What the survey found
JumpCloud surveyed 250 IT decision-makers in the U.S. and U.K. at organizations with 200 to 2,500 employees. The survey was fielded July 8–20, 2026. That detail comes from JumpCloud's own methodology notes, as does the rest of the data below. The company reports a 95% confidence level and a ±6.1% margin of error. Respondents were split evenly between Microsoft 365-primary and Google Workspace-primary shops, 125 each. All were using or testing AI assistants, agents that can take action, or both.
The headline figures:
- 59% haven't fully integrated agents into IAM. This applies to organizations using agents, a base of 137. Only 41% fully integrate agents into the same IAM policies used for humans. The rest do it partially or manage agents through separate policies and tools.
- 72% report a moderate-to-large gap between what AI is permitted to access and what they can prove it accessed. Only 4% report no meaningful gap. This is respondents' self-assessment. It does not mean 72% of AI tools exceed their permissions.
- 36% investigated an AI-related concern in the past 12 months without being able to determine what the AI had accessed.
- 92% reported at least one AI-related incident or exposure in the past year. The category is broad. It does not mean 92% suffered a confirmed agent-caused breach.
- 99% say they are confident they can audit AI activity. Only 37% say they are fully prepared with the records, logs and evidence an audit would need.
The Microsoft 365 versus Google Workspace comparison
The detail most relevant here is the platform split. JumpCloud says 21% of Microsoft 365-primary organizations land in the lowest tier of its AI Auditability Maturity Model. The figure for Google Workspace-primary organizations is 11%. Among agent users, full IAM integration is 36% for Microsoft 365-primary organizations and 45% for Google Workspace-primary ones.
This is not a verdict on Microsoft's security. Several caveats apply:
- The tiers come from a vendor-built model. It weighs evidence, permission controls, governance, security layers and audit readiness.
- The data is self-reported and observational. It doesn't show that the suite itself caused the difference.
- A sample of 125 per group, limited to mid-sized firms in two countries, can't speak for all Microsoft 365 customers.
- Google Workspace-primary organizations more often reported AI reaching financial, billing or procurement systems (42% versus 30%). They also more often reported AI reaching documents and shared drives (38% versus 26%). Their auditability foundations looked stronger on the lowest-tier measure but their AI reach was broader. Those are separate measures.
If their suite couldn't reliably audit AI, 55% of all respondents would keep it and add controls. Another 45% would reassess it or consider migrating. Microsoft 365-primary shops more often favored tightening permissions inside the existing platform (35% versus 21%). Google Workspace-primary shops leaned toward adding a separate identity or security layer (30% versus 24%).
Why agents make this harder
An assistant mostly reads. An agent can change systems, permissions, records and workflows. JumpCloud says 55% of respondents use or test agents with that kind of reach, and 34% report AI access to IAM systems. Organizations use an average of 2.7 AI tools across three types of systems or data. Microsoft Copilot and ChatGPT were each among the most widely used or tested, at 40%, with Gemini at 29%.
Other findings suggest the audit trail is thin where agents act:
- Approval: 46% let high-risk agent actions happen automatically and log them for later review. Only 18% require human approval first.
- Attribution: Only 38% can fully identify the identity that authorized an AI action. Only 36% can fully prove the AI had only the permissions it needed.
- Authentication: 63% of agent environments still include higher-risk methods such as shared service accounts or long-lived API keys. Only 37% rely exclusively on stronger methods.
- Permissions: 49% describe AI permissions as broad, hard to review or inconsistent. Only 14% review them continuously or automatically.
- Policy: Only 29% have a formal AI usage policy, and 26% govern non-human identities.
Post-action logging only works if the log can answer who, with what authority, and what happened. The survey suggests many shops can't yet answer all three.
JumpCloud's Agentic IAM Lifecycle
JumpCloud proposes four stages. This is the company's own framework, not a Microsoft standard or an independently validated benchmark.
- Discover every agent, including shadow deployments. 30% of respondents report shadow AI that IT can't fully monitor.
- Register each agent with a purpose, intended scope and accountable human owner.
- Manage access by right-sizing it and using time-bound permissions.
- Govern through continuous auditing, current entitlements and human checkpoints for high-impact actions.
What Microsoft offers on the same problem
Microsoft's identity tooling follows similar principles. Microsoft's documentation describes Entra Agent ID, which creates agent identities as accounts within Microsoft Entra ID that provide unique identification and authentication for AI agents. Microsoft says every agent identity requires a human sponsor accountable for the agent's purpose, lifecycle decisions and access reviews. That maps closely to JumpCloud's "register" step.
Microsoft also says agent identities are governed through the same entitlement management access packages used for humans, providing time-bound, auditable access with approval workflows. A Microsoft blog adds that agents historically may have used shared credentials or access rights borrowed from an employee. That is the pattern the survey's authentication figures describe.
Rollout is a practical issue. Third-party analyst Big Hat Group says Entra Agent ID reached general availability in April 2026. Devoteam cautions that agents created before the Agent ID rollout still run on legacy service principals, and moving them over is not automatic. Microsoft's own materials describe parts of agent identity governance as preview, so check current licensing and feature status in your tenant before planning around it.
A practical checklist for admins
The survey supports these steps as sensible implications. No single control guarantees safety.
- Inventory agents and the systems they connect to, including ones built by business users.
- Assign an accountable owner to every agent.
- Compare actual grants against task needs. Replace shared service accounts and long-lived keys where you can.
- Test your logs. Pick an agent action and try to reconstruct the identity, permissions and result.
- Decide which actions need approval before execution rather than review afterward.
- Schedule permission reviews so they don't rely on someone remembering.
Bottom line
The strongest takeaway is the gap between confidence and evidence, not the platform comparison. Most IT leaders in this sample believe they can audit AI, but far fewer can reconstruct what an agent did. JumpCloud sells identity infrastructure, so its framing favors a cross-platform identity layer. The directional problem still looks real: agent adoption is moving faster than identity and audit controls. Microsoft 365 administrators have native tools to close some of that gap, provided they actually apply them.