Dr. Sanjay Kumar is a GenAI & Data Science Product Leader with 15+ yrs in AI, MLOps & cloud analytics driving enterprise innovation.
There’s a growing tendency to judge a company’s AI progress by counting how many agents it’s launched. I think that’s the wrong measure. Deploying an agent is easy. Deciding where it belongs, what authority it should have and how people should work with it is much harder. Those choices separate lasting value from market-following pilots.
Agents are often described as smarter chatbots, but there’s a problem. A chatbot might summarize a report, answer a question or even draft an email. An agent can gather information from several systems, complete steps in a workflow and sometimes act without waiting for a person at every stage. That ability makes agentic AI promising but also harder to control.
Many organizations have already introduced agents into the mix. For CTOs and CIOs, the question is now whether companies can use them at scale without creating a fresh set of security, operational and governance problems.
Knowing Where An Agent Will Help
Not every process needs an agent. If a task is predictable, repetitive and governed by clear rules, traditional automation may still be cheaper and more dependable. Putting an AI agent into a simple workflow can add cost and uncertainty without improving the outcome.
Leaders should start with the business problem rather than the technology. Where do decisions repeatedly stall? Where are experienced employees doing administrative work instead of applying their expertise? Where are customers waiting because one team can’t see information that another team already has? These questions are far more useful than asking everyone to “find an AI use case.”
Cost reduction will be part of the conversation, but it shouldn’t become the entire business case. Presenting agentic AI mainly as a way to reduce headcount narrows the opportunity and creates understandable anxiety. The larger return may come from faster planning, better service, earlier risk detection or revenue opportunities that teams lack time to pursue.
Data
Many companies assume they must clean every dataset before they can begin. That sounds responsible, but it’s rarely practical. Agents don’t require perfect information. They require information dependable enough for the decision at hand along with guidance about which sources should be trusted.
An agent must understand what the organization means by customer value, revenue, risk, performance or a policy exception. It needs to know which metric takes priority when two reports disagree. Just as importantly, it must recognize when the available evidence is insufficient.
Imagine an agent asked to identify customers who may be preparing to leave. Before producing a useful answer, it needs the company’s definition of churn, its method for calculating customer value, the signals that indicate declining engagement and the rules governing retention offers. A general-purpose model can’t safely invent those definitions.
This is why context engineering is becoming an important enterprise capability. Companies need a practical way to connect current data with their definitions, policies, permissions and decision rules. Without that foundation, an agent can produce an answer that sounds convincing and is completely wrong for the business.
Architecture
Companies should be cautious about embedding agent logic deeply inside ERP, CRM and other core platforms. Those systems are expensive and difficult to change, and adding fast-evolving AI capabilities directly to them can create technical debt quickly.
An easier approach is to place an agent’s workflows, model access, memory and orchestration in modular layers above the core environment. This allows the organization to change models, replace vendors and redesign workflows without rebuilding critical systems whenever the market shifts.
Governance
Traditional AI governance has focused on privacy, bias, cybersecurity, compliance and model risk. Agentic AI adds a more immediate question: What’s this agent authorized to do? Recommending that a payment be approved isn’t the same as approving it. Suggesting a customer response is different from sending it. The distinction matters.
Autonomy should be earned gradually. A new agent might begin in shadow mode, offering recommendations but taking no action. Later, it could move into supervised execution, with an employee approving important steps. Broader authority should come only after the company has evidence that the agent performs reliably under normal, unusual and hostile conditions.
Leaders also need visibility into how the agent reached its decision. What information did it use? Which tools did it access? Where did uncertainty appear? Why did it choose one action over another? Traceability should be designed into the system from the beginning, not added when an auditor asks for it.
Failure should be tested deliberately. What happens when data is missing or instructions conflict? Can a misleading document manipulate the agent’s behavior? The goal is to understand the risk before a small weakness becomes a serious incident.
Organizational Buy-In
Employees who understand the work should help design the workflow, define acceptable performance and decide where human judgment remains essential. The best agents will capture and extend the practices of the organization’s strongest performers. This requires subject-matter experts at the beginning, not after a prototype has been built.
It also requires trust. People are more likely to use a system they helped shape and understand. Leaders should begin with focused problems, show measurable value and expand carefully. Training, accountability and process redesign must develop alongside the technology.
Conclusion
The companies that lead in agentic AI will combine ambition with judgment, choose worthwhile problems and bring employees into the transformation. Models may power agentic AI, but leadership will determine whether it creates lasting value or becomes another expensive experiment.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?