AI agents are becoming a governance problem, not just a productivity tool
As companies prepare to deploy AI agents across business workflows, security and governance are becoming central obstacles to scaling the technology safely.
An MIT Technology Review Insights article produced in association with the Deloitte Microsoft Technology Practice argues that agentic AI can open a new enterprise attack surface. The concern is that insecure agents may be manipulated into accessing sensitive systems, proprietary data, or tools beyond their intended role.
The article is sponsored content rather than MIT Technology Review editorial reporting, but it includes survey figures and a clear enterprise risk thesis. According to the Deloitte AI Institute 2026 State of AI report cited in the piece, nearly 74% of companies plan to deploy agentic AI within two years. Only 21% report having a mature model for governance of autonomous agents.
Non-human identities are multiplying
One of the article’s key points is that modern enterprises already manage a growing number of non-human identities, such as service accounts, machine credentials, automated workflows, and software actors. Agentic AI could accelerate that trend because agents may need permissions, data access, tool access, and the ability to act on behalf of users or business functions.
That creates a different risk profile from ordinary chatbot use. A conversational system that answers questions is one thing; an agent that can retrieve files, call internal tools, write to systems, or initiate actions is another. Governance has to define what the agent is allowed to do, whose authority it is using, and how its behavior is monitored.
The source article says executives are most concerned with data privacy and security, cited at 73%. Legal, intellectual property, and regulatory compliance follow at 50%, while governance capabilities and oversight are cited at 46%.
The control-plane concept is moving into AI operations
Andrew Rafla, a principal in Deloitte’s Cyber Practice, describes a control plane as a centralized layer that governs who can run which agents, with which permissions, under which policies, and using which models and tools. In his framing, without such a layer, companies have unmanaged execution rather than scalable autonomous operation.
That concept matters because enterprises rarely deploy technology in isolation. AI agents may interact with identity systems, document stores, customer records, code repositories, analytics platforms, and external services. If each deployment manages permissions and auditability differently, oversight becomes fragmented.
A functional governance system would need to answer basic operational questions: what an agent did, on whose behalf, using what data, under which policy, and whether the action can be reproduced or stopped. The article presents those questions as the minimum foundation for enterprise-scale agent use.
Governance separates pilots from production
The source argues that governance is what moves AI agents from experiments to repeatable enterprise automation. Pilot projects can often rely on close supervision, limited data, or manual guardrails. Production deployments need controls that work consistently across teams and use cases.
The risk is not only that an agent makes a single mistake. It is that a poorly governed agent system can fail unpredictably and at scale. If many agents have broad access, weak monitoring, or unclear accountability, small design flaws can become systemic exposure.
For businesses, the near-term implication is that agent deployment should be paired with identity, security, compliance, and observability planning. Treating governance as a later add-on may make early pilots easier, but it can leave organizations without the control structures needed for broader rollout.
What the article signals
The piece reflects a broader shift in enterprise AI discussion. The question is no longer only whether AI agents can automate useful work. It is whether organizations can define and enforce the boundaries under which those agents operate.
Because the source is sponsored content, its recommendations should be read in that context. Even so, the risk categories it identifies are concrete: privacy, security, legal compliance, intellectual property, oversight, permissions, and auditability. Those are likely to remain central as agentic AI moves from demonstrations into operational systems.
This article is based on reporting by MIT Technology Review. Read the original article.
Originally published on technologyreview.com
