AI Agents Need Less Supervision. So Why Do They Need More Management?
AI agents are supposed to reduce management overhead. They can work around the clock: research, analyze, draft, code, check their own work, call tools, retry failures, and coordinate across workflows without waiting for someone to assign the next task. So it seems reasonable to expect a future with less management.
But a recent McKinsey article suggests the opposite may be happening. As agentic AI expands, companies are discovering they still need to decide what level of intelligence a task deserves, what that intelligence is allowed to cost, what quality is sufficient, when the system should retry, when it should stop, which model should handle the work, what the agent may execute autonomously, what must escalate to a human, and who ultimately owns the outcome.
That does not sound like less management. It sounds like different management.
The Scarce Resource Is Changing
For much of modern business history, human labor has been one of the central production constraints. People have limited hours, hiring takes time, expertise is scarce, and capacity has to be carefully allocated. Management systems naturally focused on familiar input questions: Who does the work? How many hours will it take? How much output can the team produce?
Those questions do not disappear, but they become less central. If AI can produce analysis, software, content, research, and operational execution continuously and faster than people, the bottleneck moves upstream.
The Traditional Question: “How do we allocate scarce human capacity?”
The Emerging Question: “How do we allocate abundant machine intelligence toward the outcomes that actually matter?”
Abundant Intelligence Creates an Allocation Problem
McKinsey argues that one of the most expensive mistakes companies make is assuming every problem deserves frontier-model intelligence. Many tasks can be handled effectively by smaller models, deterministic systems, traditional software, or cheaper architectures. Their recommendation is to treat intelligence more like capital: allocate it deliberately rather than assuming more capability is always better.
When intelligence becomes abundant, allocation becomes more important—not less.
Limitless possibility creates its own constraint: choice. When an organization can generate hundreds of strategies and automate thousands of workflows, the question is no longer “Can we do this?” but “Should we do this?”
Human work moves upstream into strategic orchestration:
- Defining core business objectives
- Evaluating trade-offs and alternative architectures
- Setting deterministic constraints and risk tolerances
- Establishing spending boundaries and stopping rules
- Deciding which workflows deserve capital allocation
Activity Is Easier to Measure Than Value
When labor was the primary bottleneck, management tracked inputs and activity: headcount, hours, utilization, and throughput. AI introduces an equivalent temptation with a fresh set of operational vanity metrics:
- Tokens
- Model Calls
- Active Agents
- Retry Counts
- Task Executions
McKinsey argues that the unit of governance for machine work should be the completed business outcome, not the token, model call, or technology line item. A successful agent run is not necessarily a successful business outcome, and higher utilization does not guarantee sound economics.
What Does Outcome Management Require?
Once the management unit becomes the outcome, the system must monitor far more than simple task completion. It must connect multiple operating dimensions:
- Outcome: Was anything economically better?
- Cost: What did achieving that outcome consume?
- Quality: Was the result good enough for production?
- Risk & Confidence: What could fail, and how trustworthy is the underlying evidence?
- Authority: What actions is the system permitted to execute?
- Decision: Scale | Continue | Redesign | Route Cheaper | Add Governance | Escalate | Pause | Stop
This is no longer simple software monitoring; it is full-fledged management infrastructure.
Governance Becomes Economic Performance
Governance is frequently stereotyped as friction, compliance, or a post-hoc constraint. But agentic AI reframes governance into a core driver of economic performance.
McKinsey notes that organizations able to measure, allocate, improve, and integrate machine work effectively can create more value than competitors using the identical underlying foundation models. A firm that knows when to route to a cheaper model, when another retry is worth the expense, when autonomy is safe, and when to terminate an inefficient loop will systematically outperform a competitor deploying raw intelligence without controls.
Governance is not merely about preventing bad behavior. It is about getting better economic performance from the same intelligence.
The Management Layer AI Creates
The central leadership question is shifting from “How capable are our agents?” to “How well do we manage the decisions surrounding them?”
As AI automates routine execution, management pivots away from supervising every unit of labor toward governing authority, bounding costs, validating outcomes, and applying strategic judgment.
The more capable AI becomes at doing the work, the more critical the ultimate executive decision becomes:
What work is actually worth doing?