AI reaches an enterprise through many doors. An employee opens a chat tab. A developer installs an assistant. A SaaS vendor switches on an AI feature by default. A team connects an MCP server to production data. Each door leaves a different trace, and no single tool watches all of them.
Airia Discover combines eight discovery signals across three tiers: Workforce, Infrastructure, and Application and Code. Together they map AI usage from the browser to the kernel to the repository, without requiring browser extensions or SASE.
Why one signal leaves gaps
A tool that watches one layer inherits that layer’s blind spots. Network-layer tools see AI traffic that crosses the perimeter, and they miss thick-client AI tools, embedded SaaS features, and agents built in code. Provider consoles show AI inside one cloud. Manual inventories go stale the day someone finishes them. The AI most likely to create risk is the AI that single-signal approaches cannot see.
That is why Airia consistently finds two to four times more AI than organizations expect. The gap does not come from carelessness. AI adoption moved faster than any inventory process built for traditional software, and each signal below closes a different part of it.
Tier one: Workforce Signals
Workforce signals detect the AI that employees use, from three angles.
Identity-level detection uses identity provider logs to link AI activity to individual users and roles. You learn who uses which AI tool and which team owns the risk.
Browser-level detection tracks web-based AI activity as it happens. Teams can block, warn, redirect, or audit it, which prevents unauthorized data exposure from a pasted document or an unapproved chat tool.
Device-level detection scans installed applications and local environments for AI running outside approved frameworks. Desktop AI clients, local models, and developer tools show up here. This is also where the new endpoint detection and response (EDR) integrations with CrowdStrike, Defender, and SentinelOne add endpoint evidence.
Tier two: Infrastructure Signals
Infrastructure signals watch the systems that AI runs on and talks through.
Network-level detection integrates with SASE and network controls to trace data flows between models and enterprise systems. It shows where information leaves and which model receives it.
API-level detection finds calls to model providers such as OpenAI, Anthropic, and Amazon Bedrock. An application that quietly sends prompts to an external model leaves an API trail even when no one registered it.
eBPF-level detection goes deeper, and it earns its own section.
Detect AI usage at the kernel level with eBPF
eBPF lets a small, sandboxed program run inside the Linux kernel and observe system activity as it happens: processes starting, files opening, network connections forming. It does this without modifying the applications it watches and without routing traffic through a proxy.
For AI discovery, that vantage point matters. A workload that calls a model API does not need a browser, an extension, or a network gateway. A containerized service, a background job, or a local agent can reach a model endpoint from places the other signals never see. The kernel sees the connection no matter which application opened it.
eBPF-level detection gives Discover evidence from the workloads themselves, which narrows the room for AI to run unseen on infrastructure you operate. It pairs with the network and API signals, so a kernel-level observation can be confirmed against what the other layers report.
Tier three: Application and Code Signals
The third tier looks inside applications and at the code teams ship.
App-level interrogation queries platforms such as Microsoft 365 and Google Workspace to uncover AI embedded in business-critical tools, including features a vendor enabled by default.
MCP server detection finds the MCP servers connected to enterprise systems, such as files, SQL databases, GitHub, and Slack. An unreviewed MCP server gives an agent a path into production data, so MCP visibility is a security question as much as an inventory question.
Code repository scanning searches repositories for LLM integrations, agents, and MCP servers before they reach production. Developers building against an unreviewed MCP server widen the gap every week, and the scan catches them while the fix is still cheap.
Why eight signals beat one
The signals corroborate each other. The same AI tool might appear in identity logs, on a device, in an API call, and in a repository. Discover matches those sightings into one record instead of four partial ones, and it attaches the context that makes the record useful: owner, access scope, data types accessed, and behavioral history.
Risk Classifications then attach a risk tier to each asset, so the inventory works as a prioritized list. Security sees which agents need review first. Governance sees which deployments fall under the EU AI Act, NIST AI RMF, or SR 11-7.
How the tiers work together: a hypothetical
Picture a developer who builds a support agent over a weekend. The agent calls an external model API, connects to an MCP server that reads a customer database, and runs in a container on a shared server. No one registered it.
Code repository scanning sees the LLM integration and the MCP configuration in the repository. API-level detection sees the calls to the model provider. eBPF-level detection sees the connection form from the container, with no browser or gateway involved. Identity-level detection ties the activity to the developer’s account and team. Four signals, one agent, one record with an owner, an access scope, and a risk tier. A single-signal tool would have caught one trace of it, or none.
Teams deciding which signals to switch on first can read Not Every AI Discovery Method Deserves Equal Priority. This post covers the full map. That one covers the order of operations.
A full inventory in 24 to 48 hours
Discover returns a complete first inventory within 24 to 48 hours of deployment and updates it continuously as AI is deployed, modified, or retired. New deployments, configuration changes, and permission changes appear as they happen. Every decision rests on today’s estate instead of last quarter’s audit.
Because Discover is a capability of the Airia platform, every deployment it finds enters scope for Secure, Govern, and Optimize immediately. The AI inventory becomes the foundation that enforcement, policy, and spend tracking all work from, with nothing to export and nothing to reconcile between tools.
Discover the AI you have not found yet
You cannot govern AI you have not found. See how eight signals map your workforce, infrastructure, and code. Request a demo of Airia Discover and review your first inventory within 24 to 48 hours.
Put these ideas to work.
Schedule a 30-minute walkthrough with our team.