Enterprises are still working to bring Shadow AI under control, but the rapid rise of autonomous AI agents is creating a more difficult governance challenge. As agents gain access to corporate data, applications and critical workflows, organisations need to know who owns them, what they can access and how far their authority extends.
Shadow AI was only the beginning
For the past two years, much of the enterprise AI governance debate has centred on Shadow AI – employees using unapproved generative AI tools beyond formal oversight. It is now being joined by something harder to see and contain.
The same permission granted to an autonomous agent could also allow that information to influence actions elsewhere.
AI agents are moving into business workflows, built by central technology teams, software vendors, business units and increasingly employees themselves. An agent may remain active after the initial instruction, connect to several systems and carry out tasks.
The governance question is therefore changing. Enterprises no longer need visibility only into which AI tools people are using. They increasingly need visibility into which AI entities are operating across the organisation.
When AI stops answering and starts acting
Depending on its design and permissions, an AI agent can retrieve information, call APIs, invoke tools, update records, send communications or trigger other processes.
Microsoft warns that autonomous agentic systems can plan, execute and adapt actions towards goals with limited human intervention. Each connection between an agent, tool and service can expand the attack surface, creating risks ranging from unintended actions and data leakage to agent hijacking.
An overly broad permission granted to a chatbot may expose information. The same permission granted to an autonomous agent could also allow that information to influence actions elsewhere.
The enterprise is heading towards agent sprawl
The scale could become striking. By 2028, an average global Fortune 500 enterprise could have more than 150,000 AI agents in use, compared with fewer than 15 in 2025. Yet only 13-percent of organisations believe they have the right governance in place to manage AI agents effectively.
The visibility gap is already becoming apparent. Recent enterprise research found that 82-percent of organisations have unknown AI agents operating within their environments, while 65-percent reported an AI agent-related incident during the previous 12 months.
But agent sprawl is not simply a numbers problem. Agents can emerge across different platforms, departments and vendor ecosystems, each with its own permissions, connections and level of autonomy. Without a central inventory and clear ownership, enterprises risk losing track of what is operating inside their environments, what those agents can access and, more importantly, what they are capable of doing.
Every agent needs an identity and an owner
The first requirement is surprisingly basic – every agent should be identifiable.
Enterprises need to know who created it, who owns it, what purpose it serves, which data and applications it can reach, what credentials it uses and when those permissions were last reviewed. An agent should not inherit broad access simply because the employee or service account behind it has that access.
Unique identities make permissions and actions traceable. Ownership provides the other half of the equation. When an agent behaves unexpectedly, changes role or is no longer needed, there must be a person or team accountable for reviewing, restricting or retiring it.
That lifecycle discipline matters. The CSA survey found only 21-percent of respondents had formal processes for decommissioning AI agents.
Not every agent carries the same risk
A research agent with read-only access to approved documents is not equivalent to an agent that can modify infrastructure, approve transactions or communicate externally.
AI agents can operate at different levels of autonomy, from simply observing and providing recommendations to taking actions with human approval or operating independently within defined guardrails. As autonomy increases, so does the potential impact of an agent’s decisions and actions. Governance controls should become stronger accordingly.
Uniform controls can create problems at both ends. Excessive restrictions can make low-risk agents unnecessarily difficult to use. Weak controls around highly autonomous agents can increase operational, security and compliance exposure.
What enterprises need to put in place now
AI agent governance needs to begin before inventories become difficult to control. Enterprises need a central agent register, unique identities, clear ownership and least-privilege access based on each agent’s role.
Governance should cover the entire lifecycle, from approval and monitoring to permission reviews and decommissioning, with audit trails showing what agents accessed and what actions they took. As agents gain greater autonomy and carry more risk, stronger controls, continuous monitoring and mechanisms to stop or reverse their actions become increasingly important.
The next governance problem is therefore not whether enterprises will use AI agents. That shift is already underway. The more urgent question is much simpler – do we actually know how many AI agents are operating inside the business and what they are authorised to do?
