The Question That Changed the Meeting
Picture a routine technology update. A manager mentions that a procurement workflow has been "optimized by an AI agent." A director asks a simple question: "Who approved the vendor it selected?" The room goes quiet. The agent has evaluated several suppliers, scored them against historical performance data, and initiated a purchase order—all before any human reviewed the decision. The amount is below the approval threshold. The logic is sound. And yet something fundamental has shifted without the board's awareness.
Scenes like this crystallize what I have been sensing across conversations with peers and advisors: we are governing AI as if it were a tool that waits for instructions. Increasingly, it is not.
What Makes Agentic AI Different
Traditional AI systems respond to prompts. You ask a question; they provide an answer. You review the output; you decide what to do with it. The human remains in the loop at every consequential step.
Agentic AI operates differently. These systems can decompose complex goals into subtasks, call external tools and APIs, execute multi-step workflows, and adapt their approach based on intermediate results—all without returning to a human for approval. They do not merely recommend; they act.
The Cloud Security Alliance's "AI Agent Governance Gap" report from April 2026 projects that 40% of enterprise applications will incorporate agentic AI capabilities by the end of this year. What surprised me about this figure was not its size but its speed. We moved from chatbots to autonomous agents in roughly eighteen months.
From the operator's side of the table, this creates a governance problem that existing frameworks were never designed to address. When I built an EPC business to 450 people and 200+ industrial projects, every significant decision had a clear chain of accountability. Someone signed off. Someone was responsible. With agentic AI, that chain becomes diffuse, sometimes invisible.
The Governance Gap We Are Living Through
Morgan Stanley's "AI Governance Adoption Report 2026" found that only about 39% of Fortune 100 boards have established explicit AI oversight mechanisms. This statistic understates the problem. Having an AI policy is not the same as having governance structures that account for autonomous decision-making.
Most AI governance frameworks I have reviewed treat AI as a category of technology risk—something to be monitored, audited, and periodically assessed. They assume human review before consequential actions. Agentic systems break this assumption. They compress the time between analysis and action to near-zero. They operate at scales and speeds that make human-in-the-loop review impractical for many use cases.
What I have seen in practice is instructive. Organizations have policies covering data privacy, algorithmic bias, and model validation. Few of them address the question of delegated authority—what decisions an AI agent is permitted to make independently, under what constraints, and with what escalation triggers.
KPMG Canada's "AI Governance Principles for Boards Report" from May 2026 frames this well: the shift required is from passive policy review to active governance of delegated authority. The board's role is not merely to approve an AI strategy but to define the boundaries within which autonomous systems may operate.
Shadow Agents and Invisible Risk
The challenge compounds when you consider shadow AI. Unseensecurity.ai's "Shadow AI Report 2026" found that shadow AI was involved in 43% of security incidents last year. These are AI tools and agents deployed by employees or departments without formal IT or governance approval.
In an industrial context, this concerns me deeply. As a chair who built a board from scratch, I have spent considerable effort establishing clear lines of authority and risk escalation. Shadow agents undermine this architecture. An employee using an autonomous AI assistant to draft contracts, analyze supplier data, or communicate with clients creates exposure that the board cannot see, let alone govern.
What I still do not fully understand is how to detect these deployments without creating a surveillance culture that undermines trust. The tension between visibility and autonomy is real, and I have not seen a satisfying resolution.
Questions I Found Myself Asking
Scenes like that one prompted me to compile questions that seemed essential but were absent from most governance discussions:
- What decisions have we implicitly delegated to AI agents? Not what we intended to delegate—what has actually been delegated through deployment choices made below the board level?
- What are the escalation triggers? At what thresholds—financial, reputational, safety-related—does an agent pause and seek human approval?
- How do we audit agent reasoning? Traditional audits examine outcomes. With agents, the reasoning path matters. An agent might reach a correct conclusion through flawed logic, creating latent risk.
- Who is accountable when an agent errs? The developer? The deploying manager? The vendor? The board for insufficient oversight? Legal frameworks have not caught up, which means governance frameworks must fill the gap.
- What is our shadow AI exposure? Do we have visibility into autonomous tools operating outside sanctioned systems?
These questions do not have clean answers. But I found that simply asking them surfaced assumptions and gaps that had been invisible.
AI Literacy as Fiduciary Duty
WilmerHale's "Board Oversight and Artificial Intelligence: Key Governance Priorities for 2026" from January 2026 argues that AI literacy is now a fiduciary requirement. I am inclined to agree, though I would frame it more precisely: what boards need is not technical fluency but conceptual clarity about how autonomous systems differ from traditional automation.
The IMD Mastering Board Governance program I completed earlier this year dedicated significant attention to emerging technology oversight. What struck me was how few governance frameworks distinguish between AI that advises and AI that acts. This distinction is fundamental. Advisory AI augments human judgment. Agentic AI substitutes for it in defined domains.
Governing the latter requires different questions, different controls, and different accountability structures. It also requires boards to engage more directly with deployment decisions that might previously have been considered operational.
What We Do Not Yet Know
I want to be honest about the limits of current understanding—my own included.
We do not yet have mature frameworks for auditing agent behavior at scale. The tools exist in nascent form, but industry standards are still emerging.
We do not have clear legal precedent for liability when autonomous agents cause harm. Regulatory guidance is fragmented across jurisdictions, and the pace of deployment far exceeds the pace of rulemaking.
We do not fully understand the emergent behaviors that arise when multiple agents interact—within an organization or across organizational boundaries. The complexity grows non-linearly.
And we do not have consensus on what "meaningful human oversight" looks like when agents operate at machine speed. The phrase appears in nearly every AI governance document I have read, but its practical implementation remains elusive.
A Closing Observation
The procurement agent in that scene was, by any reasonable measure, performing well. It saved time, reduced errors, and made defensible choices. The governance gap it revealed was not about malfunction but about authority—about decisions being made in the organization's name without the organization's explicit consent.
This is the challenge boards face with agentic AI. The risk is not primarily that agents will fail catastrophically. It is that they will succeed quietly, accumulating delegated authority through convenience and efficiency, until the board discovers it has been governing a fundamentally different organization than it believed.
From the operator's side of the table, I find this both fascinating and sobering. We are building governance structures for systems that are themselves evolving. The frameworks we establish today will need revision tomorrow. What matters now is recognizing that the question has changed—from "What did the AI recommend?" to "What did the AI do?"