As AI moves from generating answers to performing work, governance has to move from policy alone to permissions, decision rights, approval gates, monitoring, evidence, and enforceable operating boundaries.
By Renee Cannon, Founder of eunoiaAI
September 28, 2026
![[Hero Alt Text]](https://cdn.prod.website-files.com/6a5e755fad03f8f3b4db484b/6aba950c01cdce19f94ba7d5_Golden-Hour%20Governance%20Desk_ChatGPT%20Image%20Sep%2028%2C%202026%2C%2011_56_42%20AM.png)
For years, AI governance could reasonably begin with a document.
Define acceptable use. Establish principles. Identify prohibited activities. Assign responsibility. Train employees. Review periodically.
That work still matters.
But something important has changed.
In September, OpenAI introduced GPT-6 Astra, describing a model designed for complex, end-to-end work across reasoning, research, coding, computer use, and document creation. In ChatGPT Work, AI can carry work across files, applications, connected tools, and browser-based tasks rather than simply returning an answer for a person to act on.
OpenAI is not alone. Microsoft, Google, Anthropic, and others are developing systems that can interact with software, use tools, and participate more directly in workflows.
That is exciting.
It should also make business leaders slightly uncomfortable.
Because once AI can act, a Responsible AI policy that says what the organization believes is no longer enough by itself.
The business also has to decide what the AI is actually allowed to do.
Imagine an organization has a clear rule:
Customer-facing communications require human approval.
Now imagine an AI system has access to customer information, email, the CRM, and the tools required to prepare and send a response.
The policy says a human must approve it.
But what actually prevents the system from sending it?
That question is becoming much more important.
A written rule describes the organization's intent. An operating control determines what happens when software is doing the work.
Those controls may include permissions, authentication, tool access, approval gates, workflow boundaries, escalation paths, monitoring, and evidence.
The gap between those two things—what the organization says AI may do and what its systems technically allow AI to do—is becoming a governance issue of its own.
Once an AI system can participate directly in work, seemingly simple implementation decisions start carrying business consequences.
Those are not merely technical configuration questions.
They are decisions about authority.
There is a comforting phrase in AI governance: human in the loop.
But where is the human?
What are they deciding?
What information do they receive?
Can they meaningfully challenge the system?
What happens if they do nothing?
And does the workflow actually stop without their approval?
These distinctions matter because even the technology providers are confronting them.
OpenAI says Astra includes monitoring intended to identify potential cases where an agent may have misunderstood instructions and can pause work for human review. Microsoft cautions that computer-use human-supervision requests are probabilistic and should not be treated as a guaranteed fail-safe. Google advises close supervision for important computer-use tasks and warns against using the capability for critical decisions or serious, irreversible actions.
The lesson is not that these systems are unsafe and should not be used.
It is almost the opposite.
They are becoming useful enough that organizations need to take their operating boundaries seriously.
A business may decide an AI system can research a customer account but cannot edit it.
That distinction can exist in permissions.
It may decide AI can prepare an outbound message but a person must release it.
That can become an approval gate.
It may allow an agent to complete routine actions under defined conditions but require exceptions to route to a named owner.
That can become workflow logic.
Not every governance principle can—or should—be converted into software.
But wherever practical, the operating environment should reflect the authority the business believes it granted.
Otherwise, the organization may discover that its AI policy and its AI operations describe two different companies.
This does not make governance frameworks, Responsible AI policies, risk assessments, or training obsolete. They establish essential organizational direction.
But there is a difference between defining a principle and operationalizing it.
A policy can say that humans retain accountability for consequential decisions.
The next question is:
Which decisions are those, exactly?
And then:
Where does the workflow stop?
Who approves?
What evidence reaches them?
What access does the AI have before and after that decision?
What happens when the expected path fails?
Those questions may require governance professionals, business owners, security teams, technology teams, or implementation partners to get considerably closer to the actual work.
Sometimes governance needs to get into the weeds.
Not because every organization needs a complicated control environment, but because a beautifully written principle cannot enforce a permission.
There is a tempting reaction to launches like Astra:
Look what AI can do now.
It is the right reaction.
But it should be followed immediately by another:
Which of those things should it do here?
Not every workflow needs an agent. Not every agent needs greater autonomy. Not every organization needs the same controls. And greater autonomy is not automatically evidence of greater AI maturity.
The appropriate level of authority depends on the work, the consequence of error, reversibility, available evidence, human judgment, and the organization's ability to operate what it deploys.
That is why AI governance is changing.
The policy still matters.
But when AI can act, governance has to govern action.
Prefer to listen? Hear this Insight explored in an AI-generated discussion from eunoiaAI.
Reference links
Explore practical governance around real workflows, decision rights, authority, and operating controls.
Passing an evaluation supports a decision to launch. Governed operation keeps that decision accountable as workflows, permissions, evidence and conditions change.
An agent's task tells you what it does. Its decision rights tell you how much authority the business has actually handed over.