AI Agent Identity Governance: Workspace IDs Expose a Gap

Agents Get Their Own Badges

Google's Gemini "coworker agents" are not just another chatbot feature. These persistent agents receive full Google Workspace identities, including email addresses on an @agents.company.com domain, calendars, Drive storage, and directory listings 5. In practice, an AI agent now sits in the corporate directory much like a new hire. That turns agent governance into something close to an HR process, with IT and security teams responsible for provisioning and offboarding 5.

Google says 90% of the Fortune 100 use Gemini Enterprise 5. At that scale, giving agents first-class identities could quickly make them some of the most numerous accounts in a large enterprise. The open question is whether organizations have the controls to manage them.

A Governance Vacuum Already in Place

The evidence suggests they mostly do not. A Cloud Security Alliance whitepaper calls non-human identity (NHI) governance the defining security gap of the agentic AI era. It argues that AI agent identities differ meaningfully from older machine identities such as service accounts and API keys 1. It cites a 2025 World Economic Forum analysis finding that 51% of organizations have no clear ownership of AI identities 1. It also points to research showing that 8% of enterprise identities lose any link to an HR system once their creator leaves. These orphaned accounts can persist indefinitely with no accountable human behind them 1.

Visibility is just as weak. Coverage of a Linx Security guide reports that only 5.7% of organizations have full visibility into their service accounts. A CSA survey cited in the same piece found that 68% cannot reliably tell AI agent activity apart from human activity 2. That piece also gives a 25x to 50x ratio for agents in modern enterprises. Its wording is ambiguous, but the point is clear: machine identities already far outnumber people 2. The CSA's own token-sprawl analysis adds that more than 16% of organizations do not track certain credentials at all 1.

Taken together, these figures sharpen the concern about the Gemini design. An agent that can send email, schedule meetings, and read shared drives under a directory-listed identity is a privileged account. If no one owns it and its actions look like a human's in the logs, a compromised or misconfigured agent could act for a long time before anyone notices. The practitioner advice is consistent across sources. Treat every agent as a privileged non-human identity with a named owner, a defined purpose, and a measurable access scope 2. The CSA layers lifecycle management, zero-standing privilege, and cryptographic workload attestation on top 1.

The Human Side of the Paradox

The governance gap is not only technical. Consultants tracking the Gemini rollout describe an "AI Paradox": employees who save hours with the agent often hide those gains. They fear being handed more work or being replaced 5.

Independent research supports this. A USC Marshall study by Eric Anicich and Jeslyn Brouwers surveyed 604 U.S. employees who use AI daily. Four in five believed sharing their techniques would help their teams. Yet nearly one in three withheld AI-related knowledge or workflows from coworkers or employers 3. Much of that reluctance came from distrust of the organization. Some workers also feared supervisors would treat efficiency gains as a reason to pile on work, or would draft them as internal AI spokespeople 3.

A Visier report points to the opposite distortion. It found that 48% of employees admit exaggerating their AI use or expertise at least some of the time, and 45% feel pressured to use AI 4. Three in five say their employer has no clear AI plan, or no plan at all. Training and transparency top workers' list of requested changes 4.

These findings can look contradictory, with some workers hiding AI use and others faking it. They share a cause, though: a low-trust environment without clear rules. In that environment, reported AI usage is an unreliable signal in both directions.

Reading the Collision

The two halves of this story point to one problem. Enterprises are being offered agents with real identities and real privileges at a moment when they cannot reliably inventory those identities. They also cannot measure how humans are actually using AI. Security teams lack ownership data and activity attribution. Managers lack honest usage data.

The practical conclusion is that agent identities should not be switched on simply because the platform supports them. Organizations that make progress will likely tie agent provisioning to existing identity-governance workflows, assign each agent an accountable human, and set expiry and review dates from day one. Just as important, they will need to tell employees plainly how productivity gains will be treated. Without that, both the security logs and the adoption metrics will keep telling leaders a story that does not match reality.