Fleet survey finds widespread AI use in device management, but slow recovery from bad changes

AI-written work is reaching employees’ devices, but undoing a bad change remains slow for many administrators. Fleet’s Apple in the Enterprise survey of more than 250 enterprise IT practitioners found that 86% allow AI-generated output onto production devices after review, while nearly 70% cannot reverse a bad configuration within an hour.

AI is already part of the administrator’s toolkit

The survey puts AI assistants alongside the tools used to manage corporate hardware. Some 57% of respondents regularly use large language model assistants, compared with 53% who use mobile device management, or MDM, consoles—the interfaces administrators use to manage devices.

The work extends beyond asking an assistant for advice. In the previous 30 days, 61% had used AI to create a device-management profile or configuration, 62% to write scripts or code, and 68% to troubleshoot problems or analyze logs. Those findings appear in Cybersecurity Insiders’ October 13 coverage of the Fleet report.

Fleet CEO and co-founder Mike McNeil describes that work as engineering, even when the job title remains administrator. In an October 8 interview with Computerworld’s Jonny Evans, he argued that AI speeds up authoring without increasing review capacity. His warning concerns the controls around deployment, not simply whether teams use AI.

69%Cannot roll back within an hour

Respondents reported being unable to reverse a bad configuration within an hour.

43%Need more than a day

Respondents said recovery from a bad configuration takes more than a day.

26%Same-day recovery, with manual work

Respondents said they could recover the same day, but only with manual intervention.

Review and recovery are different protections

Only 28% of respondents combine human review of AI-generated changes with version control, which keeps a record of file changes. That is a narrower safeguard than the survey’s broad measure of output reaching devices after some form of review: approval does not necessarily come with a tracked, reversible change.

McNeil frames the risk around two questions: how much access the code has, and how difficult recovery will be. Scripts running as root can take full control of a machine. He also warned that MDM is a trusted route that can bypass macOS security prompts, making a bad or malicious management-delivered change especially consequential.

These are respondents’ accounts of their practices and recovery capabilities, not a count of AI-caused incidents. McNeil’s argument is that slow recovery creates exposure even without such a count. The findings should therefore be read as a warning about preparedness, rather than evidence that AI-written changes caused widespread failures.

A tracked change is not always an undo button

McNeil recommends treating device configurations like application code. He says Fleet’s own management system supports this approach, while arguing that the pattern can work with any tool that has an application programming interface, or API—a way for software to interact with it. His proposed workflow has three parts:

1. Store device profiles, scripts and policies as files in a Git repository, where changes can be tracked.
2. Submit each change for a second person’s review before automation applies it to devices.
3. Reverse the recorded change when a configuration needs to be restored.

There is an important limit: reverting the record changes what a configuration says, but cannot undo a script that has already run. McNeil therefore favors declarative configuration—specifying the desired settings—over scripts that execute instructions wherever possible. The distinction separates restoring a policy from repairing consequences that code has already produced.

The survey found 78% already use or are piloting declarative device management. But administrators are also managing operational pressure: 31% identified keeping up with Apple’s release pace as the hardest part of their job, compared with 5% citing budget and headcount. And this is rarely an Apple-only workload. Just 8% of respondents manage Apple devices exclusively.