AI Security

The Shadow AI Problem: How to Find Out Which AI Tools Your Employees Are Already Using

August 19, 2026  ·  NetworkDefender Security

← Back to Blog

Ask any owner of a small professional services firm whether their staff use AI, and most will say something like "a little, maybe." Ask them which tools, on which accounts, and with what client data in the prompts, and the answer is usually silence. Not because nothing is happening. Because nobody has looked.

That gap between "some AI use" and "which AI, by whom, with what" is the actual problem. You cannot write a policy for a tool you do not know exists, and you cannot govern data you do not know left the building. This post is about closing that gap: what the current numbers say about how much you are probably missing, and five concrete ways to find out what is really running in your firm before you write a single rule.

The Gap Between What You Think Is Happening and What Is

The scale of shadow AI has moved past "some employees" territory. Verizon's 2026 Data Breach Investigations Report found regular AI use on corporate devices jumped from 15 percent to 45 percent in a single year, and shadow AI is now the third most common non-malicious insider action its investigators detect, a fourfold increase over the prior year. A separate PagerDuty survey found 66 percent of office professionals had used an AI tool at work despite believing it was against company policy, and nearly half said they would rather use AI quietly than risk being told no.

Here is the number that actually matters, though: one widely cited 2025 industry estimate puts the share of companies that can detect all the shadow AI in their organization at 12 percent. Fewer than 11 percent of AI applications in a typical workplace are visible to IT at all. That is not a failure of any particular firm's IT setup. It reflects how these tools get adopted: no install, no purchase order, no IT ticket, often just a browser tab and a personal email address.

None of this makes your staff careless. Someone in bookkeeping finds a tool that turns a messy spreadsheet into a clean summary in ten seconds. Someone drafting client correspondence finds a chatbot that fixes their tone. They are solving a real problem, and it works, so they keep using it and never think to mention it. That is exactly why asking once, in a hallway conversation, and treating the answer as complete is not enough.

Where the Risk Actually Sits

The reason this matters more than a normal software adoption question is where the money and the exposure hide.

Spend hides in places nobody reconciles against a policy. Individual AI subscriptions at 20 to 40 dollars a month sit under most expense approval thresholds, so they never trigger a second look. They show up on personal cards and get expensed, on corporate cards under a generic line item, or bundled into an upgrade tier of a tool your firm already pays for. None of that is visible from a vendor list, because none of it was ever a distinct purchasing decision.

The financial exposure is not hypothetical. IBM's 2025 Cost of a Data Breach research found breaches involving high levels of shadow AI cost roughly 670,000 dollars more on average than breaches without it, and that the large majority of the organizations involved had no governance in place to manage AI or detect unauthorized use before the breach happened. The gap is not between firms with perfect AI security and firms with none. It is between firms that know what is running and firms that find out during an incident.

And the tools themselves are getting harder to see. AI is no longer only a chatbot in a browser tab. It shows up as a feature quietly switched on inside software you already own, as a browser extension that summarizes whatever page is open, as a coding assistant on a developer's personal account, or as an add-on connected through an OAuth grant that never touched a firewall. Traditional shadow IT left a visible footprint of installed programs and network traffic to known destinations. Shadow AI increasingly does not, which is why asking "do we have a firewall rule for that" is no longer a complete answer.

Five Ways to Actually Find Out

You do not need an enterprise security budget to get real visibility. You need to check five places, and most firms have never checked any of them.

  1. Ask directly, and frame it as an inventory, not an audit. This sounds too simple to matter, and it is consistently the fastest way to surface the bulk of what is in use. A one-page form asking "what AI tools do you use for work tasks, and what do you use them for" works because staff are usually proud of the shortcuts they have found, not hiding them out of malice. The framing decides whether you get honest answers. If the first person who admits to using a personal ChatGPT account gets reprimanded, everyone else stops volunteering information and the tool stays in use anyway, just invisibly. Ask as a project to build an approved list, not as a search for who broke a rule that did not exist yet.
  2. Review expense reports and card statements for AI vendor charges. Scan for OpenAI, Anthropic, Google, Microsoft, Perplexity, and the dozens of smaller AI vendors, across both corporate cards and reimbursed personal expenses. This will not catch free tier use, since that is the point of a free tier, but it reliably surfaces the paid subscriptions your firm is already funding without knowing it, including tools bundled into a software upgrade nobody flagged as an AI purchase.
  3. Check your DNS and web traffic logs for AI destinations. If your firm runs its internet traffic through a filtering or SASE platform, most of these now log connections to known AI domains as a matter of course, separate from whatever content filtering policy you have set. Reviewing that log for a few weeks shows you which AI tools staff are reaching, from which devices, without needing to install anything new on an endpoint. This is where a platform like Cloudflare's Secure Web Gateway earns its keep even if you never configured it specifically for AI, since app visibility and shadow AI detection are part of what it already sees at the DNS and HTTP layer.
  4. Audit OAuth grants and connected apps in your core platforms. Microsoft 365 and Google Workspace both keep a list of third party applications employees have granted access to their account, and AI tools that summarize email, draft documents, or read a calendar almost always request that kind of access to work. Most owners have never opened this list. It takes a few minutes and routinely turns up connections nobody remembers approving.
  5. Look at browser extensions on managed devices. AI writing assistants, meeting summarizers, and "chat with this page" tools distribute almost entirely as browser extensions now, which sidesteps app installation entirely and leaves nothing for traditional endpoint tools to flag. A quick inventory across your firm's devices, even a manual one if you are small enough, will surface tools that never touched a network log because their AI processing happens through the browser itself.

Run all five, even loosely, and you will have a real list instead of a guess. Compare that list against what a proper business tier and a signed data agreement look like, the distinction our earlier post on using AI without leaking client data covers in detail, and you know exactly where your actual exposure sits instead of where you assumed it did.

What to Do With What You Find

Discovery is the first step, not the whole project. Once you have a real list:

  1. Sort what you found into three buckets. Tools already on an approved business tier with the right data terms, tools that could be moved to one, and tools that have no safe version and need to stop.
  2. Do not punish what surfaces. The fastest way to make your next discovery effort useless is to make people regret being honest in this one.
  3. Publish the approved list you now have evidence for, instead of one built from a guess about what staff might be using.
  4. Recheck on a cadence, not once. New tools spread inside a firm in weeks, not years. A light version of these five checks twice a year keeps the list honest.
  5. Feed what you learn into your written AI policy and risk assessment, so the paper trail reflects what is actually running, not what you hoped was running.

The Bottom Line

Most firms do not have an AI problem so much as a visibility problem. The policy, the approved tool list, the vendor questions, none of it means much if it is written against a guess. Five checks, most of which take an afternoon and cost nothing, turn that guess into a real inventory. From there, governing AI use is the same discipline you already apply to every other vendor with access to client data: know where it is, decide on purpose, write it down.

NetworkDefender Security helps Wisconsin CPA firms, law firms, and insurance agencies find out what AI tools are actually in use, build the approved list and policy around real findings, and configure the logging and tenant controls that make the next review easier than the first. Contact us to schedule a security review.

Key Resources

Sources