
As AI agents proliferate across the enterprise, traditional approaches to identity and access management are being rethought. CIOs and CISOs are looking for ways to effectively manage AI agents and the operational complexities they introduce. One of the key challenges is handling “AI orphans,” agents operating outside the visibility of the IT team or left behind by former employees.
When a former employee’s user account is disabled, the IT team must consider whether to disable all associated agent accounts or continue managing them. This is further complicated by the fact that some agents are connected to a service account instead of a single human, making permission decisions more complex. Hanah-Marie Darley, chief AI officer at AI agent management company Geordie AI, notes that many larger organizations have multi-use or multi-workflow agents tied to multiple human users, requiring multiple sets of permissions.
Stop the sprawl: 5 steps to wrangling AI orphans
According to Darley, one of the biggest challenges in transitioning from machine identity to non-human identity is the total volume and dynamic nature of how agents access tools and systems. This is different from machine identities, which used to be set and always accessed the same resources. There’s also the challenge of gaining visibility into the status of unused agents, which can leave unregulated agents lying dormant in the ecosystem.
Effective AI agent management starts with gaining visibility into all agents used within the enterprise and applying control policies to manage agent activity. Darley recommends that IT teams take a unified approach to understanding agentic operations, including where agents are, whether they’re active or not, and how their identities are working in practice. Allen Houchins, CIO of Fleet Device Management, emphasizes the importance of visibility into agent operations, including understanding how employees use their corporate devices and the finer details of agent instruction files and plugins.
The first step in simplifying identity management for AI agents is to inventory all agents, including internally built and third-party agents. IT teams need to communicate closely with internal development and technology teams to get a clear picture of all agents in use. Darley notes that most organizations don’t have a unified understanding of agentic operations, which is critical to governing agents throughout their lifecycle.
Agent Lifecycles and Ownership
Once inventory is complete, IT teams should map out agent lifecycles, including determining the business outcomes they want agents to deliver and how to measure their performance. This requires managing the full lifecycle of an agent, not just the initial deployment process, and continuously assessing whether agents are aligned with business outcomes and demonstrating ROI. Darley emphasizes the importance of tying all lifecycle management to value, as agents are judged by outcomes.
Every agent needs an owner, just like human employees report to a manager. Darley’s customers often have an accountable owner or human “parent user” responsible for an agent. Houchins agrees, stating that every single agent needs a named owner.
Control and governance are also essential in managing AI agents. Darley notes that larger customers have thousands of agents working within their enterprise, making it impractical for an AI governance team to manually review every agent. Independent AI systems can help speed up the process of managing agents. Rob Hughes, CISO at IAM solution provider RSA, emphasizes the importance of having the right governance of AI agents at the API or Model Context Protocol level to keep agents operating within a controlled environment.
Hughes shares an example of a case where an agent ID connected to an API gave it unfettered access to all of an organization’s data, highlighting the need for clear control policies governing which data and systems AI can access. Danny Jenkins, CEO of ThreatLocker, agrees that IT executives need to ensure they have control policies in place to define which files, documents, and applications AI agents can access.
According to Darley, the audit process must also be updated for agentic AI, with regular review processes to ensure agents are operating within set bounds. This requires thinking at both the micro and macro levels, assessing agent performance in isolation and as a cohesive whole. Darley shares an example of a customer that realized it needed to rethink its agent lifecycles after learning it was using about 19% of its AI budget on just automating scheduling, highlighting the importance of continuous assessment and governance.
With the pressure to quickly deploy AI within the enterprise, some agents will inevitably slip through the cracks, even with controls and governance in place. Darley notes that everybody will have orphan agents at some point, but it’s essential to think about how to manage the lifecycle of agents rather than just getting them up and running. The process of managing AI agents is complex, and Darley emphasizes the need for a unified approach to understanding agentic operations, including inventorying agents, mapping out lifecycles, and assigning ownership, as well as implementing control and governance policies.
Implementing Control Policies
Control policies should define which data and systems AI agents can access, and under what circumstances. This includes setting permissions at a file-permission level, application-permission level, and AI control level.
