Dr. Sanjay Kumar, a GenAI and Data Science Product Leader with more than 15 years in AI, MLOps, and cloud analytics, published an opinion piece on Aug 28, 2026, at 06:00am EDT. The article, hosted on Forbes through the Forbes Technology Council, argues that the real challenge in agentic AI is not technical deployment but leadership judgment.
The piece, written for CIOs and CTOs, contends that judging AI progress by counting the number of agents launched is the wrong measure. Deploying an agent is easy, Kumar writes. Deciding where it belongs, what authority it should have, and how people should work with it is harder.
Why Agentic AI Differs from Chatbots
Agentic AI represents a significant step beyond simple chatbots. Agents can gather information from several systems, complete workflow steps, and act without waiting for a person. This autonomy makes them promising but harder to control.
Many organizations have already introduced agents into their operations. Yet Kumar warns that not every process needs one. Putting an AI agent into a simple workflow can add cost and uncertainty without improving the outcome. For predictable, repetitive tasks, traditional automation may be cheaper and more reliable.
The article advises leaders to start with business problems, not technology. Cost reduction should not be the entire business case. Larger returns may come from faster planning, better service, earlier risk detection, or revenue opportunities.
Presenting agentic AI mainly as a way to reduce headcount narrows the opportunity and creates anxiety among employees. Kumar suggests a broader vision that includes growth and quality improvements.
Data, Context Engineering, and Architecture
Agents do not require perfect data, but they do need dependable data and guidance on trusted sources. Cleaning every dataset before beginning is rarely practical, the author notes.
A general-purpose model cannot safely invent definitions like churn, customer value, revenue, risk, performance, or policy exception. Agents must understand company-specific definitions and know which metric takes priority when reports disagree.
They must also recognize when evidence is insufficient. Without context engineering, an agent can produce an answer that sounds convincing and is completely wrong for the business.
Context engineering is becoming an important enterprise capability, according to the article. This discipline involves defining the boundaries, rules, and priorities that guide agent behavior.
Companies should be cautious about embedding agent logic deeply inside ERP, CRM, and core platforms. Doing so can create technical debt quickly.
Kumar recommends building modular layers above the core environment. These layers handle workflows, model access, memory, and orchestration. Modular design allows organizations to change models, replace vendors, and redesign workflows without rebuilding critical systems.
This approach also supports gradual evolution as agent capabilities improve. Leaders can test new models or adjust workflows without disrupting the underlying business systems.
Governance, Authorization, and Autonomy
Traditional AI governance focuses on privacy, bias, cybersecurity, compliance, and model risk. Agentic AI adds a new question: authorization.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Kumar highlights the difference between recommending versus approving a payment, and suggesting versus sending a customer response. These distinctions determine how much risk an organization accepts.
Autonomy should be earned gradually. The article outlines a path from shadow mode, where agents observe but do not act, to supervised execution, and finally to broader authority. Broader authority should come only after evidence of reliability under normal, unusual, and hostile conditions.
Traceability should be designed into the system from the beginning. It should not be added when an auditor asks for it.
Failure should be tested deliberately. This includes scenarios with missing data, conflicting instructions, and misleading documents. The goal is to understand risk before a small weakness becomes a serious incident.
People, Trust, and Adoption
Employees should help design workflows and define acceptable performance. Subject-matter experts should be involved at the beginning, not after a prototype exists.
People are more likely to use a system they helped shape and understand. Trust is necessary for adoption, and trust comes from involvement and transparency.
The best agents will capture and extend the practices of the strongest performers. This means studying how top employees work and encoding those approaches into agent behavior.
Training, accountability, and process redesign must develop alongside the technology. Leaders cannot simply deploy agents and expect employees to adapt without support.
A Leadership Test
Kumar concludes that leadership determines whether agentic AI creates lasting value or becomes an expensive experiment. Leaders should begin with focused problems, show measurable value, and expand carefully.
The article is a Council Post, which means it carries expertise from Forbes Councils members and is operated under license. The opinions expressed are those of the author. Membership in the Forbes Technology Council is fee-based.
The Forbes Technology Council is an invitation-only community for CIOs, CTOs, and technology executives. The article also includes a voice experience generated by AI, reflecting the broader use of generative tools in media production.
For technology leaders in 2026, the message is clear. The number of agents in production matters less than the quality of decisions about where they operate, how much authority they hold, and how humans integrate with them.
The path forward requires patience, deliberate testing, and a commitment to involving people in the design process. Technology alone will not deliver value. Leadership will.