You can’t govern what you can’t see. Here’s why ephemeral discovery, a centralized AI Hub, and browser plugins outperform persistent endpoint agents for enterprise AI governance.

Before an organization can govern AI usage, it must first answer a fundamental question:  

What AI tools and services are being used?  

The proliferation of AI-enabled SaaS applications, browser-based AI assistants, API-connected copilots, and shadow AI tools has made this question increasingly difficult to answer through traditional IT asset management and traditional security toolsets.  Traditional security tools were built for a world of devices, users, endpoints, and human identities.  All constructs AI has rapidly redefined. 

Answering the question (and converting the answer into real governance) means resolving three foundational challenges:  

1. How do I accurately discover and inventory which AI tools and services are being used?   

2. How can I obtain visibility and governance across AI workflows?  

3. How are new workflows onboarded, secured and ultimately trusted? 

This whitepaper addresses these challenges and critical design decisions for successful AI deployments. 

Day One Discovery 

Continuous discovery should be done in the most comprehensive and frictionless way possible. The resulting inventory becomes the foundation of your AI Manifest, which is a living record of every AI tool, identity, service, and workflow in use across the organization. 

Three primary options exist for performing this initial discovery:   

1. Ephemeral Discovery 

2. Standalone Persistent Agent  

3. Existing Toolsets 

Ephemeral Discovery 

A non-persistent method that deploys lightweight software on demand, performs a targeted set of discovery actions, and automatically removes itself upon task completion. Collected data is transmitted in encrypted form to a centralized repository for storage, search, and follow-up. No persistent credentials are stored at the endpoint. 

This asymmetry matters particularly for AI discovery, where the landscape of services, models, and interaction patterns is evolving faster than other methods can adopt.  Since Ephemeral Discovery can be updated in near real-time and deployed at the next scheduled discovery window, the pace of change is minimized.  JetStream provides this capability natively within the platform; AI resources discovered via this method are automatically cataloged into the AI Manifest. 

This stands in contrast to a persistent standalone agent: software that is installed on a system, registered as a service or daemon, and remains resident across reboots. Persistent agents carry an ongoing cost. Every second one runs; it is a target, a maintenance obligation, and a compatibility risk that compounds with every OS patch, browser update, and tool deployment. 
The ephemeral approach eliminates the cost. When the discovery task concludes, the agent dissolves. No process remains in memory. No service persists across reboots. No privileged credentials sit waiting to be harvested. The attack surface does not grow. 

Persistent Standalone Agent 

A persistent agent is software that is installed on a host system and remains resident across power cycles and operating system reboots. It operates continuously or on a scheduled basis, typically as a registered service, daemon, or background process deeply integrated with the host operating system.  Persistent agents typically require deep integration with the host operating system to function effectively across their continuous lifecycle.  

This integration commonly includes: 

• Registration as a system service (Windows Service, Linux systemd unit, macOS LaunchDaemon) 

• Access to kernel-level APIs for real-time monitoring and interception 

• Hooks into OS event subsystems (file system filters, network stack callbacks, process creation events) 

• Registry or configuration store entries that persist across reboots 

• Elevated or SYSTEM-level privileges to operate beneath user session boundaries 

This depth of integration can create interrogability and operational challenges.   

Finally, since persistent agents run continuously, they impose a tax on the host system, consuming CPU cycles, memory, disk I/O, and network bandwidth.  In high-volume enterprise environments, the cumulative resource footprint of multiple co-resident persistent agents is not merely a nuisance; it creates measurable performance degradation that directly impacts end-user productivity and system reliability.  

Existing Tooling  

Organizations that have invested in IT and security platforms such as EDR (CrowdStrike, SentinelOne, Microsoft Defender, or similar).  These tools commonly offer application inventories and can be used to search for AI-related services by keyword or file hash.  However, these approaches require additional effort.  

Teams must build and continuously maintain search criteria to keep pace with the rapidly changing AI landscape. More critically, discovering certain AI components, such as Model Context Protocol (MCP) configurations, requires inspecting file content (e.g., searching JSON for pointers and references), which is generally not possible with EDR solutions. This leaves gaps that ephemeral collection is designed to close.  With this as a backdrop, while using existing tooling is viable, you must determine if the work is justified.  You may determine that this approach is best for your goals. 

Read more: Why Enterprise AI Governance Doesn’t Need Another Endpoint Agent 

From Discovery to Governance 

Once you have moved past initial discovery and have developed a maintained AI manifest, we must begin thinking about how to achieve AI Governance.  This is where a centralized AI Hub Control Plane and Browser Plug-ins play a critical role in maximizing viability and efficiencies while removing friction.     

AI Hub Control Plane

Once discovery has been performed and an AI manifest has been created, an AI Hub model should be deployed to monitor, control, and enforce AI policies across the enterprise.  Specifically, the AI Hub serves as a proxy for all AI services.  By routing through this control plane, organizations gain visibility, monitoring, alerting, and enforcement.  Several added benefits include financial tracking and enforcement, key management, and drift identification. To ensure policy enforcement with users, a browser-based plugin approach is recommended.  This is needed to ensure users are not exposing sensitive information to web-based AI chatbots, etc.   

A proxy deployed as the routing layer for AI traffic delivers governance capabilities that a persistent endpoint agent cannot replicate at scale. This approach also removes the friction caused by a persistent agent approach.   

This approach enables organization to: 

• Monitor and enforce AI policies at scale 

• Manage and rotate API keys to AI services centrally rather than distributing them to individual endpoints 

• Audit every AI interaction with request and response metadata 

• Alerting in real-time on anomalous AI usage patterns across the entire organization simultaneously 

• Detect drift across AI usage 

• Benchmark and control costs  

The persistent agent approach to AI governance requires successful deployment to every managed device, meaning any unmanaged device, contractor machine, mobile endpoint, or BYOD asset falls outside governance coverage entirely. The proxy approach covers all these scenarios.  Any device that routes through the AI Hub is subject to the same policy enforcement regardless of what software is installed on it. Think of it as the equivalent of a network control point. The same way VPNs govern network traffic; the AI Hub governs AI traffic. 

Persistent agents also create a moving target for maintenance: every OS update, driver change, or security patch is a potential compatibility event that may require agent remediation. The proxy has no per-endpoint dependency and requires no coordination with endpoint management cycles.  

Browser Plugins 

Browser Plugins provide a means to gain visibility and enforcement for AI actions performed via a user’s web browser.  Users interact with AI services through web browsers, browser extensions, web applications, and API calls that are invisible to conventional endpoint inventory tools. The result is a significant governance blind spot: organizations cannot enforce policy over AI components they cannot see.  By using a Browser Plugin, visibility and enforcement can be instrumented.  This includes redacting sensitive information from end user prompts to chatbots, monitoring and alerting violations, etc.   

Comparative Summary 

The following table summarizes the key differences across the dimensions examined in this paper. 

AI You Can Trust (and actually deploy)  

AI governance does not require a persistent agent on every endpoint. By combining ephemeral discovery with a proxy control plane and browser-based enforcement, organizations can achieve comprehensive visibility, policy enforcement, and financial control across their AI landscape.  This approach does not introduce an expanded attack surface, and maintenance burden. 

JetStream is using this approach: a control plane for AI that delivers real-time governance.  The result is AI you can trust and deploy as a true competitive advantage.