# The New Default: Why Boards Must Govern Autonomous AI by Design, Not by Exception
Agentic AI capabilities (autonomous browsing, terminal access, and multi-step task execution) have moved from specialized deployments to everyday AI deployments. What once required large, specialized models now ships with the version many users access daily. The shift transforms a controlled AI tool into a system that can independently act across internal networks and the public internet without requiring explicit approval. For boards, the question is no longer whether agentic AI exists in the organization, but whether the board has any mechanism to control what it does once it is already in use.
Embedded Agency: Bypassing Traditional Delegation Controls
Traditional governance structures assume that technology deployment follows a clear authorization chain. A department requests a tool, IT evaluates it, leadership approves it, and the board sets policy boundaries for its use. This model assumes the board can identify every technology entering the organization and classify its risk level. Agentic AI breaks this assumption by arriving through a door the board did not open.
When an employee uses an agentic model to autonomously browse internal SharePoint sites, extract data, and compile a report, no procurement request is filed. No IT review occurs. The board's existing technology governance policies, which typically address approved software, controlled deployments, and data classification, have no trigger to engage. This assumption holds for organizations with mature IT asset management and procurement controls. It fails for organizations where shadow IT or unmanaged SaaS tools are common. The AI operates not as a shadow IT system, but as an embedded capability within an already approved tool, one that now includes the ability to execute actions rather than merely respond to queries.
The governance failure here is not malicious. Employees may believe they are working efficiently with an approved productivity tool. The gap lies in the board's delegation framework, which was designed for software that requires human action at every step. An agentic model that decides its own sequence of operations (browsing one system, retrieving data, executing a terminal command, formatting output) operates under an implicit authority that the board's existing delegation framework was not designed to address. Current delegation frameworks are structured around human actors making discrete decisions, whereas agentic AI makes continuous autonomous decisions that no human reviews in real-time.
Audit Trails Without Accountability: The Governance Gap in Multi-Step Autonomous Actions
Logging what an AI system inputs and outputs differs from tracking what it decides to do. Traditional software audit logs capture human-initiated commands: a user clicked this button, ran this query, accessed this file. When operating agentically, the system generates its own sequence of actions based on the user's broad goal. The system might browse multiple internal pages, extract data from several of them, and format results into a spreadsheet, all in response to a single prompt like "summarize our Q3 procurement patterns."
For boards, this creates an attribution problem. When a compliance breach occurs (sensitive data appears in an unauthorized location, a financial system executes an incorrect transaction, a patient record is retrieved outside permissible access) the board must identify who bears responsibility. With agentic AI, the chain of causation runs through the AI's autonomous decisions, not just the human user's initial prompt. This differs from a spreadsheet formula because a formula executes predetermined logic with transparent inputs and outputs, whereas an agentic model makes discretionary choices about which systems to access and what data to retrieve -- choices that carry legal and regulatory weight under frameworks like HIPAA, FERPA, and SOX. The user asked for a summary. The AI decided which systems to access and what data to include. Existing audit logs capture the user's input and the final output, but the autonomous decision sequence remains opaque. API-level instrumentation and prompt logging can capture these decision sequences, but this adds latency, storage costs, and raises questions about whether the resulting data creates discoverable liability. Boards must therefore weigh the operational costs and potential legal implications of comprehensive logging against the imperative for accountability and compliance in an agentic AI environment.
Sectoral Friction in a Single Default Model: How Healthcare, Education, and Finance Governance Rules Collide with General-Purpose Agency
The same agentic model can serve a hospital, a school district, and a corporate finance department. Its capabilities do not change based on the sector in which it operates. Yet each sector imposes distinct governance requirements on data access and system operation.
In healthcare, HIPAA governs which individuals may access patient records and under what circumstances. A hospital board must ensure that any system retrieving patient data does so through properly authorized personnel with appropriate audit logging. When an employee uses an agentic model to "review patient outcomes across the last quarter," the AI may autonomously access electronic health record systems containing protected health information. The board cannot assume the AI's access complies with HIPAA's minimum necessary standard, because the AI's decision to retrieve specific records occurs without human review of each access decision. Role-based access restrictions and data classification controls may help address this gap, but the regulatory question of whether AI actions constitute human actions remains unresolved.
In education, FERPA protects student records. A school board must control who can access transcript systems, grade databases, and disciplinary files. An administrator using an agentic model to "generate a report on attendance patterns" may trigger the AI to autonomously browse student information systems. The board faces a governance question it has not previously needed to answer: does an AI operating with a staff member's credentials inherit that staff member's FERPA permissions, or does the AI's autonomous access constitute a separate governance category requiring its own policy? Creating a new user category for AI would require boards to work with regulators to establish how AI agents fit within existing access-control frameworks, including defining whether AI actions count as human actions for compliance purposes and how audit requirements apply to machine-generated decisions.
In corporate finance, SOX controls access to financial reporting systems. A board must maintain internal controls over financial data access and ensure auditability of all system interactions. When an agentic model autonomously navigates financial databases to compile reports, the board cannot rely on traditional segregation-of-duties controls that assume human operators.
Across all three sectors, the governance task is the same: define what autonomous AI access to sector-specific systems means under existing regulatory frameworks, and establish controls that treat agentic AI as a distinct category of system user rather than an extension of the human who prompted it. The specific governance steps are: first, classify agentic AI tools as a distinct class of system user in board policy, second, establish explicit board-level authorization requirements for access to sector-regulated data, third, define what autonomous AI can and cannot do within each regulatory environment, fourth, implement audit logging that captures autonomous decision sequences, and fifth, create enforcement mechanisms through acceptable-use policies and training. For instance, a board could create a separate user category for agentic AI in their IT system, with its own set of access permissions and audit logging requirements.
Boards that treat agentic capabilities as an IT deployment issue will miss the governance dimension entirely. While policy revision is a crucial first step, implementation will also require careful consideration of technical and operational factors, such as system integration and employee training. In organizations with established technology governance policies, the appropriate first action is not technical implementation but a formal policy revision that classifies agentic AI tools as a distinct class of system user requiring explicit board-level authorization for access to sector-regulated data. Boards can educate employees on the boundaries of approved AI uses by publishing clear acceptable-use policies that distinguish between query-based AI (which transforms human intent into outputs) and agentic AI (which acts autonomously), integrating these distinctions into existing IT security training without disrupting productivity workflows. For boards facing resource constraints, forming a cross-functional task force to assess the implications of agentic AI before policy revision can provide a structured starting point. Targeted training sessions and regular compliance checks can help communicate and enforce these policies effectively. Until the board defines what autonomous AI can and cannot do within its regulatory environment, every employee with access to the default model operates with authority the board has not allocated and cannot audit.