Authority Can Be Delegated. Accountability Cannot.

Why the rise of AI agents requires a new discipline of decision stewardship

Artificial intelligence is creating extraordinary opportunities for organizations to strengthen how they operate, make decisions, and serve people. These systems can analyze information at a scale most teams can't match, reduce administrative burdens, identify patterns, and extend people's capacity across an organization. Much of the conversation has understandably focused on what that increased capability can produce. The next phase will require leaders to pay equal attention to what organizations are willing to authorize.

AI systems are beginning to move beyond tools that provide information or recommendations toward agents capable of acting on behalf of people and institutions. That shift is already visible in commerce. Visa has developed infrastructure intended to allow approved AI agents to interact with merchants and make payments on behalf of account holders, while Mastercard has introduced its own infrastructure for agentic transactions. Visa has also reported early evidence of agents paying for services, booking travel, reordering inventory, and purchasing computing resources. These developments matter beyond payments. They demonstrate what changes when artificial intelligence moves from informing a decision to executing one.

An AI system that recommends a transaction is one thing. A system authorized to initiate it is something different. The technological difference may appear incremental, but the institutional difference is significant because action introduces questions of authority, responsibility, and consequence. As AI becomes capable of doing more on behalf of organizations, leaders will have to determine not simply what the technology can do, but what it should be authorized to do.

That requires a new management discipline, which I call Decision Stewardship.

Capability Is Not Authority

Organizations delegate authority every day. Boards delegate responsibilities to management. Executives establish decision rights across teams. Financial institutions set approval limits that allow people to act without seeking permission for every transaction. Well-governed organizations do not treat delegation as the absence of control. They establish boundaries within which authority can be exercised. AI requires the same discipline, but with an important complication. Technological capability can advance faster than the governance structures surrounding it. A system may be able to evaluate vendors, analyze credit risk, monitor financial performance, allocate certain resources, or execute routine transactions. None of those capabilities answers the governance question of whether the system should have the authority to act. That determination still belongs to the institution deploying it.

The distinction becomes more consequential as autonomous systems improve. Organizations may discover that AI can perform increasingly complex work before leadership decides which decisions should remain recommendations, which can be executed within established parameters, and which require human judgment because the consequences extend beyond what the system can evaluate. The opportunity is not to restrict AI simply because its capabilities are expanding. It is to become more disciplined about the relationship between capability and authority. An institution should be able to explain not only what its AI systems can do, but why they have been permitted to do it.

Human Oversight Must Mean More Than Approval

The phrase "human in the loop" has become a familiar response to concerns about AI governance. For consequential decisions, human involvement remains important. But having a person somewhere in a process does not necessarily mean human judgment is governing it. Consider a system that identifies financial transactions for potential fraud. An employee may technically review the recommendation before taking action. Meaningful oversight depends on more than an approval step. The employee needs enough information to understand why the transaction was flagged, sufficient authority to disagree with the system, and a process for determining what happens when the model and human judgment diverge. Someone must also set the risk threshold and remain responsible for determining whether it continues to produce acceptable outcomes.

This is where the difference between human presence and human judgment becomes important. If people routinely approve recommendations they do not understand, cannot meaningfully challenge, or lack the authority to override, the institution may still have humans in the process while practical decision authority has shifted elsewhere. That risk grows as AI systems become faster and more capable. People working under time pressure may begin to defer to systems that appear more comprehensive, consistent, or technically informed than their own analysis. Over time, what began as decision support can become the institution's default judgment without any formal decision to make it so. The governance challenge, then, is not merely to preserve human involvement. It is to preserve meaningful judgment where judgment matters.

Decision Stewardship

Decision Stewardship in this context is the intentional allocation of intelligence, judgment, authority, and accountability across human and AI decision systems. The distinction among those functions matters. AI may provide intelligence without exercising authority. It may be given bounded authority without determining the policy that created that authority. A person may retain formal approval authority while exercising little practical judgment. When these functions are treated as interchangeable, accountability becomes increasingly difficult to locate. Decision Stewardship requires institutions to make those relationships explicit. That begins with defining what a system is authorized to do. Permission to analyze a decision differs from permission to recommend an action, and permission to recommend differs from permission to execute. The greater the potential consequence, the more deliberate an institution should be about where those boundaries sit and how it reviews them.

Institutions must also determine what evidence is sufficient before action occurs. Not every decision deserves the same standard.

A routine operational choice that can be easily reversed carries different consequences than a decision affecting someone's employment, financial security, access to capital, or ability to participate in an essential service. Decision architecture should reflect those differences rather than treating automation as a single category. The same discipline applies to human judgment. Some decisions involve competing values, unusual circumstances, relationships, context, or consequences that cannot be resolved simply by optimizing for a measurable outcome. Leaders need to identify where those conditions are likely to exist before deploying systems, rather than discovering the limits of automation after people experience the consequences.

Most importantly, the institution must know who owns the outcome. Technology can act. It cannot assume institutional accountability for the policy that authorized the action, the conditions under which authority was delegated, the quality of the system's performance, or the consequences that follow. This is what Decision Stewardship is ultimately intended to protect: the connection between the authority an institution delegates and the responsibility it still carries.

Stewardship Before Scale

The conversation surrounding AI often focuses on how autonomous these systems will become. Executives should be equally concerned with how capable their institutions become alongside them. An organization can automate thousands of decisions without improving the quality of its judgment. It can increase speed while making accountability harder to locate. It can generate more information without clarifying who has authority, what standards govern that authority, or how the institution learns when a decision produces an unintended result.

AI did not create these problems. Unclear decision rights, fragmented accountability, weak feedback systems, and poorly defined authority existed long before artificial intelligence. AI can expose those weaknesses more quickly and magnify their consequences as decisions occur at greater speed and scale. That is why stewardship should precede scale. The principle extends beyond technology. Capital alone does not produce economic development. Data alone does not produce intelligence. Strategy alone does not produce execution. Resources create lasting value when people and institutions have the capability and discipline to steward them well. AI belongs in that same category. Its value will depend not only on what the technology can accomplish, but on the institutional conditions surrounding its use.

For executive teams, this changes where the AI conversation can begin. Instead of starting exclusively with possible use cases, leaders can examine the consequential decisions already moving through their organizations. They can look at who currently has authority, what evidence informs those decisions, where judgment materially affects the outcome, what happens when a decision is wrong, and who is responsible for learning from that failure. Once that architecture is clear, organizations are better positioned to determine where AI can responsibly expand their capacity. In some areas, greater autonomy may be entirely appropriate. In others, AI may be most valuable as a source of intelligence and analysis rather than action. The answer does not need to be the same across an institution because the consequences are different.

This is also why AI governance cannot remain solely a technology function. The decisions being affected belong to organizations and the people they serve. Executives and boards must determine how authority moves through their institutions, including when technology exercises that authority. Financial institutions, universities, government, philanthropy, community organizations, and other institutions will have to make similar determinations in environments where the consequences, values, and people affected may differ greatly. The people affected by these decisions also have a legitimate interest in how those boundaries are established. A decision does not become less consequential to a person because software made it faster.

As AI becomes more capable, organizations will almost certainly delegate more meaningful work to it. Some will delegate forms of authority that would have seemed unrealistic only a few years ago. That expansion can strengthen institutions, improve service, and let people focus their judgment where it adds the most value. But greater autonomy changes the distance between institutional leaders and individual decisions. A policy established in one part of an organization can eventually govern thousands or millions of actions performed elsewhere by systems operating at a speed no leadership team could replicate manually. At that scale, accountability cannot depend on someone reviewing every decision after it occurs. It has to be built into how authority was designed in the first place.

That may become one of the defining leadership responsibilities of the agentic AI era. The question will not only be how much autonomy technology can achieve. It will be whether our institutions become equally capable of governing the authority they give it. AI can extend an institution's intelligence, reach, and ability to act. It can change who, or what, performs the work. It does not change who is responsible for deciding where that authority begins, where it ends, and what happens when the consequences reach people.

Authority can be delegeted. Accountability cannont.