AI Agents are no longer just tools that answer questions. They can independently read emails, open tickets, query databases, change settings, and complete an entire task—often without a person reviewing each step along the way.
According to a recent NewCore analysis, enterprise identity governance—from onboarding and role changes to offboarding—was originally designed entirely around people. Today, however, AI Agents enter the enterprise in fundamentally different ways, without corresponding governance processes, creating an accountability gap.
Why Can’t Existing Identity Governance Keep Up with AI Agents?
Traditional service accounts execute only predefined logic, making their behavior predictable. AI Agents are different: they assess the situation and choose their next action independently. NewCore describes this distinction as a combination of “real permissions plus independent judgment,” which makes governance based solely on permission lists inadequate.
This also makes three previously straightforward questions much harder to answer: Which Agents exist in the environment? Which systems can they access? How can the enterprise control what they actually do? Agents are distributed across different platforms, each with its own management interface. A developer can build an Agent in a single afternoon, while a business user can assemble one using a low-code platform. Agents generated by automation platforms may even create other Agents, making even basic inventory management difficult.
More specific risks emerge in two scenarios. The first is “borrowed permissions.” If an Agent operates using credentials stored by its creator, anyone who can interact with the Agent may effectively borrow those permissions. This type of access usually does not appear in the user’s own access review report. The second is a “delegation chain”: a person invokes an Agent, the Agent invokes a tool, the tool triggers another service, and that service invokes another Agent. With every additional hop, it becomes harder to trace authorization responsibility back to the original person.
How Quickly Are Machine Identities Growing?
NewCore cites Palo Alto Networks’ 2026 Identity Security Landscape survey, which included 2,930 cybersecurity decision-makers. The article identifies it as a preview report, and because the figures are self-reported, they should be interpreted cautiously.
For every human identity, there are an average of 109 machine identities, 79 of which are AI Agent identities. Over the next 12 months, Agent identities are expected to grow by 85%, far outpacing the 56% growth projected for human identities. About 40% of AI Agents can already access organizational data, yet fewer than half of organizations apply basic lifecycle controls to Agents, such as behavior monitoring or credential revocation. Only a small minority classify machines with access to sensitive data as privileged identities.
The same article also cites a Gartner forecast reported by SD Times: by the end of 2027, more than 40% of agentic AI projects will be canceled because of inadequate risk controls. At the same time, Gartner predicts that by 2028, approximately one-third of enterprise software applications will include built-in agentic capabilities. Taken together, these figures suggest that governance is failing to keep pace—not that demand is fading.
A Governance Blueprint: Embedding Accountability into the Architecture
NewCore proposes several governance principles. It is important to note that NewCore is itself an identity governance service provider, and the latter part of its article discusses the design of its own products. The following summary focuses only on principles with broad relevance:
- Treat each Agent as an independent, first-class identity for inventory and governance purposes. Do not hide it inside an application or simply classify it as a service account
- Assign a clearly accountable person to every Agent. Ownership should transfer when organizational responsibilities change, rather than leaving behind unmanaged orphan accounts
- An Agent’s maximum permissions should never exceed those of the person it represents. When that person’s permissions change, the Agent’s permissions should change accordingly
- Govern “behavior,” not just “applications”: define which actions can be performed directly, which require human approval, and which are entirely prohibited. Bulk exports and irreversible changes should be subject to additional restrictions
- Maintain a complete record of every action, including the person involved, the Agent used, the tools invoked, the systems affected, and the basis for approval or rejection
NewCore summarizes these trade-offs with a clear principle: when uncertainty arises, the system should issue a clear, auditable “no,” rather than quietly allowing a “yes.”
What This Means for Traditional Industries Adopting AI Agents
When traditional industries discuss AI Agent adoption, system integration is usually the first concern: connecting to an ERP, a CRM, or a particular internal database. NewCore’s analysis shows, however, that the real source of problems is often not whether systems can connect, but how work is handed from people to Agents—and then back to people. Integrating human workflows is just as important as integrating systems, yet it is much more easily overlooked.
Before deployment, enterprises should ask themselves several questions: Is every Agent currently in use mapped to a clearly defined role or accountable owner? Should human approval be required before a task moves to the next step? When that person leaves the role, what happens to the Agent and its permissions? If these questions cannot be answered, the governance design has not kept pace with the speed of adoption.
This type of task orchestration, which must account for both systems and people, is precisely the kind of scenario products such as EgentWrX are designed to address. A digital counterpart is tied to a role rather than an individual, so when a person leaves, it remains in place awaiting handover instead of becoming an unmanaged account. Workflows between tasks can include human approval gates before handoff, rather than running automatically from beginning to end. When cross-functional support is needed, each party responds through channels governed by its own scope of authority, instead of exposing a single Agent as a shared account for the entire company. Thinking through these issues before deployment is more important than deciding which system to integrate first.
Governance and system integration should be designed together, rather than relying on documentation to fill the gaps after deployment.
FAQ
Are these figures based on a single vendor’s survey, making them unrepresentative of the broader industry? The Palo Alto Networks survey included 2,930 cybersecurity decision-makers, and the figures are self-reported. NewCore also cautions readers to interpret them conservatively. However, the overall direction aligns with other industry observations: the number of Agents is rising rapidly, while governance mechanisms are generally lagging behind the pace of deployment.
Why are Agents more difficult to govern than service accounts? Service accounts execute only predefined logic, so their behavior is predictable. AI Agents assess situations and independently choose their actions. NewCore describes this as a combination of real permissions and independent judgment, making traditional governance based solely on permission lists insufficient.
Should enterprises pause AI Agent adoption until governance mechanisms mature? The article does not recommend pausing adoption. Instead, it argues that accountability should be designed into the architecture: map every Agent to a clearly accountable owner, add human approval before handoffs, and maintain complete decision records. These controls can be implemented alongside deployment rather than added retroactively.
Does Gartner’s forecast that 40% of agentic AI projects will be canceled mean the trend is losing momentum? The forecast points to project cancellations caused by inadequate risk controls, not a decline in the broader trend. The same source also indicates that approximately one-third of enterprise software will include built-in agentic capabilities by 2028. The problem is that governance is not keeping pace—not that demand is disappearing.
References
- NewCore, “It’s Not Me, It’s My Agent”: Governance and Accountability in the Age of AI Agents, a blog analysis by an identity governance service provider (citing Palo Alto Networks’ 2026 Identity Security Landscape survey, a Gartner forecast reported by SD Times, and perspectives from the Google technical blog)
