Finding the AI Your Organization Already Uses
In the last week of January, a reasoning model from a Chinese lab went from little-known to the top of the mobile app charts. Within days, security leaders were fielding the same question from their executives: are our people using it, and what have they put into it?
For most organizations the honest answer was “we’re not sure”. That’s not a criticism of anyone. AI tools arrive through browsers, mobile apps, browser extensions, features switched on inside existing SaaS products, and developer libraries. None of those paths goes through procurement on its own.
The episode is a useful prompt. An organization that can answer the question quickly this time will be able to answer it for whatever arrives next month.
Where AI usage actually shows up
You don’t need a new tool to get a first picture. Most of the evidence is already sitting in systems you run.
Network and secure web gateway logs. Traffic to AI service domains, broken down by department and volume, is the fastest starting point. It shows reach, though not what was shared.
Identity provider logs. Users signing in to AI services with corporate credentials, or granting OAuth consent to AI applications, show up here. OAuth grants are worth particular attention because they can give a third-party service ongoing access to mail, files or calendars.
SaaS admin consoles. Many products the organization already licenses have added AI assistants over the past year, sometimes enabled by default. Check the admin settings for each major platform and note which AI features are on, who can use them, and where the data goes.
Endpoint and browser inventories. Desktop AI applications and browser extensions with AI features appear in software inventory. Extensions deserve scrutiny, since some request permission to read every page a user visits.
Code repositories and package manifests. Development teams pull AI SDKs and model libraries into applications. A search of dependency files will show which projects call external models, and from where.
Expense reports. Individual subscriptions paid on corporate cards are a reliable signal that a team found something useful and didn’t wait for approval.
Turn the findings into an inventory
A list of domains doesn’t help anyone make a decision. For each AI service or feature you find, record:
- who uses it, and roughly how much
- what kind of data could reach it, based on the users’ roles
- whether it’s covered by an enterprise agreement with data-handling terms, or used under consumer terms
- whether the provider retains or trains on submitted data under those terms
- who in the business would own a decision about it
That last field matters most. Shadow AI usually exists because people found a way to work faster. The goal is to bring that usage into the open and give it a home, with an owner who can weigh the benefit against the risk.
Policy follows the inventory
Organizations that wrote an AI acceptable use policy in 2023 often find it no longer fits. The tools have changed, and blanket prohibitions tend to push usage onto personal devices where it can’t be seen at all.
A policy written after the inventory can be specific. It can name approved tools for common tasks, explain what kinds of data can go into them, and give people a quick way to request something new. Clear, workable options reduce shadow usage more effectively than restrictions alone.
This also matters for regulatory reasons. The EU AI Act’s first obligations, including the expectation that organizations ensure an appropriate level of AI literacy among staff who work with AI systems, began applying on February 2. Organizations with EU operations or customers should expect questions about how they know which AI systems their people are using.
Keep it current
The inventory is only useful if it stays current. Schedule a monthly review of the same sources, and add AI usage to the questions asked in vendor reviews and software requests. When the next model tops the charts, the question from leadership becomes a five-minute lookup.
That’s the first step in building AI security maturity: knowing what’s in use before deciding what to protect, permit or restrict. Each later step depends on it.