
Voluntary gateway routing is not a shadow AI strategy because the employees who opt in are the ones who were already low-risk, while developers, power users, and teams running local agents keep working outside the gateway. Real shadow AI visibility requires active discovery that does not depend on employees, combining endpoint visibility, identity provider analysis, and network inspection.
Key Takeaways:
- Voluntary gateways self-select for low-risk users, missing the highest-risk AI activity
- Power users and developers building local agents are least likely to route through sanctioned tools
- True visibility requires active detection: endpoint monitoring, IdP analysis, and network inspection
- Local AI agents remain a detection challenge—no vendor has fully solved this problem
- Effective shadow AI governance cannot depend on employee self-reporting
Why do enterprises rely on voluntary AI gateways?
When enterprises first confront shadow AI, the common response is to stand up a sanctioned AI gateway, publish usage guidelines, and ask employees to route AI interactions through it. The logic seems sound, but it depends on employees choosing to comply, and the users who create the most exposure are the least likely to.
When enterprises first confront the reality of shadow AI, the most common response follows a predictable pattern: stand up a sanctioned AI gateway, publish usage guidelines, and ask employees to route their AI interactions through the approved channel.
The logic seems sound. Give people a secure, governed path to AI tools, and they’ll use it. Publish the policy, communicate the benefits, and adoption will follow.
But there’s a fundamental flaw in this approach—one that an enterprise security leader articulated directly during a recent evaluation: “If we have to ask people to route through the gateway voluntarily, we’ve already lost.”
He was right. And understanding why requires looking closely at who actually complies with voluntary policies—and who doesn’t.
Why does voluntary gateway routing miss high-risk AI use?
Voluntary compliance self-selects for low-risk users: the employees who read the policy and default toward compliance. Developers building local agents, power users with multiple AI subscriptions, and teams experimenting with open-source models are least likely to change workflows that already work for them, so the highest-risk activity stays outside the gateway.
The employees who voluntarily route through a sanctioned AI gateway are, almost by definition, the ones who were already low-risk. These are the users who read the acceptable use policy, attended the security awareness training, and default toward compliance when given the choice.
They’re not your shadow AI problem.
The users generating the most significant enterprise exposure—developers building local AI agents, power users maintaining multiple AI subscriptions, engineering teams experimenting with open-source models—are precisely the ones least likely to change their behavior based on a policy memo. Not because they’re malicious, but because they’ve already invested time building workflows that work for them. They’re moving fast. And in many cases, they’ve concluded that the sanctioned tools don’t meet their needs.
This creates a selection bias that fundamentally undermines the value of voluntary gateway routing. The gateway ends up capturing traffic from the employees who posed the least risk in the first place, while the highest-risk activity continues to flow outside your visibility entirely.
What is the shadow AI discovery gap?
The discovery gap is the difference between the AI usage a voluntary gateway sees and the usage that actually happens. Gateway dashboards show activity and policy metrics, so everything looks governed, but local coding assistants, personal subscriptions, and self-hosted models never appear. Audits and risk assessments built on that data undercount real exposure.
A gateway that only sees what employees choose to send through it produces something worse than no coverage at all—it produces a false sense of security.
Security leaders reviewing gateway logs see activity. They generate reports. Dashboards show usage patterns, token consumption, and policy enforcement metrics. Everything looks governed.
But those metrics only represent the fraction of AI usage that voluntarily surfaced. The developer running a local coding assistant, the analyst with a personal ChatGPT Plus subscription, the team experimenting with self-hosted models—none of that appears in your governance layer.
This is the discovery gap that voluntary approaches create. You end up with a system that’s highly effective at monitoring the traffic you didn’t need to worry about, and completely blind to the activity that actually matters.
For CISOs and security architects, this gap represents a significant risk posture problem. Audit reports based on gateway data will show compliance metrics that don’t reflect reality. Risk assessments will undercount exposure. And when a data exfiltration incident traces back to an unsanctioned AI tool, the gateway logs won’t help you understand what happened.
What does real shadow AI detection require?
Real shadow AI detection uses methods that work independently of employee behavior. Endpoint visibility shows the AI applications running on managed devices, identity provider analysis reveals sign-ins to AI services, and network inspection detects AI API traffic outside sanctioned channels. Combined, they give an accurate picture of AI usage whether or not employees report it.
Real shadow AI visibility requires detection methods that operate independently of employee behavior. This means building a discovery layer that doesn’t depend on users opting in.
Endpoint visibility provides direct insight into what applications and processes are running on managed devices. This captures local AI tools like Claude Desktop, Cursor, and other desktop applications that never touch a network gateway.
Identity provider (IdP) analysis reveals authentication patterns to AI services, surfacing SaaS AI tool usage that employees access with corporate credentials—or personal accounts linked to work email addresses.
Network inspection detects traffic patterns associated with AI API calls, even when users aren’t routing through sanctioned channels. This won’t capture everything, but it closes gaps that endpoint-only approaches miss.
Combined, these methods create an active discovery capability that produces an accurate picture of AI usage across the enterprise—regardless of whether employees chose to report it.
Can you detect local AI agents?
Only partially. Tools like Claude Desktop, Cursor, and locally hosted models can run entirely on the endpoint without calling known AI services, so full enforcement remains an unsolved problem across the industry. Active discovery improves coverage dramatically compared to voluntary approaches, and knowing the remaining gaps lets you add compensating controls.
Here’s where most vendors stop being truthful: full enforcement of local AI agents remains an unsolved problem, even with active detection methods in place.
Tools like Claude Desktop, Cursor, and locally-hosted models present genuine detection challenges. They may run entirely on the endpoint without making network calls to known AI services. They may use encrypted local storage. And in some configurations, they simply don’t generate the signals that traditional security tools are designed to capture.
Any vendor claiming complete visibility into all local AI agent activity is overstating their capabilities. The honest position acknowledges that active discovery dramatically improves coverage compared to voluntary approaches—while being transparent about the gaps that remain.
This matters because security leaders need accurate information to make risk decisions. Understanding that local agents represent a partially-addressed challenge allows you to implement compensating controls, adjust acceptable use policies, or make informed decisions about endpoint restrictions. Believing you have complete coverage when you don’t leads to blind spots in your risk posture.
How do you build a shadow AI strategy without self-reporting?
Start with active discovery that surfaces AI usage across the enterprise without employee self-reporting, then use that visibility to understand actual exposure, prioritize enforcement, and decide which risks to accept or mitigate. Airia provides active discovery methods that operate independently of employee behavior, and is transparent that local agents remain a challenge.
The path forward requires accepting that shadow AI governance cannot be built on voluntary compliance. The employees who need governance most are the ones who will comply least—not out of malice, but out of momentum.
Effective shadow AI strategy starts with active discovery: detection methods that surface AI usage across the enterprise without depending on employee self-reporting. This provides the accurate visibility needed to understand actual exposure, prioritize enforcement efforts, and make informed decisions about which risks to accept and which to mitigate.
Airia’s approach to AI governance is built on this principle. Rather than relying on users to route through sanctioned channels, the platform provides active discovery methods that operate independently of employee behavior—giving security teams visibility into AI usage patterns they would otherwise never see. When discovery finds local MCP client configurations on employee devices, it can replace them with Airia’s MCP Gateway, so those tool connections come under policy without anyone having to opt in.
That’s not a claim of perfect coverage. As noted, local agents remain a genuine challenge for the entire industry. But active discovery closes the most significant gaps that voluntary approaches leave wide open.
The enterprise security leader who said “we’ve already lost” if compliance is voluntary understood something important: governance that depends on the goodwill of the governed isn’t governance at all. It’s hope. And hope isn’t a shadow AI strategy.
What is AI FinOps? AI FinOps applies financial accountability to AI spend: who spent it, on what, and whether it stayed inside budget. Airia runs AI FinOps in the path of every AI request, so cost is attributed and budgets are enforced before the money is spent.
Ready to stop relying on employees to self-report their AI usage?Book a demo to see how Airia’s active discovery capabilities surface shadow AI across your enterprise—giving your security team real visibility instead of voluntary compliance metrics.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.