By Madhu Sudan R, AVP, Modern Work, Infosys Microsoft Practice
The average global Fortune 500 company deployed fewer than 15 AI agents in 2025. By 2028, that number is estimated to go over 150,000. Yet, only 13 percent of enterprises are confident that they have the right AI agent governance framework in place. This is as per a recent Gartner research.
With agentic AI, work is increasingly speeding up, and workflows are also changing fundamentally. Agentic systems plan tasks, coordinate workflows, make decisions, and execute them in real time, often with minimal human intervention, intertwining governance and innovation intricately.
To hold up to scrutiny, every agent in the organization should be traceable, explainable, and consistent with organizational policies. Without this, the cracks can show up quickly. Agentic workflows may be redundant, with multiple agents doing the same job. Workflows may not synchronize with each other. Work output may vary for no clear reason.
An ungoverned system is harder to untangle every quarter it runs unchecked. Gartner estimates that by 2027, 40 percent of enterprises are expected to demote or shut down autonomous AI agents due to governance gaps that will be identified only after production incidents take place.
Leaders looking to scale their agentic deployment must consider various governance models.
Establish a centralized dashboard
The agentic control tower is an intelligent layer that operates as a centralized command and governance dashboard that can discover, monitor, and manage all the agents of the enterprise. It creates a single control plane to track operational telemetry and token consumption costs. It also ensures compliance with security protocols, while orchestrating multi-agent workflows.
In a supply chain management system, such an operations layer can watch every shipment and often predict service level agreement breaches before they happen, while triggering the next best action automatically. The action may be a call to a carrier or alerting a customer, while simultaneously raising a debit note.
Build an adaptive way to govern
Adaptive governance replaces static policies with fluid, context-aware guardrails. It empowers teams to operate with autonomy. Decision-making authority moves closer to the operational flow, while maintaining built-in observability and human-in-the-loop (HITL) escalation pathways.
In this model, governance focuses on current agentic use cases, embedding guardrails and “agent passports”- dictating what data agents can access, what APIs they can call, as well as their parameter schemas. Security and legal teams typically sit alongside technology teams in working groups, ensuring governance is a feature, not a bolt-on. Such programs iterate continuously, as risks and capabilities evolve.
Measure experiences alongside outcomes
An experience-driven model operationalizes Experience Level Agreements (XLAs), where controls are typically embedded seamlessly into the employee and developer experience.
Digital Employee Experience Management (DEXM) solutions fall in this category. Forrester reports that organizations are increasingly using such systems to improve how work actually gets done. The focus is on enablement, not restriction. For instance, insights from the DEXM system help guide employees towards approved AI tools instead of just blocking or restricting access.
Good agentic governance should account for what changing collaboration patterns and workflows mean for employees on the ground. Experience Management Offices (XMOs) are well positioned to track that shift, since operational data and employee feedback already fall within their scope. Pulling those threads together lets them shift focus from “Is this faster?” to “Is AI actually making work better?”
A project manager, for instance, could use an agent to pull project updates, consolidate milestones and risks, and automatically generate a status report, allowing them to focus on reviewing the output and making important decisions.
Balance autonomy with accountability
In enterprises where governance is ingrained, HITL earns its place, providing critical oversight for edge cases and high-stakes decisions, while routine activities often run autonomously.
The Reserve Bank of India has prioritized this approach. Its Guidance on Regulatory Principles for Model Risk Management, 2026 requires Regulated Entities (REs) to ensure: Board-approved Model Risk Management Framework (MRMF) for AI/ ML governance, independent validation for all models, and robust human oversight for AI models, among other things.
In conclusion
Every enterprise chasing agentic AI will eventually face the same question: not how many agents are running, but how much you can trust what they decide. Governance by design, human oversight, and an operating model that keeps pace with the technology – these aren’t constraints on speed. Instead, they are what make speed sustainable. In the agentic enterprise, trust is the real infrastructure.