AI agents are becoming enterprise insiders. Is your business ready to manage them?

For enterprises, that creates opportunities to automate complex work and improve operational efficiency. It also introduces a new challenge: how do you maintain control when software can make decisions and act across business systems?

That question has become increasingly urgent.

Microsoft recently described AI agents as a new category of enterprise insider, highlighting the need for dedicated identities, restricted permissions and continuous supervision. Days later, ServiceNow published research suggesting that many organizations are already struggling to manage the security risks associated with agentic AI.

The message for enterprise leaders is clear. As AI agents become more capable, the systems and teams responsible for managing them need to evolve.

The enterprise AI security gap is already emerging

According to ServiceNow's October 2026 research, based on a survey of more than 3,700 security leaders, 90% of organizations reported experiencing an agentic AI security incident in the previous year.

Perhaps more concerning, only 13% met the study's criteria for having a defensible agentic AI security posture.

The research also found that 29% of security teams had discovered AI agents operating in production without being registered or previously known to them.

These findings point to a growing visibility problem.

Organizations are introducing AI agents through enterprise platforms, development environments, third-party applications and individual teams. As adoption accelerates, it becomes increasingly difficult to maintain a complete picture of which agents are operating, what they can access and who is responsible for their actions.

An AI agent that summarizes internal documents presents a different level of risk from one that can modify customer records, approve transactions or interact with financial systems.

Yet both may be described simply as AI tools. Understanding that distinction is essential to managing enterprise AI safely.

Why AI agents need a different approach to security

Traditional enterprise security models are built around people, applications and clearly defined access permissions.

Agentic AI introduces another layer.

An agent may interpret instructions, select tools, retrieve information and execute several actions to complete a task. Its behavior can depend on the information it encounters along the way, including content from external or potentially untrusted sources.

Consider an AI agent supporting accounts payable. It might retrieve supplier invoices, compare them with purchase orders, identify discrepancies and prepare transactions for approval.

To do that effectively, it needs access to financial records, supplier information and relevant business applications. But what happens if the agent encounters incorrect information, receives a manipulated instruction or attempts an action outside its intended responsibilities?

The answer depends on how the system has been designed.

A production-ready agent needs clearly defined boundaries around the information it can access, the actions it can perform and the circumstances that require human approval.

Microsoft's October 2026 security guidance emphasizes this principle, recommending distinct agent identities, least-privilege access and permissions that can be limited to specific tasks.

For businesses, the implication is significant: agent governance needs to be part of the technical architecture from the beginning, rather than added after deployment.

Five questions enterprises should answer before deploying AI agents

Organizations don't need to eliminate every possible risk before putting AI into production. They do need a practical framework for understanding and managing the risks associated with each use case.

1. Do you know which AI agents are operating across your business?

Before an organization can govern its AI agents, it needs visibility into where they're being used.

That includes agents built internally, introduced through enterprise software, deployed by third-party providers or created by individual development teams.

A useful starting point is an inventory covering each agent's purpose, business owner, connected systems, access permissions and operational status.

Without that visibility, businesses risk creating an expanding collection of AI-powered processes that nobody fully understands or oversees.

2. Can each agent access only what it needs?

An agent that prepares customer briefings shouldn't automatically have permission to modify commercial agreements.

Similarly, an agent supporting invoice processing may need to review transactions without having unrestricted authority to approve payments.

Applying least-privilege principles helps limit the consequences of mistakes, unexpected behavior or malicious instructions.

Permissions should reflect the specific task an agent is performing, with additional authorization required for sensitive actions.

3. Where does human approval remain essential?

Not every AI-assisted action carries the same level of risk.

Generating a draft report may require relatively little oversight. Approving a significant financial transaction, changing customer entitlements or modifying production infrastructure demands stronger controls.

Enterprises should establish clear thresholds for when an agent can act independently, when it must request approval and when an action should be prohibited entirely.

These decisions should reflect business risk, regulatory obligations and the potential consequences of an error.

4. Can you explain what an agent has done?

When an automated process produces an unexpected result, teams need to understand what happened.

That requires more than recording the final output.

Organizations should be able to trace relevant inputs, tool interactions, decisions, approvals and actions across the workflow, with appropriate protections for sensitive information.

Effective monitoring also makes it easier to identify recurring failures, improve agent performance and demonstrate accountability.

5. Who is responsible when something goes wrong?

AI agents may operate across several business functions and technical systems.

That can make accountability unclear.

Business leaders need ownership of the outcomes agents are intended to deliver. Technology teams need responsibility for architecture, integration and reliability. Security and governance teams need visibility into the controls protecting those systems.

Clear ownership helps organizations respond more effectively when an agent behaves unexpectedly and ensures that ongoing performance remains somebody's responsibility.

Forward Deployed Engineers can help bridge the gap between AI capability and enterprise reality

Building a convincing AI demonstration is relatively straightforward compared with deploying an agent inside a complex enterprise environment.

Production systems introduce challenges involving legacy applications, fragmented data, authentication, business rules, operational dependencies and security requirements.

This is where Forward Deployed Engineers (FDEs) can play an important role.

FDEs work directly with business stakeholders and technical teams to understand operational challenges and build AI solutions around real workflows.

Their responsibilities can span solution architecture, application development, integration, testing and deployment.

For example, an FDE working on an AI-powered customer service workflow might need to connect an agent to a CRM platform, establish appropriate data access, implement escalation rules and test how the system responds to incomplete or conflicting information.

The goal is to develop a solution that works reliably within the customer's existing environment. That includes designing safeguards into the workflow, validating performance against realistic scenarios and ensuring internal teams understand how the solution operates.

Through its specialist Forward Deployed Engineering capabilities, Tenth Revolution Group helps organizations bring together practical AI engineering and business requirements to move solutions toward production.

This embedded approach can also support knowledge transfer, helping businesses develop the internal capability to maintain and improve their AI systems over time.

Governance also depends on having the right people and skills

Technical safeguards are essential, but they represent only part of the challenge.

As AI agents become more integrated into business operations, organizations need people who understand how to design, manage and work alongside them.

That includes several different capabilities:

• Engineering and architecture teams need expertise in AI integration, identity management, system reliability, testing and observability
• Business teams need practical training on what agents can do, where their limitations lie and when human intervention is necessary
• Technology leaders need the ability to evaluate AI opportunities against business objectives, implementation costs and operational risk
• Security and governance professionals need to understand how autonomous systems introduce risks that differ from conventional software applications

For many enterprises, developing all these capabilities internally will take time.

A combination of specialist hiring, targeted workforce training and external delivery expertise can help close immediate gaps while building longer-term capability.

Tenth Revolution Group supports organizations through its three connected service areas:

• Talent: Access specialist technology professionals with the engineering, AI, data and security skills needed to support enterprise delivery
• Training: Develop the practical capabilities employees need to adopt, manage and work effectively with emerging technologies
• Transformation: Bring together advisory, implementation and specialist delivery expertise to help organizations design and deploy technology around business outcomes

The right combination will depend on an organization's existing technical maturity, available resources and AI ambitions.

Building AI governance into delivery from day one

One of the risks facing enterprises is treating AI governance as a separate activity that begins after implementation.

By that stage, agents may already be connected to critical systems, operating with excessive permissions or creating dependencies that are difficult to change.

A more effective approach is to incorporate governance throughout the delivery process:

• During assessment, organizations should identify the business problem, understand the systems involved and evaluate the potential risks
• During design, teams should establish access controls, approval requirements, monitoring arrangements and accountability
• During development and testing, engineers should validate how agents behave under realistic operating conditions, including unexpected inputs, system failures and attempted misuse
• During deployment, organizations need appropriate monitoring, incident response procedures and the ability to restrict or disable an agent when necessary

And after deployment, teams should continue reviewing performance, permissions, costs and business outcomes as requirements evolve.

This approach also helps organizations avoid deploying AI agents simply because the technology is available. A clearly defined use case, supported by appropriate technical controls and measurable objectives, provides a stronger foundation for long-term value.

The next phase of enterprise AI requires greater accountability

AI agents are moving beyond isolated productivity tools and into the operational processes that businesses depend on.

That transition creates significant opportunities, but it also raises expectations around reliability, security and accountability.

The latest developments from Microsoft, ServiceNow and OpenAI demonstrate how much attention is now being directed toward controlling agent behavior, managing permissions and strengthening the environments in which AI systems operate.

For enterprise leaders, the priority should be to understand where agents can deliver meaningful value, establish appropriate safeguards and build the technical and workforce capabilities required to support them.

Forward Deployed Engineers, specialist delivery teams and targeted training can all contribute to that process, helping businesses connect AI innovation with the realities of enterprise operations.