treated as the accountable organisational actor simply because it performed the action. Responsibility remains with the organisation and the people who design, deploy, authorise and oversee the system. Governance must therefore make ownership explicit. There should be identifiable individuals or functions responsible for an agent's purpose, permissions, performance, monitoring and continued authorisation. The more consequential the agent's activities, the more important this allocation of responsibility becomes.

The resulting governance challenge can be understood as a progression from model governance to action governance. Model governance asks whether a system is appropriately designed, tested and monitored. Action governance asks an additional and more consequential question: what is the AI system authorised to do within the organisation? This encompasses identity, permissions, task boundaries, escalation, monitoring, auditability, human intervention and the ability to revoke authority.

The expansion of AI autonomy therefore requires an expansion of organisational control. Organisations cannot assume that approving an AI model or application is sufficient when that system can independently interact with enterprise resources. Governance must extend to the authority granted to the system and the conditions under which that authority can be exercised.

The agentic enterprise consequently represents a shift from governing AI as a technology to governing AI as an operational participant in organisational processes. The central principle is not to prevent autonomy, but to ensure that autonomy remains bounded by organisational authority. As agentic systems become more capable, the organisation must be able to demonstrate who authorised the system, what it was permitted to do, what it actually did and how its authority can be constrained or withdrawn. In this sense, the ultimate requirement of agentic AI governance is not simply trustworthy models, but controlled action.

12. Discussion: How Much Governance Is Enough?

The analysis suggests that the question of “how much AI governance is enough?” cannot be answered by a fixed number of policies, committees, approval stages or technical controls. AI governance is not an absolute organisational state. The appropriate level of governance depends on what AI is capable of doing, where it is deployed, what information it can access, how much autonomy it has and what consequences may follow from failure. Governance must therefore be understood as a proportional organisational capability rather than as a universal set of requirements.

This conclusion is consistent with the broader literature on AI governance. Batool, Zowghi and Bano (2025) demonstrate the multidimensional nature of AI governance and the range of organisational concerns involved in governing AI systems. Papagiannidis, Mikalef and Conboy (2025) similarly emphasise that responsible AI governance requires organisational structures and processes rather than reliance on isolated principles. Gartner (2026) approaches governance as a maturity journey in which organisations progressively develop their ability to manage AI as adoption and organisational exposure increase. Taken together, these perspectives suggest that governance adequacy is contextual and should evolve with the organisation's use of AI.

Governance can therefore be considered sufficient when an organisation can demonstrate meaningful control over its AI exposure. This includes knowing where AI is being used, establishing clear ownership, understanding the data and systems involved, constraining AI according to risk, monitoring relevant behaviour, detecting material failures and intervening when necessary. The objective is not to eliminate all uncertainty or prevent all AI-related failure. Rather, the objective is to ensure that failures remain sufficiently visible, bounded and manageable that organisational accountability and resilience can be maintained.

This produces a more useful conception of governance maturity. At an early stage of AI adoption, organisations primarily require visibility and basic boundaries. They need to know which AI tools are being used, what forms of use are acceptable and what information may be shared. As experimentation expands, governance needs to become more consistent and transparent so that employees and business units can understand how AI can be adopted safely. The prevalence of Shadow AI demonstrates the consequences of failing to establish such visibility and usable boundaries (Silic, Silic and Kind-Trüller, 2025).

As AI becomes embedded in organisational workflows, governance must become more formalised. Responsibilities need to be allocated, risk assessments need to become repeatable, monitoring needs to become systematic and controls need to be incorporated into the processes through which AI is actually used. Sarkar (2026) and Hughes et al. (2026) both emphasise the importance of organisational and workflow conditions in moving generative AI from experimentation towards meaningful organisational adoption. Governance must develop alongside this transition rather than remaining an external approval process.

As AI scales further, governance increasingly needs to become integrated into operational and technical infrastructure. Data controls, identity and access management, monitoring, audit logging, approval thresholds and incident-response mechanisms can increasingly be embedded within the systems through which AI operates. This is particularly important for agentic AI, where governance must extend beyond the quality of outputs towards the authority and actions of AI systems (Hosseini and Seilani, 2025; Kshetri, 2025). At this stage, governance is no longer simply about deciding whether an AI system should be permitted to operate. It is about continuously controlling what the system can access and do.

At the highest level of maturity, AI governance becomes part of ordinary organisational management rather than a separate compliance activity. AI-related accountability, risk management, procurement, architecture, cybersecurity, data governance and business continuity become interconnected elements of how the organisation operates. Gartner's (2026) maturity perspective is useful in this respect because it presents governance as an evolving organisational capability rather than a one-time implementation exercise. Mature governance is therefore characterised less by the existence of a comprehensive policy than by the extent to which appropriate controls have become embedded in normal organisational processes.

This progression should not, however, be interpreted as a race towards maximum governance. More governance is not automatically better governance. The objective is fitness for purpose. A low-risk internal productivity application should not necessarily be subjected to the same approval requirements as an AI system that influences financial transactions, employee decisions or other high-consequence activities. Conversely, a critical AI-enabled process should not be governed as though it were simply an experimental productivity tool. The intensity of governance should reflect the level of organisational exposure.

The concept of proportionality is therefore central to answering the question of how much governance is enough. Relevant dimensions include the sensitivity of the data involved, the autonomy of the AI system, the significance of the decisions it influences, the criticality of the workflow, the extent of external dependency and the organisation's ability to detect and recover from failure. Governance becomes more demanding as these factors increase. Radanliev, Santos and Ani (2025) and Neri, Niccolini and Virili (2026) reinforce the importance of resilience and the ability to manage disruption, while Kshetri (2025) highlights the particular control challenges created by increasingly autonomous intelligent systems.

Governance that is too weak creates obvious risks. Uncontrolled AI adoption can expose sensitive information, create security vulnerabilities, introduce unreliable decisions, obscure accountability and allow AI systems to become embedded in critical workflows without adequate oversight. Shadow AI illustrates one dimension of this problem, while the expansion of agentic systems introduces additional risks associated with machine identity, permissions and autonomous action (Silic, Silic and Kind-Trüller, 2025; Kshetri, 2025).

However, governance that is unnecessarily heavy can also create organisational problems. Excessive approval requirements can delay legitimate experimentation, increase administrative costs and encourage employees to circumvent formal processes. Where employees perceive governance as impractical or disconnected from how work is actually performed, AI adoption may move outside approved channels rather than stop altogether. Governance must therefore be sufficiently strong to control material risks while remaining usable enough to support legitimate organisational activity.

This creates a fundamental relationship between AI capability and organisational control. As AI capability increases, the organisation's ability to understand, constrain and intervene must increase accordingly. A system that merely generates low-consequence text requires a different control environment from an agent capable of accessing enterprise systems and executing consequential actions. Similarly, an AI application operating on non-sensitive information presents a different governance challenge from one processing confidential data or influencing high-impact decisions.

The question of governance sufficiency can therefore be translated into a set of practical organisational questions. Where is AI being used? What data does it access? Which systems can it interact with? Which decisions does it influence? What actions can it take? Who owns the use case? What permissions does the AI possess? What external providers does the process depend upon? How are its outputs and actions monitored? What happens when the system fails? Can its actions be reconstructed? Who has the authority to intervene or stop it? Can the organisation continue operating if a critical AI service becomes unavailable?

These questions are important because they shift attention from documented intention to demonstrated control. An organisation may possess a comprehensive AI policy while having little visibility into actual AI use, limited control over employee adoption or no effective mechanism for intervening when an AI system behaves unexpectedly. Conversely, an organisation with fewer formal documents may demonstrate stronger governance if accountability, permissions, monitoring, intervention and recovery are effectively embedded in its operating processes.

This distinction is particularly important in the agentic enterprise. When AI systems can act rather than simply advise, governance must demonstrate control over action. When AI depends on external providers, governance must demonstrate control over critical dependencies. When AI is embedded in workflows, governance must demonstrate control over the processes in which it operates. When AI is used by employees outside formal channels, governance must demonstrate sufficient visibility to understand and manage that exposure. Governance therefore becomes progressively less about controlling a technology in isolation and more about maintaining organisational control across an interconnected AI-enabled environment.

The concept of maturity should consequently be understood as the relationship between AI capability and organisational control. A highly capable AI environment with weak visibility, unclear accountability and limited intervention mechanisms represents immature governance regardless of how extensive its policy documentation may be. Conversely, a complex AI environment in which responsibilities, permissions, monitoring, resilience and intervention are demonstrably embedded may represent a mature governance capability even though not every AI use is subject to identical controls.

The answer to “how much governance is enough?” is therefore neither minimum governance nor maximum governance. It is the amount of governance necessary to maintain accountability, control and resilience at the level of AI capability and organisational consequence being introduced. As AI becomes more autonomous, more deeply integrated and more consequential, the organisation must increase its ability to understand, constrain, monitor and intervene. Where AI remains low-risk and limited in scope, governance can remain proportionate and lightweight.

Ultimately, governance maturity is demonstrated not by the volume of governance activity but by the organisation's ability to remain in control. The strongest evidence of mature AI governance is therefore not that an organisation has written comprehensive rules, but that it can identify where AI operates, understand what it can do, assign responsibility, constrain its authority, reconstruct its actions, respond to failure and maintain operations when critical dependencies are disrupted. In this sense, “enough” governance is the point at which organisational control remains demonstrable as AI capability and organisational reliance increase.

13. Conclusion

The central question of this paper has been how much AI governance is enough as organisations move from experimentation with generative AI towards increasingly autonomous and embedded forms of artificial intelligence. The analysis demonstrates that this question cannot be answered through a universal set of policies, committees, approval stages or technical controls. The appropriate level of governance depends on the relationship between AI capability and organisational exposure. As AI becomes more autonomous, accesses more sensitive information, influences more consequential decisions, becomes more deeply embedded in workflows or creates greater dependency on external providers, the organisation's capacity for control must increase accordingly.

This argument requires a shift in how AI governance itself is understood. Governance cannot remain primarily a statement of principles or an approval process applied at the beginning of the AI lifecycle. The literature increasingly points towards governance as an organisational capability involving accountability, processes, technical controls and continuous oversight (Batool, Zowghi and Bano, 2025; Papagiannidis, Mikalef and Conboy, 2025). Gartner's (2026) maturity perspective similarly suggests that governance develops alongside organisational AI adoption. The important measure of maturity is therefore not the existence of formal documentation but whether the organisation can demonstrate effective control over AI in practice.

Several implications follow from this perspective. First, AI governance must move beyond the model itself. AI systems operate within broader socio-technical environments involving data, applications, identities, APIs, workflows, employees and external providers. Governance must therefore address the environment in which AI operates rather than treating the model as an isolated object of control. This is particularly important as organisations move from experimentation towards workflow-level integration (Sarkar, 2026; Hughes et al., 2026).

Second, governance must account for the reality of Shadow AI. Employees will often adopt AI tools in response to genuine organisational needs, particularly where formal governance processes fail to provide accessible and timely routes for legitimate experimentation. Shadow AI should therefore not be understood solely as a compliance failure. It can also reveal a mismatch between organisational demand and governance capability (Silic, Silic and Kind-Trüller, 2025). Effective governance must combine visibility and boundaries with practical mechanisms that allow useful AI adoption to occur safely.

Third, the increasing role of AI in organisational decision-making makes human judgement more important, not less. Generative AI can alter how information is produced, evaluated and incorporated into decisions, creating both opportunities and new forms of decision-making risk (Zercher et al., 2025; Schulte and Kanbach, 2026). Human oversight is therefore meaningful only when people retain sufficient information, competence, authority and opportunity to challenge or override AI-supported decisions. A nominal human-in-the-loop arrangement is insufficient if the human lacks the practical ability to exercise judgement.

Fourth, the emergence of agentic AI changes the nature of governance itself. Agentic systems can increasingly interpret objectives, select actions, invoke tools and interact with organisational systems (Hosseini and Seilani, 2025). Governance must consequently move from model governance to action governance. The critical questions become what an AI agent is authorised to do, which identity it operates under, which permissions it possesses, what actions require additional approval and how its authority can be monitored or withdrawn. Kshetri (2025) highlights the importance of identity, security and oversight in governing increasingly autonomous intelligent systems. The organisation must remain accountable for actions performed by AI even where those actions are initiated autonomously.

Fifth, AI governance must increasingly become dependency governance. Enterprise AI frequently relies on external foundation-model providers, cloud infrastructure, APIs, data services and other third-party technologies. These relationships create operational and strategic dependencies that may affect security, availability, cost, data governance and organisational resilience. Governance must therefore extend into procurement, architecture, vendor management and business continuity. The relevant question is not simply whether a provider satisfies current compliance requirements, but whether the organisation understands the dependency it is creating and can remain operational if that dependency fails or changes. This reflects the broader importance of resilience identified by Radanliev, Santos and Ani (2025) and Neri, Niccolini and Virili (2026).

These developments reinforce the importance of proportionality. Governance that is too weak can expose organisations to uncontrolled risk, while governance that is unnecessarily restrictive can create friction, delay innovation and encourage employees to circumvent formal processes. The appropriate objective is therefore risk-adjusted speed: low-risk AI applications should be capable of moving rapidly through proportionate governance processes, while higher-consequence applications require stronger evidence, oversight, monitoring and intervention. Governance should be calibrated according to factors such as data sensitivity, AI autonomy, decision consequence, workflow criticality, external dependency and the organisation's ability to detect and recover from failure.

This also changes how governance effectiveness should be evaluated. The existence of an AI policy does not demonstrate that an organisation has effective AI governance. A more meaningful assessment asks whether the organisation knows where AI is being used, understands what data and systems it can access, has assigned clear ownership, can constrain its authority, can reconstruct its actions, can detect material failures, can intervene when necessary and can maintain critical operations when an AI service becomes unavailable. These are indicators of demonstrable organisational control.

The resulting conception of governance maturity is therefore not a progression towards maximum bureaucracy. It is a progression towards greater organisational capability. Early-stage organisations require visibility and basic boundaries. As adoption expands, governance becomes more repeatable and transparent. As AI becomes embedded in critical workflows, accountability, monitoring and technical controls become increasingly formalised and integrated into operational systems. At higher levels of maturity, governance becomes part of ordinary organisational management rather than a separate compliance function (Gartner, 2026).

The paper therefore concludes that there is no universal answer to how much AI governance is enough. Governance is sufficient when it is proportionate to the organisation's AI exposure and when the organisation can demonstrate that it remains in control. The relevant threshold is reached when accountability is clear, AI authority is bounded, material behaviour is observable, intervention is possible and critical failures can be contained and recovered from.

The central implication is that the future of AI governance is not simply about governing artificial intelligence. It is about governing the organisational systems, decisions, dependencies and actions through which artificial intelligence creates consequences. As AI becomes increasingly autonomous and embedded, governance must evolve accordingly. The ultimate measure of maturity is therefore not how many rules an organisation has written, but whether it can continue to exercise meaningful organisational control when AI systems become more capable, more interconnected and more consequential. In this sense, “enough” governance is the level of control required to preserve accountability, resilience and operational continuity as AI capability and organisational reliance increase.

References

Batool, A., Zowghi, D. and Bano, M. (2025) ‘AI governance: a systematic literature review’, AI and Ethics, 5, pp. 3265–3279. doi: 10.1007/s43681-024-00653-w.

Gartner (2026) ‘AI governance: How much is enough?’, Cybersecurity Awareness Month 2026. Gartner.

Hosseini, S. and Seilani, H. (2025) ‘The role of agentic AI in shaping a smart future: A systematic review’, Array, 26, 100399.

Radanliev, P., Santos, O. and Ani, U.D. (2025) ‘Generative AI cybersecurity and resilience’, Frontiers in Artificial Intelligence, 8, 1568360. doi: 10.3389/frai.2025.1568360.

Silic, M., Silic, D. and Kind-Trüller, K. (2025) ‘From Shadow IT to Shadow AI–Threats, Risks and Opportunities for Organizations’, Strategic Change, pp. 1–16.

Zercher, D., Jussupow, E., Benke, I. and Heinzl, A. (2025) ‘How Can Teams Benefit From AI Team Members? Exploring the Effect of Generative AI on Decision-Making Processes and Decision Quality in Team–AI Collaboration’, Journal of Organizational Behavior, 47(8), pp. 1369–1396. doi: 10.1002/job.2898.

Nir Kshetri (2025), ‘Governing Agentic AI: Security, Identity, and Oversight in the Age of Autonomous Intelligent Systems’, Computer, 58(8), pp. 123–129. doi: 10.1109/MC.2025.3572173.

Papagiannidis Emmanouil, Mikalef Patrick, Conboy Kieran, (2025), ‘Responsible artificial intelligence governance: A review and research framework’, Journal of Strategic Information Systems, 34

Sarkar, A. (2026) ‘From AI Hype to Workflow Reality: A Strategic Framework for Integrating Generative AI Across Organizational Functions’, Organizational Dynamics, 55(1), 101202. doi: 10.1016/j.orgdyn.2025.101202.

Hughes, L., Davies, F., Li, K., Madugoda Gunaratnege, S., Malik, T. and Dwivedi, Y.K. (2026) ‘Beyond the hype: Organisational adoption of Generative AI through the lens of the TOE framework–A mixed methods perspective’, International Journal of Information Management, 86, 102982.

Neri, M., Niccolini, F. and Virili, F. (2026) ‘Organizational cyber resilience: toward an integrative conceptual framework’, Management Review Quarterly, 76(1), pp. 789–840.

Niklas Schulte & Dominik K. Kanbach ‘Rethinking organizational decision-making: The emerging roles and tasks of generative artificial intelligence’ (2026), Management Review Quarterly

Contact

Reach out via email for inquiries.

Email

Subscribe to newsletter

info@grcadvisory.ch

© 2025. All rights reserved.