
The usual response to shadow AI is to tighten AI governance, often resulting in a stricter set of policies and a longer approval process. In the client programs I work with, that often tends to backfire. Staff members still have deadlines, and when the approved route is slow, they opt to use whatever tool is fastest irrespective of whether or not it has been approved by the company.
The programs that work begin by finding out which AI tools people already use and putting an executive in charge of the risk. They decide how tight the rules should be after that. My colleague Aaron O’Reilly and I recently ran a client webinar on AI governance gaps, and when we polled attendees on their biggest gap, not knowing which tools were in use tied for first.
Your web logs show only part of the AI your people use
The difficult reality is that shadow AI is harder to find than shadow IT ever was. Much of it runs in a browser, where asset management tools can’t see it, or it manifestsas an assistant built into software you already license. Web traffic logs show you who visited which AI sites, but they won’t show you the AI feature someone switched on inside a tool you bought two years ago.
The clients I see making real progress start by casting a wide net to identify AI use, combining insights from several sources that already exist within their companies. Identity systems record new OAuth consent grants, which can show someone signing into an AI service with a work account. Browser telemetry can list installed AI-enabled browser extensions. Endpoint inventories catch desktop apps, and data loss prevention (DLP) tools have started flagging data that employees send to AI tools.
The most useful, but often overlooked sources I’ve seen came from a client’s finance department. After exhausting the above tools, their security team started meeting with procurement and purchasing to go through expense reports and purchase orders, looking for anything AI-related, including products where AI is a feature and not the selling point. It’s the same approach we recommend for finding hidden vendors, since teams that skip IT still have to pay for what they use.
Spending is a security signal too. We’ve seen cases where attackers stole AI API keys and ran up large usage bills in a very short time.Spending limits and usage alerts on every AI account can give you a way to catch and prevent that API exploitation, and the same billing data can tell you what’s in use.
Once you can see usage, you can measure it. One of the most important metrics you can track is the share of each department using approved tools in the approved way. The way to ensure that share increases and ensure a commensurate decrease in the use of unapproved AI is to build a governance process that is led from the top, thorough, but not overly cumbersome.
An AI policy needs an executive behind it
Too often, when we discuss AI risk ownership with clients, there’s a prevailing sense that it’s largely the concern of someone else within the org. . Without a clear owner, the AI policy ends up with IT by default, and that’s where it stalls, because the rest of the business treats a rule from IT as an IT preference. Aaron and I have both watched IT teams try to enforce AI policies without executive sign-off and get nowhere.
The crucial and imperative fix is to secure vocal and outspoken support at the executive level for AI governance and risk reduction.. That person needs to support, often and out loud, why the AI governance program matters to the business, and for that they don’t need to be an AI or security expert. I’ve written about why security programs stall without that kind of sponsor, and AI governance programs are no different; they stall in exactly the same way.
An executive owner also settles a question many policies leave open, which is what happens when someone breaks the rules. An employee who pasted customer data into a chatbot because nobody told them not to needs training. Someone who knowingly went around the controls is a different conversation, and ownership at the executive level gives the governance program and its associated policies the weight they need to remain effective..
Slow approvals send people around your controls
If getting an AI tool approved takes weeks, people will sign up for a free account and use it anyway. It’s the same story we’ve heard about shadow IT for years. Aaron describes the risk as security falling back into its old reputation as the office of no. “If you create that type of environment, people are just going to find ways to get around your controls,” he said.
The clients handling this well make the approved route faster than the workaround. They ask for a business purpose up front, which keeps requests short and ties spending to something the business wants. They also treat the policy as a draft. Aaron tells clients to expect two or three rewrites in the first two years, because the tools change that fast.
Training works better when it’s specific to the job. Finance and HR teams handle sensitive data every day, so they need the clearest guidance on where AI helps them and where it’s risky. It also helps to find the people already getting good results from AI and build the approved version around what they’re doing.
Vendor contracts are your view into AI you didn’t build
A lot of the AI in your environment arrives inside vendor software, and you can’t inspect those models, so you have to get your visibility through the contract. Ask for a clause that bars the vendor from training models on your data. You’ll also want to know who else can access that data, and you’ll want notice of material model changes and security incidents.
In Aaron’s experience, vendors usually accept no-training terms and breach notification terms and push back on audit rights and access to model internals. You probably won’t change the terms with the large model providers, but smaller vendors adding AI features to their products will usually negotiate.
Time your approval process against a free signup
One quick way to check your own program is to time it. Find out how long it takes an employee to get a new AI tool approved, and compare that with how long it takes to sign up for a free account. When the gap runs to weeks, your policy is producing shadow AI, and tightening it will widen the gap. In that case, begin by identifying bottlenecks in the approval process and then working to remove them. In my experience, an effective governance process is one that balances the speed and agility that the business demands with the controls and protections that the business requires.
Aaron and I cover more of this, including where the NIST AI Risk Management Framework and ISO 42001 fit, in the full webinar recording.
Frequently asked questions
What is shadow AI and how do you find it?
Shadow AI is any AI tool employees use without approval from IT or security, including AI features built into software an organization already licenses. Security teams can find it by combining OAuth consent grants from identity systems, browser extension and endpoint inventories, data loss prevention alerts, and expense and procurement records.
Who should own AI risk in an organization?
AI risk should have an executive owner who visibly supports the governance program and gives the program and its policies the weight they need to be followed.
Why can strict AI governance increase shadow AI?
When approving an AI tool takes much longer than signing up for a free account, employees tend to skip the approval process. Faster approvals tied to a business purpose, role-specific training, and regular policy updates help keep AI use visible.



