The Economics of Machine Authority
Imagine an AI agent that can recommend a $5,000 customer concession. Now imagine that the same agent can approve one.
The intelligence may be identical. The economic consequence is not. The difference is authority.
For much of the AI era, enterprises have focused on capability. Can the model write the proposal, analyze the account, detect the anomaly, recommend the next action, or summarize the relevant information? Those questions still matter, but agentic AI changes the discussion because the system increasingly does not stop at recommendation. It can act.
An agent may change a price, issue a refund, alter a production schedule, provision access, commit funds, modify a customer record, disable an account, or execute a transaction. At that point, AI capability becomes enterprise authority, and that changes the economics of AI.
Every agent therefore has two business cases.
The first is familiar: What value can this agent create if we allow it to act? Perhaps it eliminates approval delays, reduces cycle time from days to minutes, expands the number of decisions we can make economically, or makes scarce expertise available at scale. That is value enabled.
The second business case is less familiar: What enterprise value becomes exposed when we grant the agent that authority? What happens if the agent makes the wrong decision, acts on incorrect information, is manipulated, exceeds its intended authority, or optimizes the objective we gave it in an unanticipated way? That is value at risk.
I use the term value at risk here in a managerial rather than a financial-statistical sense. It is the economic consequence attached to granting machine authority.
The two cannot be separated. If every meaningful AI action requires human approval, the enterprise may preserve control while eliminating much of autonomy's value. But simply removing the human does not create responsible autonomy. It transfers authority.
That transfer should have an economic rationale.
The unit of analysis is changing
Traditional technology evaluation usually begins with the system itself. How accurate is the model? How fast is it? How much does it cost? How does it compare with alternatives?
Those questions remain useful, but they become incomplete once the system can act. The more meaningful unit of analysis becomes the decision or action that the system has been allowed to make.
What is the agent permitted to do? Under what conditions? With what information? Using which tools? Within what financial, operational, customer, security, or reputational boundary? Can the action be reversed? Can the agent be interrupted? What happens when the situation falls outside the conditions for which authority was granted?
An agent that drafts a customer email and an agent that changes a customer’s credit limit may rely on similar underlying technology, but they do not have similar economic significance. The difference is not merely technical capability. It is the consequence of the authority granted.
This suggests a useful distinction among three layers: capability, authority, and accountability.
Capability describes what the system can technically do. Authority describes what the organization permits it to do. Accountability determines who remains responsible for the consequences. The first is primarily a technology question, the second is a management and governance question, and the third is an enterprise question.
Confusing these layers creates predictable mistakes. Organizations may grant authority because a model appears capable, or restrict authority because governance is immature, without explicitly evaluating the economic tradeoff.
Authority is not binary
Organizations often discuss autonomy as though it were a yes-or-no decision: either a human remains in the loop, or the AI is autonomous.
In practice, authority is graduated.
An agent may be permitted to observe but not act, recommend but not execute, execute low-consequence actions automatically, act within a defined threshold, act unless a specified exception occurs, operate across multiple systems, or initiate actions that create commitments for the enterprise.
Each increment of authority can create additional economic value. It can also create additional exposure.
The relevant question, therefore, is not whether an agent is autonomous in the abstract. It is how much authority the agent has, over which decisions, under what conditions, and with what controls.
An agent that can automatically approve a $25 expense may not be appropriate to approve a $25,000 purchase order. An agent that can reset a password may not be appropriate to change an administrator’s privileges. A customer service agent may be trusted to issue a small refund while a larger concession still requires escalation.
Authority should be calibrated to consequence.
This is where the discussion shifts from AI capability to what I think of as Authority Economics.
The technology industry is assembling the control system
Several developments across the enterprise technology market are converging around this problem.
Microsoft is creating identity and governance specifically for AI agents, extending authentication, authorization, lifecycle management, least-privilege access, and human sponsorship to nonhuman identities. Palo Alto Networks is moving control closer to the moment of action through agent identity, privilege, tool access, and runtime enforcement. ServiceNow is approaching the problem through enterprise workflows, combining AI visibility, governance, risk, security, performance, and value measurement. Gartner is increasingly emphasizing that agents operating with different levels of autonomy and access cannot all be governed in the same way.
These may appear to be separate technology categories, but I think they are becoming components of a new enterprise management system for machine authority.
The architecture is beginning to emerge. What remains less developed is the economic logic that determines how much authority to grant in the first place.
That is the missing layer.
Governance has two forms of economic value
Security and governance have traditionally been justified largely through avoided loss. A control reduces the probability or consequence of something bad happening.
That remains important, but agentic AI introduces another source of economic value.
Suppose management is unwilling to allow an AI agent to issue customer refunds because the organization cannot adequately identify the agent, constrain its privilege, monitor its behavior, establish clear decision boundaries, or reverse inappropriate actions. Human approval is therefore required.
Now suppose the enterprise introduces controls that make management comfortable delegating refunds up to $500 under defined conditions. The organization can identify the agent, restrict its authority, monitor its actions, preserve the audit trail, route exceptions to a human, and reverse an inappropriate transaction.
Those controls have done more than reduce risk.
They have enabled economically valuable authority that management previously rationally withheld.
That means governance can create value in two directions. It can protect value by reducing exposure, and it can unlock value by allowing greater delegation.
This distinction matters because it changes the business case for AI governance. Governance is not simply a cost imposed on AI after the value case has already been established. In many agentic workflows, governance may be part of the mechanism through which the value becomes possible.
The stronger the control environment, the more authority management may be willing to delegate. The more authority that can be delegated safely, the larger the portion of AI capability that can become economically useful.
That is a very different way to think about security.
The economics of delegation
The value of delegation depends on more than how many tasks an agent can complete. It depends on the decisions that can be made faster, more consistently, or at greater scale; the cost of human review; the value of time saved; the value of decisions that might otherwise not be made; and the consequences of error.
A useful conceptual expression is:
Value Enabled = Decision Value × Delegated Volume × Autonomy Benefit
The autonomy benefit may come from speed, scale, availability, consistency, or lower operating cost. But gross value enabled is not the same as realized enterprise value.
Realized value must account for the controls required to make the delegation acceptable, the cost of monitoring and intervention, the residual exposure after those controls are applied, and the cost of exceptions and remediation.
Conceptually:
Realized Value = Value Enabled − Control Cost − Residual Expected Loss − Exception and Remediation Cost
This is not intended as a universal financial formula. It is a management discipline. It forces the organization to evaluate both sides of the authority decision.
If the control cost exceeds the value of delegation, the authority may not be worth granting. If the potential value is substantial but the residual exposure remains unacceptable, management may need stronger controls, narrower authority boundaries, or a different workflow. If controls reduce the exposure to an acceptable level, the enterprise can rationally grant more authority.
The business case is therefore not simply whether AI can perform the task. It is whether the enterprise can create a sufficiently trustworthy boundary around the authority required to perform it.
Value at risk is a function of authority
The value at risk associated with an AI agent is not determined solely by the model’s error rate.
It also depends on what the agent can affect.
One useful conceptual relationship is:
Value at Risk = Consequence × Exposure × Propagation × Irreversibility
Consequence is the impact of a wrong or unauthorized action. Exposure reflects the number and type of decisions the agent can make. Propagation captures whether one action can trigger additional actions across other systems. Irreversibility reflects how difficult it is to detect, stop, or undo the result.
This explains why even a highly accurate agent may require strict controls. A small error rate applied to a high-consequence, high-volume, highly connected, or irreversible workflow can create substantial enterprise exposure.
It also explains why a less capable agent may be perfectly acceptable in a narrow, reversible, low-consequence workflow.
Accuracy matters, but authority determines its economic significance.
The control stack for governable autonomy
Granting authority responsibly therefore requires more than a model evaluation. It requires a control stack that connects the agent to the enterprise.
The organization must be able to establish identity, assign accountable sponsorship, define authorization, enforce least privilege, determine the context in which actions are permitted, prevent inappropriate actions at runtime, observe what the agent is doing across tools and systems, preserve an audit trail, intervene when necessary, reverse actions where possible, route exceptions appropriately, and measure whether the agent is actually producing the expected economic value.
These controls are not all-or-nothing, nor should they be identical for every agent. An agent that drafts text does not require the same governance structure as one that can commit funds, alter production systems, or change administrative privileges.
As authority increases, the burden of proof should increase with it.
This is the basis of what I call governable autonomy: enough authority to capture the economic value of AI, bounded by controls appropriate to the value at risk.
Governable autonomy is not simply a security posture. It is an enterprise capability.
The human-in-the-loop question changes
This also changes how we should think about human oversight.
Organizations frequently talk about keeping a human in the loop as though human involvement is inherently safer. Sometimes it is. Human approval may be an essential control because judgment, accountability, contextual knowledge, or ethical responsibility genuinely requires human participation.
But sometimes the human is there for a different reason: management does not yet trust the technology or the surrounding control environment enough to delegate the decision.
Those are not the same thing.
One is a control to preserve. The other may be a compensating control for an immature system.
As identity, authorization, observation, runtime enforcement, and reversibility improve, some human approvals can move from mandatory review to exception-based intervention. That is where governance begins to enable value rather than merely constrain risk.
The objective is not to remove humans indiscriminately. It is to put human judgment where it creates the most value: at consequential boundaries, unusual exceptions, ambiguous cases, and decisions that require accountability beyond the agent’s design.
A management framework for granting authority
Before granting an agent additional authority, executives should be able to answer five questions.
First, what decision or action is actually being delegated? Describe authority in operational terms, not vague statements such as “manage the account” or “optimize the workflow.”
Second, what value does delegation create? The organization should estimate the value of faster decisions, greater scale, reduced review cost, improved availability, increased consistency, or decisions that otherwise would not be made economically.
Third, what value is exposed? Management should identify the financial, operational, customer, regulatory, security, and reputational consequences of error, misuse, or unintended behavior.
Fourth, what controls make the exposure acceptable? The required identity, authorization, boundaries, monitoring, intervention, reversibility, and accountability should be explicit.
Finally, how will the enterprise know whether the delegation is working? Measurement must extend beyond activity and productivity to decision quality, exceptions, incidents, reversals, control performance, and realized economic value.
This turns autonomy from a technology setting into a management decision.
The next enterprise AI question
For the past several years, executives have largely asked, What can AI do?
Increasingly, that will not be the difficult question.
The more important questions will be what we should allow AI to do, what additional value becomes possible when we grant that authority, what enterprise value that authority exposes, what controls would allow us to grant more safely, and who remains accountable when the machine acts.
That is why I don't believe value and governance are competing AI agendas. They are becoming the same management problem.
As AI becomes better at accomplishing goals, enterprise advantage will increasingly depend on an organization’s ability to delegate consequential authority without losing control of the consequences.
The objective is neither maximum autonomy nor maximum control. It is governable autonomy.
Governable autonomy is not a permanent state. It is a capability that develops as the enterprise matches authority to controls, measures outcomes, learns where exceptions occur, and increases delegation as it earns confidence.
The organizations that benefit most from agentic AI will not necessarily be those that deploy the most agents. They will be those that make the best decisions about where authority belongs, how much to grant, what controls are required, and when the boundary can safely expand.
The central economic question is therefore no longer simply whether an AI agent is capable.
It is whether the organization can make its authority trustworthy enough to use.
And perhaps the most important question for the next phase of enterprise AI is surprisingly simple:
What is the next increment of machine authority worth?
Comments