The Admin Console Shows What Your Agent Did. Not What It Promised.
OpenAI made agent activity visible in May.
Business, Enterprise, and Edu admins can now open the admin console and see which workspace agents exist across their organization, what they’ve done, what they’ve connected to. The feature went GA on May 22. The free period runs through July 6. After that, credits.
It’s a real feature. It matters. And here’s what it doesn’t show.
The thing missing from every admin console
The admin console logs the action. Not the commitment inside it.
An agent sends a client email. The log records: sent, timestamp, recipient, sender. Policy governance: authorized, within scope, compliant. The audit trail is complete.
What the audit trail doesn’t capture: the email contained a delivery commitment. “We’ll have this to you by June 30.” That commitment is now real. It has a counterparty. It has a deadline. It belongs to someone in your organization.
The admin console doesn’t know it exists.
Three weeks later: the June 30 deadline approaches. The agent that made the commitment is running other tasks. The human it works for is managing 47 other threads. The commitment exists somewhere — in an email thread, in a client’s expectation, in a vendor’s planning calendar — and nowhere in your system.
That’s not a visibility failure. The action was visible. That’s a commitment tracking failure.
What Gartner is measuring
Gartner published in May: 40% of agentic AI projects will be canceled by 2027 due to governance gaps found only after production incidents.
The same week: Agentic AI Institute put numbers on it. 72% of enterprises have agents in production. 60% have no formal governance.
Coverage treated these as the same problem. Build governance, close the gap, avoid the cancellations.
They’re not the same problem.
The 60% without governance — those failures will look like access control incidents. Unauthorized actions. Out-of-scope behavior. The governance layer catches them. They’ll build governance and fix that.
The 40% Gartner predicts aren’t coming from there. They’re coming from the 72%. The organizations that built the governance. That have the admin console. That can see what their agents are doing.
Their failure mode: perfect action logging, zero commitment tracking. Agents running authorized workflows, inside policy bounds, logging every action — while making commitments nobody is tracking and nobody can account for.
The distinction the industry keeps missing
Access governance: who authorized this agent, what can it touch, what policies govern it, what did it do.
Commitment governance: what did the agent commit to during those actions, to whom, by when, and was it fulfilled.
Every major platform is building the first. OpenAI’s admin console, Microsoft’s Agent 365, Google’s compliance tooling — all of it answers the access question with increasing precision.
Nobody is building the second.
In a world where agents are making external commitments at machine speed — sending proposals, quoting terms, booking resources, scheduling follow-ups — the question that surfaces in every production incident isn’t “was the agent authorized?”
It’s “what did it commit to — and where is that tracked?”
The audit trail shows the action. The action carried a commitment. The commitment has a deadline and a counterparty. None of that is in the log.
Gartner called the failure mode correctly: governance gaps found only after production incidents. The governance gap they didn’t name is the one that will cause most of them.
Eliran Keren — Founder of Deeplica, building coordination infrastructure for the agent era.
Sources: OpenAI — Workspace Agents GA, May 2026 · Gartner — Uniform AI Agent Governance Will Lead to Failure · Gartner — 40% of Agentic Projects Canceled by 2027 · Agentic AI Institute — 72% Production, 60% Governance Gap
