Most mid-market IT managers discover their shadow AI problem the same way: someone in finance mentions they’ve been using an AI tool to summarize contracts for six months, and you’re hearing about it for the first time. That’s not a people problem. It’s a process problem — specifically, the absence of any systematic approach to finding out what AI tools are actually running inside your organization.
on Unsplash
Shadow AI discovery is the foundational step in any real AI governance program. You can’t assess risk on tools you don’t know exist. You can’t assign ownership to tools nobody has claimed. And you can’t demonstrate compliance with emerging AI regulations — including the EU AI Act’s requirements for organizations with EU exposure — if your tool inventory is a blank page. This guide covers how to do the work without a dedicated GRC platform or a compliance team.
Start With What You Already Have Access To
Before you run any scans or send any surveys, pull the data sources you already control. Your SaaS spend management tool, your SSO provider, and your expense reports are collectively going to surface more AI tools than any single discovery method.
Start with your SSO logs. If your organization uses Okta, Azure AD, or a similar identity provider, look at what applications have active OAuth connections. Filter for anything that was added in the last 18 months and cross-reference against your approved software list. Anything that doesn’t appear on that list is a candidate for your shadow AI inventory, AI-powered or not — but right now you’re looking specifically for AI tools, so flag anything with a known AI component.
Next, pull three to six months of corporate card and expense data. Look for SaaS subscriptions in the $10–$150/month range that aren’t running through IT procurement. Tools like ChatGPT Plus, Jasper, Copy.ai, Perplexity, Otter.ai, and similar services are almost always expensed directly by individual contributors or department heads. This is one of the fastest ways to find AI tools that never touched your procurement process.
Finally, if you manage or have visibility into browser extension deployments, audit those too. AI writing assistants, meeting summarizers, and research tools frequently enter organizations as browser extensions installed by individual users with no IT involvement.
Run a Structured Department Survey
Tech discovery alone won’t catch everything. Some AI tools are being accessed through personal accounts — no SSO, no corporate card, no extension. The only way to surface those is to ask directly.
Send a short, plainly worded survey to department heads and team leads. Keep it to five questions or fewer. Ask what AI tools their team uses, how often, for what purpose, and whether any of those tools handle customer data, employee data, or anything confidential. Don’t frame it as a security audit — frame it as an inventory exercise to help IT support the tools people are actually using. You’ll get higher response rates and more honest answers.
You’re not trying to catch people using unauthorized tools. You’re trying to build a complete picture. Make that clear in how you introduce the survey, and you’ll find people are often relieved to have a legitimate channel to surface the tools they’ve been quietly using.
One practical note: send this survey from a named person, not a generic IT alias. A message from “the IT team” gets ignored. A message from a specific person asking a direct question gets answered.
What Information to Capture for Each Tool
Once you’re pulling results from SSO data, expense reports, and department surveys, you need a consistent schema for what you’re recording. A spreadsheet works fine at this stage. The goal is a structured record, not a polished system.
For each AI tool you identify, capture at minimum: the tool name and vendor, the category of use (writing, coding, data analysis, customer interaction, etc.), the department or team using it, whether it’s accessing or processing any company data (and if so, what type), how the tool was procured (IT-approved, department-procured, personal account), and who the business owner is. That last field matters more than most people realize — more on that in a moment.
You’ll also want to note whether the vendor has a data processing agreement in place, and whether the tool falls into a higher-risk category under any framework you’re tracking. If your organization has EU customer or employee data in scope, the EU AI Act classification for that tool type is worth recording even at this early stage. The EU AI Act establishes risk tiers that vary significantly by use case — a general-purpose writing assistant sits in a very different category than an AI tool being used to screen job applicants or evaluate creditworthiness.
Don’t let perfect be the enemy of done here. An incomplete inventory that captures 80% of your AI tools is vastly more useful than a perfect template that never gets filled in.
Assigning Ownership Is the Step Most Organizations Skip
A tool inventory without ownership is just a list. It doesn’t tell you who’s responsible when something goes wrong, who should be reviewing the vendor’s data practices, or who needs to be involved when a regulation changes the compliance requirements for that tool.
Ownership at the tool level should default to the department head or team lead whose team is the primary user. IT owns the infrastructure; the business owns the tool and its use. This distinction matters when you’re trying to get action taken — an IT manager telling a department that their AI tool has a data exposure risk lands differently when the department head has been named as the tool owner in your inventory.
For tools that multiple departments use, assign a primary owner and document the other stakeholders. The primary owner is accountable for reviewing the vendor’s terms, confirming the data processing agreement is in place, and flagging any changes in tool behavior or vendor policy. You’re not asking department heads to become privacy lawyers. You’re asking them to be the point of contact and to care about the risk.
When you assign ownership, send a brief one-paragraph summary of what that means in practice. Most department heads have never been asked to own an AI tool before. A short explanation of what you’re asking — review the vendor’s data terms once a year, notify IT if you expand the tool’s use to new data types — removes the ambiguity and makes the assignment stick.
Maintaining the Inventory Without a Dedicated Process
An AI tool inventory that gets built once and never updated is only marginally better than no inventory at all. The challenge for mid-market IT teams is that AI tool adoption is moving fast — new tools get trialed and adopted on timescales that outpace annual reviews.
The most practical solution is to make AI tool disclosure part of existing processes rather than a standalone exercise. Add a field to your new software request form that asks whether the tool has AI capabilities. Add a line to your onboarding checklist that asks new hires to disclose any AI tools they plan to use. Route any AI-related expense under $50/month through a lightweight approval email rather than a formal procurement process — the goal is awareness, not friction.
Quarterly, pull a fresh pass of your SSO and expense data and diff it against your existing inventory. New tools that appear in that diff get added. Tools that show no usage for 90 days get reviewed for decommission. This takes about two hours per quarter for most mid-market organizations, not a full-time role.
You should also establish a simple intake path for employees to self-report AI tools they’re considering using. A short form, a shared email alias, or even a dedicated Slack channel works. The barrier to report should be lower than the barrier to just start using the tool without telling anyone.
Do This Week
If your current AI tool inventory is empty or outdated, start here: pull your SSO OAuth connections and your last 90 days of corporate expense data, filter for anything that looks like an AI tool, and build a flat list in a spreadsheet. Don’t try to categorize or risk-rate anything yet. Just get the list.
Once you have 10 to 20 tools on that list — and most mid-market organizations will find at least that many — you have enough to start assigning ownership and filling in the schema fields described above. That first pass of real data, done this week, is worth more than any governance framework document you could write from scratch.
A Note on Tooling
Everything described in this guide is achievable with a spreadsheet, a survey tool, and access to your existing IT data. You don’t need an enterprise GRC platform to start.
That said, once your inventory exists, the next challenge is risk-rating the tools in it — evaluating vendor data practices, flagging high-risk use cases, and tracking whether your controls are actually in place. That’s where a purpose-built tool starts to pay off. InfoDefenders’ AI Risk Assessor is built specifically for mid-market IT teams who have an inventory and need a structured way to assess what’s on it, without the overhead of an enterprise platform.
But that’s a next step. Right now, the inventory comes first.
