The enterprise conversation surrounding artificial intelligence (AI) is shifting into a more serious phase. The initial enthusiasm for copilots and digital assistants is evolving into a pressing operational reality: as AI agents proliferate across various functions, systems, and workflows, the discussion transitions from novelty to control. Harpreet Singh Dhariwal, Solution Engineer at Salesforce, identifies this as the primary challenge of the agentic enterprise.
Dhariwal’s central assertion is clear: organizations are not heading towards a future defined by a singular omniscient enterprise assistant. Instead, they are moving towards an ecosystem where “every experience being built is leading to an agent,” resulting in a “plethora of agents” collaborating to address business challenges. This transformation alters the landscape of enterprise AI. While a single agent can be treated like a tool, a diverse array of agents constitutes a self-governing system, which necessitates robust governance.
The foundation of this governance structure lies in data. Dhariwal highlights three critical layers essential for effective governance within an agentic enterprise: data governance, API governance, and agent governance. Chief among these is data governance, which he identifies as fundamental. If agents leverage language models informed by enterprise data, the accuracy of their outputs relies heavily on the quality, context, and lineage of that data. As Dhariwal emphasizes, “it’s crucial to ensure that the data fed to LLMs is contextually accurate and has proper lineage.”
Consequently, concepts such as golden records, privacy, metadata cataloging, and lineage are no longer merely considerations of backend hygiene; they are now pivotal to business operations. In an agent-driven environment, poor-quality data can lead to flawed actions, poor decisions, and unreliable workflows.
Governance must extend beyond the data layer into the systems layer, which brings us to the second layer: API governance. Dhariwal makes an important distinction: enterprise agents not only provide responses, but they also execute actions—whether updating systems, creating records, triggering workflows, or interacting with APIs and other technical infrastructures. The governance question thus shifts from whether an agent can generate a correct response to whether it can engage safely and predictably with the broader technological ecosystem.
The full scope of the challenge becomes apparent with agent governance. Dhariwal underscores the necessity for discoverability, policy implementation, orchestration, and observability. He articulates that enterprises need “the ability to discover,” “the ability to apply policies,” and “the ability to orchestrate those agents.” This sequence is crucial; governance cannot be effective without first understanding which agents exist, their locations, their interactions, and the decisions they influence.
This is particularly significant as agents will be developed across various platforms, functions, and environments. Without a centralized method to discover and catalog these agents, governance will inevitably be fragmented from the outset.
The overarching takeaway is that the next phase of enterprise AI will be determined more by the discipline surrounding agent management than by the sheer number of agents deployed. While visible elements may include intelligence, automation, and speed, the foundational elements that ensure trust, scalability, and sustainability will be rooted in governance. Within the agentic enterprise framework, governance is not merely a supporting function; it serves as the control layer.
Disclaimer: The views expressed in this article are solely those of the speakers from the ETCIO Cloud Summit 2026, and do not necessarily reflect the views of ETCIO.
