40% Have Scaled AI Agents. Only 6% Generate 5%+ of EBIT. What Is Missing?
AI adoption is moving faster than enterprise value.
McKinsey’s 2026 State of AI research, cited in its new AI operating system report, found that 40% of respondents at large companies have scaled AI agents. Only 6% say AI is generating at least 5% of EBIT. McKinsey: The AI Operating System
That is a striking gap.
Companies are clearly getting better at deploying AI. The harder question is whether they are getting better at turning AI into:
- Decisions
- Action
- Outcomes
- Financial value
McKinsey’s conclusion is important: the challenge is no longer simply getting people to use AI; it is changing how the enterprise itself works. McKinsey: The AI Operating System
That changes the AI strategy problem.
The Problem May Not Be AI Capability
For much of the last few years, enterprise AI strategy has focused on capability:
- Can the model summarize?
- Can it generate code?
- Can agents automate the task?
- Can AI reduce cycle time?
- Can we deploy at scale?
Those questions still matter. But once companies have dozens of use cases and agents in production, they stop being the most important questions.
The management questions become harder:
- Which decisions actually matter?
- What evidence should inform them?
- What should AI be allowed to recommend or execute?
- Where does human judgment remain essential?
- What level of risk is acceptable?
- Which initiatives deserve more capital?
- What should stop when it no longer creates value?
AI can generate intelligence. The enterprise still has to decide what to do with it.
McKinsey Is Describing a Management Layer
McKinsey calls its answer an AI operating system.
It describes the AI OS as a continuously operating management layer that maintains enterprise context, reasons across functions, recommends or executes decisions, and learns from outcomes. McKinsey: The AI Operating System
That is a very different proposition from simply adding more AI tools. It means connecting:
Strategy → Context → Intelligence → Decision → Action → Outcome → Learning
The report explicitly says an AI operating system is not simply an executive dashboard. Instead of reporting what already happened, it should help leaders understand what is happening now, what may happen next, which actions are available, and how those decisions should be coordinated. McKinsey: The AI Operating System
That distinction matters. A dashboard can surface a problem, an AI system can analyze it, and an agent can recommend an action. But none of those things, by themselves, answer: What should the organization actually do?
Start With the Decisions That Matter
One of the strongest recommendations in the report is also one of the simplest: McKinsey argues that companies should begin by identifying the decisions that matter, then determine what data, insights, and AI systems are required to support them. McKinsey: The AI Operating System
That reverses a common enterprise AI sequence. Instead of:
Technology → Use Case → Adoption → Maybe Value
start with:
Business Outcome → Decision → Required Evidence → Intelligence → Action
That is a much more disciplined operating model. The objective is no longer to find more places where AI can be used. It is to identify where better decisions can create measurable enterprise value.
This Is Where the Decision Layer Becomes Critical
McKinsey’s AI operating system is broad. It includes:
- Enterprise context
- Semantic understanding
- Data
- Models
- Agents
- Orchestration
- Workflow redesign
- Management processes
- Learning
But inside that operating model, one problem remains unavoidable: What does the available intelligence actually justify doing?
That is the decision layer. It has to connect evidence, economics, risk, confidence, judgment, and authority to action.
This is where Decision Architecture becomes important. The model may generate the analysis, and the agent may be capable of executing. But leadership still needs explicit answers to questions such as:
- Is the evidence strong enough?
- Does the economics justify action?
- How much risk are we accepting?
- Who has authority?
- Does this require human judgment?
- What happens next?
That is the difference between intelligence and decision.
Intelligence Is Not Authority
McKinsey makes this especially clear in its recommendations to CEOs. The report says leaders should identify:
- Which decisions AI and agents may make or execute
- Which decisions require human judgment
- When issues should escalate to senior leadership McKinsey: The AI Operating System
That distinction is fundamental. An AI system may be capable of recommending a pricing change. An agent may be capable of executing it. But capability does not automatically determine permission.
Intelligence is not authority.
The enterprise needs explicit rules for where AI ends and human accountability begins.
The Economics Belong Inside the Decision
The report also makes a point that may become increasingly important as AI scales. McKinsey notes that token unit costs have fallen by more than 95%, yet AI spending at many companies has still increased by more than 3x. McKinsey: The AI Operating System
That sounds contradictory until you look at what drives total AI economics:
- Model choice
- How tasks are structured
- Reasoning consumption
- Human support
- Quality
- Cost of errors
McKinsey’s recommendation is explicit: AI economics should be incorporated into the operating system’s decision-making process. McKinsey: The AI Operating System
Cost should not sit in a separate report reviewed after deployment; it should help determine the decision itself.
Do not pay for more intelligence than the decision requires.
That means the system should be able to distinguish between high-consequence decisions that justify stronger intelligence and lower-risk workflows that should route to cheaper models or simpler processes. That is not merely technical optimization. It is capital allocation.
The System Should Learn From Outcomes
McKinsey’s AI OS also includes a continuous learning loop. The report describes a shift from periodic reviews and model updates toward continuous learning from decisions, actions, and outcomes so future recommendations improve. McKinsey: The AI Operating System
This matters because a good AI system should not only help make the current decision—it should improve the next one.
The operating loop becomes:
Decision → Action → Outcome → Learning → Better Future Decision
That is how enterprise experience begins to compound.
The Market Is Moving From AI Adoption to AI Decision Management
The 40% versus 6% gap is not evidence that AI does not work. It is evidence that deployment alone is not enough. The enterprise has to connect AI capability to:
- Business outcomes
- Decision rights
- Economic tradeoffs
- Authority
- Workflow redesign
- Feedback
- Learning
That is a management problem. And increasingly, it is becoming the central enterprise AI problem.
The next phase of AI strategy will not be defined by who can deploy the most agents. It will be defined by which organizations can consistently answer:
What does the evidence justify doing next?
Because intelligence can be abundant. Value still depends on the quality of the decision.