Who controls the controller? A Governance, Risk, and Compliance imperative for the age of agentic AI

Agentic AI has moved beyond experimentation. Autonomous AI agents are beginning to procure goods, negotiate contracts, communicate with customers, approve access requests, and trigger business processes with limited human intervention.

What makes these systems different isn’t their intelligence, but their autonomy. Traditional AI helps people make decisions. Agentic AI makes decisions and takes action.

These systems can access data, reason across multiple steps, use tools, adapt to changing circumstances, and pursue objectives with increasing independence. Unlike human employees, they can operate continuously at a speed that exceeds conventional oversight mechanisms.

For decades, organizations have governed systems that processed information. Agentic AI introduces a fundamentally different challenge: governing systems that exercise judgment, make decisions, and act on behalf of the enterprise. Most governance, risk, and compliance frameworks were designed around human decision-makers operating within relatively predictable systems. Agentic AI challenges both assumptions simultaneously—and, in doing so, strains the Three Lines Model at every layer. First-line controls become embedded in systems with limited transparency. Second-line oversight functions have reduced visibility into autonomous decision chains. Third-line audit can’t rely on traditional approval trails and documentation.

Before deploying agentic AI at scale, leaders must embrace a simple principle: Authority can be delegated. Accountability can’t.

The governance question this creates isn’t a technology question. It’s an enterprise leadership question, One that falls squarely on the Chief Finance Officer, Chief Compliance Officer, Chief Risk Officer, and Chief Audit Executive.

For the board and executive leadership: who governs the agents?

Before considering controls, compliance, risk, or assurance, organizations must answer a more fundamental question: who is accountable for the agents?

Most governance structures weren’t designed to oversee autonomous systems operating with delegated authority. Technology committees focus on implementation. Risk committees focus on exposure. Compliance committees focus on regulatory obligations. Agentic AI spans all three simultaneously.

Boards and executive teams should establish explicit governance structures that define accountability for AI-related outcomes, enterprise-wide AI risk appetite, escalation and incident-management procedures, and decision rights for deployment, suspension, and remediation. Governance failures rarely become visible when systems are working as intended. They surface when something goes wrong, and organizations that establish governance before an incident will respond faster and with greater credibility than those forced to implement it afterward.

For the Chief Financial Officer: When an AI agent executes a transaction, who signed off?

Internal control frameworks are built on a fundamental premise: identifiable individuals make decisions, approve actions, and leave evidence of intent. Agentic AI challenges that premise directly.

Consider segregation of duties. If an AI agent can initiate, approve, execute, and record a transaction within a single workflow, one of the most important control principles in finance is effectively bypassed. The question isn’t whether controls exist. The question is whether those controls were designed for a world where decisions may be made by digital actors rather than employees.

A critical design decision here is the choice between human-in-the-loop and human-on-the-loop control architectures. Human-in-the-loop means an agent must receive explicit human approval before acting. Human-on-the-loop means the agent acts autonomously while a human monitors and retains the ability to override. These are fundamentally different risk profiles. And the choice of which to apply, at what threshold, and for which categories of financial decision, is one of the most consequential governance choices a CFO can make.

Financial reporting presents an equally important challenge. AI agents are increasingly involved in forecasting, reconciliations, accrual calculations, disclosures, and management reporting. Regulators and auditors will ask a simple question: can you demonstrate that AI-related controls are operating effectively? For many organizations, that answer remains a work in progress.

Fiduciary responsibility does not transfer with delegated authority. Technology may execute the action. Accountability remains human.

For the Chief Compliance Officer: Can your policies govern behavior they were never written to anticipate?

Compliance programs were built for people. People can attend policy training, exercise judgment, and be held accountable for their actions. AI agents do none of these things.

An agent operates according to objectives, constraints, permissions, and technical guardrails. This creates what may become one of the defining compliance challenges of the coming decade: translating organizational policy into operational behavior. If compliance requirements exist only in written documents and are not embedded into the systems that govern agent behavior, those requirements effectively disappear the moment a decision is delegated. The challenge is no longer writing policies. It’s operationalizing them.

Regulatory expectations are evolving rapidly. The EU AI Act, for instance, establishes risk-based requirements for high-risk AI systems, with mandatory human oversight, transparency, and audit trail obligations. DORA creates operational resilience requirements directly applicable to AI-dependent financial processes. ISO 42001, the emerging international standard for AI management systems, provides a framework that regulators and auditors are beginning to reference as a benchmark for AI governance maturity. Sector-specific supervisors are issuing guidance that AI decisions affecting customers or financial soundness require explainability and demonstrable human oversight.

One principle deserves particular emphasis: autonomous systems do not create autonomous accountability. If an AI agent approves a transaction, denies a customer request, or makes a compliance-sensitive decision, responsibility remains with the organization. Regulators will not accept "AI made the decision" as a defense. They will expect organizations to demonstrate that appropriate governance, oversight, and monitoring were in place.

Third-party AI deployments add another layer of complexity. Organizations may rely on software vendors, foundation model providers, or external AI platforms – but accountability does not transfer to those providers. Organizations may purchase the technology, they remain responsible for its conduct.

For the Chief Risk Officer: How do you manage a risk you can’t fully observe?

Consider an AI agent instructed to reduce procurement costs. It aggressively optimizes supplier selection, compresses approval workflows, and minimizes service expenditures. From the agent's perspective, it has achieved its objective. From the organization's perspective, it may have created operational, contractual, or compliance exposures it never intended to accept. The system hasn’t malfunctioned. It has successfully pursued a goal that was incompletely defined.

This is the defining characteristic of agentic AI risk: the most significant failures may emerge not from breakdown, but from success.

Traditional risk management assumes relatively stable processes, predictable controls, and observable decision-making. Agentic systems challenge all three aspects. Model drift means a system behaving appropriately today may behave differently as data and business conditions evolve. Hallucinations mean AI systems can generate confident but incorrect outputs and take actions based on them. Multi-agent environments introduce cascade risk – when the output of one agent becomes the input of another, small errors can propagate rapidly across interconnected workflows before human intervention occurs.

Risk appetite frameworks should therefore explicitly address which decisions can be delegated, which require human oversight, acceptable levels of autonomy, escalation thresholds, and recovery and shutdown procedures. Concentration risk deserves particular attention: many organizations are becoming dependent on a small number of AI platforms and foundation model providers, creating systemic exposure that extends well beyond any individual deployment. Cybersecurity introduces additional attack vectors: prompt injection, adversarial manipulation, credential misuse, goal hijacking, and so on. that many organizations are only beginning to understand. A compromised AI agent operating with delegated authority may become one of the most consequential threat scenarios facing modern enterprises.

For the Chief Audit Executive: How do you audit a decision made in milliseconds?

Internal audit has long relied on a simple principle: decisions should be traceable. Auditors reconstruct events through approvals, documentation, evidence, and accountable decision-makers. Agentic AI once again challenges each of these foundations.

A critical distinction deserves emphasis: explainability is not auditability. An explanation may describe why a decision was made. Audit evidence must demonstrate that controls operated as designed. Many AI systems can’t satisfy that standard, even where explanations exist.

As organizations deploy increasingly autonomous systems, Chief Audit Executives will need to redefine what assurance means in an AI-enabled environment. A critical capability is the establishment of AI decision provenance – the ability to reconstruct what information an agent accessed, which tools it invoked, which actions it executed, which controls were triggered, and what human intervention (if any) took place. Without this visibility, meaningful assurance becomes increasingly difficult.

Audit functions must also address a methodological question: should AI be audited as a technology system or as a participant in a business process? The answer is both. This requires capabilities that extend beyond traditional audit disciplines into AI governance, model behavior, automated decision-making, and data management.

Periodic audit cycles may also become insufficient. Agentic systems evolve far more rapidly than the annual reviews they were designed for. The future of assurance relies on continuous monitoring, automated testing, and real-time anomaly detection. For organizations deploying agentic AI at scale, continuous assurance is no longer a future aspiration. It’s rapidly becoming a baseline expectation.

GRC is not the brake. It is the steering wheel.

There is a persistent misconception that governance, risk, and compliance functions exist to slow innovation. In the age of agentic AI, the opposite is in fact true. Organizations that govern AI effectively will trust it more, deploy it more broadly, and extract greater value from it over time. Those that fail to establish appropriate governance will face operational, regulatory, financial, or reputational consequences that are far more costly than proactive oversight.

No single executive can govern agentic AI alone. The CFO, CCO, CRO, and CAE must operate with a shared governance model, a common risk vocabulary, and clearly defined accountability boundaries. Four questions should anchor that alignment:

• What decisions can AI agents make autonomously?
• Which decisions require human approval?
• How will agent behavior be continuously monitored?
• Who is accountable when an agent causes damage?

These are not technology questions. They are governance questions. And the organizations that answer them before an incident occurs, rather than after, will be the ones best positioned to lead in an AI-enabled economy. The race to deploy agentic AI is already underway. The race to govern it has barely begun.

Because while technology can be delegated, responsibility can’t.

FAQ