- JumpCloud's survey finds 99% of IT leaders are confident they can audit AI activity, but only 37% are fully prepared with the records and logs an audit needs.
- Agents change what they can reach, and shared credentials blur who acted. A policy says what should happen. Only a log shows what did.
- Leaders should ask whether each agent has its own identity, how often permissions are reviewed, and whether the team can answer an auditor from records.
Ask an IT leader whether they could audit what their AI agents did last month, and almost all say yes. Ask them to produce the records, and the answer changes. That distance between confidence and evidence is the main finding in new research from JumpCloud. It matters more than any single number in it.
What the survey found
JumpCloud is an identity and device management vendor. It surveyed IT leaders about AI access. Almost all respondents, 99%, said they could audit AI activity. Far fewer, 37%, said they hold the full set of records, logs and evidence that such an audit would demand.
Permissions tell a similar story. JumpCloud says just 12% of organizations call their AI permissions tightly scoped and easy to review. Close to half describe them as broad, hard to review or uneven.
The survey also asked about proof. Nearly three in four organizations, 72%, see a moderate or large gap. On one side is what AI is allowed to reach. On the other is what they can show it actually reached.
Only 36% can fully prove that AI held just the permissions it needed. Fewer than half can fully document what an agent did. The same goes for which identity authorized it and whether the action followed policy.
Why agents strain a familiar rule
Least privilege means giving an identity only the access its job needs. It works well for people with stable roles. A finance clerk needs the finance system. A yearly review catches most drift.
JumpCloud argues that agents break that pattern. One workflow may touch several applications, data sources and APIs. APIs are the connections software uses to talk to other software. Each new integration can widen what the agent can reach. The permission set keeps moving.
That makes review timing important. According to JumpCloud, close to half of organizations review AI permissions monthly or quarterly. Only 14% review them continuously or through automation.
JumpCloud calls a periodic review a snapshot of an access model that may keep changing. The survey links weak review habits to messy permissions. Among organizations with weaker or reactive reviews, 61% describe their permissions as broad or inconsistent.
The shared key problem
Attribution is a more basic issue. Among organizations that use agents, 63% rely on riskier methods. These include shared service accounts and long-lived API keys. They sit alongside stronger sign-in methods, not in place of them.
Think of a door code that a whole team shares. The log shows someone entered at 2 a.m. It cannot show who. A shared account or a key that never expires works the same way. JumpCloud says these methods make it harder to tie activity to the agent responsible.
The exposure is not small. JumpCloud reports that more than half of the organizations surveyed use or test agents that can change systems, permissions, records or workflows. Among those using agents, 59% have not fully brought them under the identity policies that apply to human staff.
Policy is a promise, the log is the proof
The lesson here is that a policy and an audit trail answer different questions. The policy says what should have happened. The trail shows what did. A firm can hold a careful policy and still fail an audit, because nobody can show the agent stayed inside it.
This changes how to read the 99% figure. Confidence is an opinion. A log is a fact. With only 37% fully prepared, executives who approve AI deployments may be leaning on the opinion.
Limits of the data
This is vendor research. JumpCloud sells identity tools. Its report page also promotes a maturity model and a free trial of its platform.
The material available here does not give the sample size, the survey method or who responded. Read the figures as one company's survey. They are not an industry measurement.
Questions to put to your team
First, can we list every AI agent that can change a record, a permission or a workflow? Does each one have a named owner? Second, does each agent have its own identity, or do several share a service account or long-lived key?
Third, how often do we review agent permissions? Could the review run continuously instead of quarterly? Fourth, suppose an auditor asks what one agent did last Tuesday. Could we show who authorized it and whether it followed policy? Could we do so from records within a day?
Fifth, have we tested that answer, or are we assuming it? That last question is the one that counts. An organization that cannot show what its agents did has not yet earned confidence in them.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: JumpCloud.
