Stop the sprawl: 5 steps to wrangling AI orphans

Stop the sprawl: 5 steps to wrangling AI orphans
AI agents create a management problem before they create a technical one: they blur ownership, access rights, and accountability. For CIOs, the immediate priority is to treat agents as a governed class of enterprise identity, not as add-ons to user or service-account processes. That means establishing clear decision rights for who can approve, deploy, modify, suspend, and retire agents—especially those embedded in SaaS platforms or created by business teams outside central IT.

The bigger implication is operating-model design. An inventory alone will not control sprawl unless it is linked to lifecycle policy: business sponsor, named operational owner, approved data scope, expected outcome, review cadence, and decommission trigger. This is where many organizations will find their existing IAM, CMDB, and software asset management disciplines only partially fit for purpose. Leaders should expect to define new controls for orphan detection, inherited permissions, and cross-agent interactions.

There is also a portfolio question. Individually useful agents can still become collectively wasteful, duplicative, or risky. Continuous audit should therefore measure not just technical behavior but business contribution, cost-to-run, and overlap with other automation. Some agents will merit preservation as reusable assets; others should be shut down quickly to reduce attack surface and spend.

Practical next questions for leadership:

  • Who is accountable for every production agent, including vendor-supplied ones?
  • What events automatically trigger access review or retirement?
  • Where do agent permissions exceed the human owner’s intended authority?
  • Which metrics show value at both agent level and portfolio level?

As AI agents proliferate across the enterprise, the old rulebooks are being thrown out. CIOs and CISOs are rethinking traditional approaches to identity and access management to effectively address AI agents — and the operational complexities they introduce.

The process of managing AI agents includes determining how to handle “AI orphans,” agents operating outside the visibility of the IT team and those left behind by former employees. When a former employee’s user account is disabled, the IT team must also consider whether to disable all agent accounts or continue managing them.

Employees aren’t the only ones flooding the enterprise with AI agents. Enterprise IT teams must also secure and manage agents introduced via third-party vendor software. To add to that complexity, some agents are connected to a service account instead of a single human, which can complicate permission decisions.

At their most complex, agents can have numerous owners and functions simultaneously. Many larger organizations have multi-use or multi-workflow agents tied to multiple human users, explained Hanah-Marie Darley, chief AI officer at AI agent management company Geordie AI. An individual agent might therefore require multiple sets of permissions, according to which person is using it and for which workflow.

“One of the biggest challenges transitioning from machine identity into non-human identity has been the total volume and dynamic nature of how agents access tools and systems,” Darley said. “It’s really different from machine identities because they used to be just kind of set — they would always access the same resources.”

There’s also the challenge of gaining visibility into the status of unused agents, Darley said. For example, an employee might use a Claude agent but later abandon or shelve it in favor of a different agent, without alerting IT to the switch. This can leave an unregulated agent lying dormant in the ecosystem.

Effective AI agent management therefore starts with gaining visibility into all agents used within the enterprise and applying control policies to manage agent activity. The following are five steps to simplifying identity management for AI agents:

1. Inventory your agents

IT teams need to get visibility into both internally built and third-party agents, according to Darley. Internal development and technology teams need to be in close communication with the security teams managing third-party agents. If enterprises can’t see their agents, they won’t be able to govern them, she said.

“Most organizations don’t tend to have that fully kind of unified understanding of agentic operations, and that’s really critical to governing agents throughout your lifecycle — understanding where you have agents, whether they’re active or not, how their identities are working in practice, and if you need to manage them out,” Darley said.

Allen Houchins, CIO of Fleet Device Management, explained that visibility into agent operations requires IT teams to have a clear understanding of how employees use their corporate devices.

“It’s really about visibility and inventory of all of your devices — and not only the agents and the applications that are running but really getting into the finer details of agent instruction files, plugins, MCPs, extensions, even local [AI] models that people are installing,” he said. “If you don’t have this visibility, you can’t govern it, control it or make any business decisions about it.”

2. Map out agent lifecycles

Once inventory is complete, there may be some visible gaps in operations where AI could be introduced. But before deploying any new AI agents, CIOs should have a clear picture of the business outcomes they want those agents to deliver, as well as a way to measure their performance.

To accurately assess whether agents are aligned with business outcomes and demonstrate ROI, CIOs need to manage the full lifecycle of an agent, not just the initial deployment process. This must also be a continuous assessment, not a one-off measurement.

“You have to tie all lifecycle management to value because ultimately you judge an agent by outcomes,” Darley said.

This also means following the lifecycle through to its end: There will be instances where an agent no longer delivers on business outcomes or becomes too expensive to maintain.

Good lifecycle management should flag these in near-real time, to minimize financial and operational impact. To prepare for this, CIOs can create policies for when it’s time to retire or delete an agent, such as closing off that agent’s permissions and ensuring it isn’t still communicating with other agents.

Alternatively, IT teams might decide it’s best for the agent to be “saved in a storage bucket somewhere, because there was all this thought and knowledge that went into it and the next person in that role might have an interest in it,” said Rob Hughes, CISO at IAM solution provider RSA. Whichever option CIOs choose, there should be a clear process to update the agent’s access.

3. Every agent needs an owner

Just like human employees report to a manager, AI agents need supervisors. Many of Darley’s customers have an accountable owner or human “parent user” responsible for an agent.

Houchins echoed the sentiment. “You need to give every single agent a named owner,” he explained.

Raj Ananthanpillai, CEO of identity screening company Trua, recommends assigning employees a “trust token,” which they obtain by completing biometric and government ID screenings. That way, a verified human is linked to the agent making an online transaction, for example. He likened it to TSA Precheck, in that the Transportation Security Administration has vetted passengers and assigned each a unique Precheck number that can be validated when a flight is booked.

“People should bind all of their agents in the company to the individual,” Ananthanpillai said.

4. Control and governance

Some of Darley’s larger customers have thousands of agents working within their enterprise, so there’s no practical way for an AI governance team to manually review every agent, she said. Speeding up the process of managing so many agents requires the use of independent AI systems, Darley added.

Having the right governance of AI agents at the API or Model Context Protocol level is also important for keeping agents operating within a controlled environment, Hughes said.

“We’ve seen a case where an agent ID connected to an API, and the API actually gave it unfettered access to all the [organization’s] data,” Hughes said. “That was not even the access that the agent ID had or was expected to have. It was expected to have the access of the owner of the agent, the person that created it.”

Danny Jenkins, CEO of cybersecurity company ThreatLocker, agreed that IT executives need to ensure they have control policies governing which data and systems AI can access.

“Make sure that you are very clearly defining, at a file-permission level, at an application-permission level, a ring-fencing level, at an AI control level — whatever way you want to do it — which files, which documents, which other applications on your computer this AI agent can talk to, and when it does talk to them, that those permissions flow through,” he said.

5. Agentic audits

The audit process must also be updated for agentic AI. Auditing agent performance shouldn’t be a one-and-done exercise, but a regular reviewal process, Hughes said.

“There’s so much more complexity wrapped up in these new non-human identities than prior,” Hughes said. “It’s no longer enough to do a ‘standard kind of audit’ of service accounts to review agent ownership quarterly or yearly. Really what you need is a continuous review of what the agents are doing and making sure they’re not going out of the bounds that you set.”

CIOs also need to think at both the micro and macro levels. Performance assessments of agents shouldn’t be conducted only in isolation but also as a cohesive whole.

Darley shared 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. Those agents were performing effectively on the individual level but weren’t efficient when considered as a fifth of AI investment — meaning they ultimately weren’t delivering the needed outcomes.

Still, with the pressure CIOs are under to quickly deploy AI within the enterprise, some agents will slip through the cracks — even with controls and governance in place.

“Everybody at some point will have orphan agents,” Darley said. “That’s OK, so don’t feel bad. But think about how you want to manage the lifecycle of agents rather than just getting them up and running.”

How are you handling orphaned agents in your enterprise? Let us know at [email protected].

Original Post

Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.

Leave a Reply