Businesses are giving artificial intelligence tools and agents access to corporate systems faster than they are establishing consistent controls and audit records, according to new research by identity management company JumpCloud.
The company’s Enterprise AI Access Is Outpacing Control study, published on 7 October, surveyed 250 IT decision-makers in the United States and United Kingdom. The sample was divided evenly between organisations using Google Workspace and those primarily using Microsoft 365.
JumpCloud found that 55% of surveyed organisations use or test AI agents capable of changing systems, permissions, records or workflows. Among respondents from organisations using agents, 59% said the agents were not fully integrated into the identity and access management policies applied to human users.
That 59% is a finding about the subset using agents, not 59% of all companies or necessarily of the full 250-person sample. The report describes 41% of agent-using organisations as having fully integrated agents into the relevant human-user identity policies.
Across the wider sample, 72% reported a moderate-to-large gap between the information their AI systems were permitted to access and what their organisations could prove those systems had actually accessed. This is a self-reported assessment of auditability, rather than an independent inspection of system logs.
The survey also found that 99% of respondents expressed confidence in their ability to audit AI activity, while only 37% met the research’s definition of being fully prepared with the records and evidence such an audit would require.
These measures address different questions. Confidence records how respondents assess their capabilities; preparedness reflects JumpCloud’s criteria; and the 72% finding concerns the perceived difference between permissions and demonstrated activity. They should not be combined as though they measure one identical control failure.
The report says 34% of surveyed organisations allow AI access to identity and access management systems. Such platforms can determine who receives permissions, how credentials are managed and when access is withdrawn, making the ability to identify automated activity particularly important.
Traditional identity procedures are often designed around an employee signing in to a known service. An agent may operate continuously, act on delegated permissions and call several applications within one workflow. Establishing which human authorised those actions and whether they stayed within the intended task can require more detailed records.
For example, an agent authorised to retrieve information for a customer report might access a document repository and a CRM system. The existence of valid credentials would not by itself show which records were consulted, what information was copied or whether the resulting report contained material beyond the employee’s authority.
Businesses deploying such software may therefore need to define the agent’s identity, owner, permitted tasks, access duration and approval points. They also need logs sufficient to reconstruct sensitive actions and determine whether a user, administrator or automated process initiated them.
JumpCloud reported that 92% of respondents had experienced at least one AI-related incident or exposure in the previous 12 months, while 36% had investigated an AI-related concern without being able to determine what the AI had accessed. These are vendor-survey responses covering a range of incidents and exposures, not verified breaches affecting 92% of all organisations.
The analysis also compared respondents by productivity platform. Some 21% of Microsoft 365-primary organisations fell into the lowest tier for AI auditability preparedness, compared with 11% of Google Workspace-primary organisations in this survey.
That comparison should not be treated as proof that one software provider is intrinsically more secure. The sample is limited, respondent organisations may differ in other ways, and the results concern the audit arrangements reported by their IT decision-makers.
JumpCloud further found that respondents reported AI access to an average of three types of corporate system or data. The categories included customer records, financial and procurement applications, endpoint administration and identity systems, illustrating how automated tasks can extend beyond a single productivity platform.
When records of AI activity are distributed across several applications, investigators may have to connect information about the agent, the employee who authorised it and each system the agent visited. Missing or inconsistent logs can make it difficult to establish precisely what happened after a security incident or internal complaint.
These questions are relevant to operational governance as well as cybersecurity. A business may need to demonstrate why an automated purchase, customer record change or access decision occurred even when the action was technically permitted by the system.
The research was produced by a vendor selling identity and device management services, and the published summary does not provide a basis for treating its 250 respondents as a representative sample of all UK and US employers. Its percentages should therefore be read as findings within that respondent group.
The immediate practical issue is whether organisations can connect an AI agent’s permissions to its actual actions and to an accountable owner. JumpCloud’s survey suggests that many respondents see a gap between those requirements and the evidence available from their current systems.
You must be logged in to post a comment.