Dr. Chih-Han Yu is the CEO and co-founder of Appier, an AI-native agentic-AI-as-a-service (AaaS) company.
AI agents are moving from pilots into everyday business operations. According to McKinsey’s The State of AI in 2026 survey, 40% of respondents at organizations with more than $1 billion in annual revenue said they were scaling AI agents in at least one business function, up from 27% a year earlier. Yet among all respondents, the share reporting that AI contributed positively to their organization’s EBIT held at 37%, essentially unchanged from 2025.
Adoption is rising, but enterprise-level impact has yet to follow at the same pace.
In my experience, this gap is often less about the technology itself. Today’s models are highly capable, and many enterprises have already done the hard work of getting agents into their workflows. What often lags is how those agents are operated once they’re in place: how strategic goals are set, how decisions are checked, how tasks are assigned and how systems improve over time.
Scaling agentic AI, in other words, depends as much on how it’s operated as on how widely it’s adopted. Here are five principles I believe enterprises should build around.
1. Strategy: Set Clear Business Goals And Metrics
Every AI deployment should begin with a clear answer to one question: What is this meant to achieve? Revenue growth, customer conversion, cost reduction and productivity improvement each call for different metrics and different timelines.
Without that clarity, teams tend to measure what’s easiest to see, such as activity, usage or the number of agents deployed. Those numbers can rise steadily while business results stay flat. It’s far more useful to agree on the outcome and the evaluation criteria before deployment, so AI is judged on what it delivers rather than how busy it looks.
This also makes later decisions easier. When the goal is clearly defined, it’s much simpler to tell whether an agent should be expanded, refined or retired.
2. Calibration: Assess Confidence And Risk Before Acting
Once the goal is clear, the next question is how reliably AI can do the job. As agents take on more decisions, they should be designed to assess their own capability and confidence, and weigh potential risks before acting.
In enterprise settings, small errors can have meaningful consequences. A mispriced offer, a misrouted customer request or an incorrect data entry can ripple through downstream systems. That’s why checking only the final output isn’t enough. Organizations should evaluate AI decisions at critical points across a workflow and put guardrails in place where the stakes are highest.
An agent that knows when to pause, flag uncertainty or hand a task back to a person is often more valuable than one that always produces an answer.
3. Adaptation: Match Models And Agents To The Task
Knowing AI can handle a task is only half the equation. The other half is deciding which model or agent should handle it.
Applying a single model to every scenario is simple but rarely optimal. A quick classification task and a complex, multistep analysis have very different requirements. Model routing, which directs each task to the most suitable model, helps match capability to need. For more complex work, agent orchestration allows multiple agents with distinct roles and expertise to work together, much like a well-structured team.
The aim is fit, not firepower. The most powerful model isn’t always the right one, and the right one may change as tasks and models evolve.
4. Learning: Turn Context And Feedback Into Better Decisions
For agentic AI to create lasting value beyond one-off automation, it needs to get better over time. That means feeding context, user feedback and the results of previous actions back into how agents make future decisions.
Many deployments still treat each task as a fresh start. The agent completes the work, and whatever could have been learned from the result is lost. Building feedback loops into workflows from the beginning lets systems improve with use, under appropriate human oversight, rather than depending entirely on periodic manual updates.
This is where much of the long-term value can come from. An agent that improves with every cycle compounds its value. One that doesn’t stay at the level it launched with.
5. Efficiency: Balance Performance And Resource Use
Efficiency matters more as agentic systems grow more complex. Multistep reasoning, repeated model calls and interactions between agents all consume compute, and those costs can climb quietly as usage scales.
The goal isn’t to minimize resource use for its own sake. It’s to deliver consistent outcomes without using more than the task requires, matching models and compute to the job and reviewing costs regularly as real usage data comes in.
Raising The Bar With More Autonomy
These SCALE principles reinforce one another. Strategic goals and metrics tell you what to measure. Calibrated confidence helps AI know when to act and keeps it reliable. Adaptive model selection makes it effective. Learning helps it improve, and efficient resource use keeps that improvement sustainable.
There’s a common assumption that as AI becomes more autonomous, it will need less management. I believe the opposite is true. The more decisions and actions AI takes on, the more sophisticated enterprises need to be about how they govern and operate it.
Autonomy alone isn’t the measure of success. What matters is whether AI can be operated reliably and efficiently, and whether its business impact keeps growing over time.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?