Attestation
The Access-Proof Gap: Why 72% of Firms Can't Document What Their Agents Did
New research from JumpCloud, covered by Compare the Cloud, puts a number on a problem the agent-protocol world has been circling for months: a gap between what AI agents can reach and what organizations can actually prove about that access. More than half of organizations now run agents capable of changing systems, permissions, or files, but 72% report they can't verify agent access the way they verify human access — and 59% haven't folded agent activity into existing identity governance at all.
JumpCloud's Joel Rennich frames the shift plainly: "AI agents are quickly becoming another class of identity in the enterprise. They can access sensitive systems, change records, trigger workflows, and act on behalf of people, but many organizations still aren't governing them with the same rigor they apply to human users".
This is a distinct failure mode from the ones already circulating in agent-accountability discourse. It isn't about audit logs naming the wrong actor, or about agents going rogue and getting quarantined after the fact. It's upstream of both: companies are deploying agents with real write access to production systems — changing records, triggering workflows — without the instrumentation to answer a basic question afterward: what did this thing actually touch?
The gap matters because access and proof-of-access have historically been bundled for human identity. When an employee logs into a system, the same IAM stack that grants the permission also generates the log entry. Agentic systems break that bundling in practice, not in principle — nothing stops an org from logging agent actions with the same rigor, but most haven't built the pipeline. An agent might call five tools in a single reasoning chain, several of which are MCP servers or third-party APIs outside the identity provider's visibility, and the resulting audit trail ends up sparse, delayed, or simply absent.
What's notable here is that this is a survey of lived enterprise experience, not a vendor's hypothetical threat model. 59% of surveyed organizations have agents live in production — reshaping records and permissions — while treating them as outside the perimeter of identity governance that every human account is subject to. That's not a future risk. It's a present operational blind spot, and it's one more data point suggesting that the accountability layer for agentic AI is lagging well behind the deployment layer, across a wide swath of ordinary enterprises rather than a handful of edge cases.
The practical takeaway: before organizations debate which attestation standard or audit-trail format wins, many still need to close the more basic gap — building the logging pipeline that lets them answer, after the fact, what an agent did and why.