Context
The company came to us convinced their AI footprint was small and manageable. About 200 employees, mostly in software development and professional services, with a handful of go-to-market and support teams. IT had approved two AI tools organizationally: ChatGPT Team for general productivity and GitHub Copilot for the engineering org. Both were licensed, both were on the books, and leadership considered the AI governance question essentially settled.
They weren’t wrong to start there. Most mid-market companies that have done any AI governance work have done exactly this: approved one or two flagship tools, written a short acceptable use paragraph into the employee handbook, and called it a program. That’s not negligence—it’s a reasonable first step when you don’t have a dedicated compliance function and your IT team is already carrying a full load.
What nobody had done was look at what else was running.
Problem
The engagement started with a standard scoping conversation. When we asked the IT manager to walk us through the company’s AI tool inventory, he pulled up a spreadsheet with two rows. ChatGPT Team. GitHub Copilot. That was it.
We asked a few follow-up questions. Did marketing use any video or image generation tools? He wasn’t sure—that was managed by the marketing director. Did HR use any AI-assisted sourcing or screening tools? Possibly a free tier of something, but nothing formal. Did customer support use any browser extensions or summarization plugins? He genuinely didn’t know.
Those three questions told us what we needed to know before we even started the audit: the two-row spreadsheet wasn’t an inventory. It was a list of tools IT had approved. The actual AI surface area of the organization was an open question.
This is the core problem we see repeatedly at this company size. AI tool adoption has decentralized faster than governance has scaled. Individual contributors and team leads are making AI tool decisions daily—not to circumvent IT, but because the tools are free or cheap, they solve an immediate problem, and nobody told them to ask first. By the time IT finds out a tool exists, it’s already embedded in a workflow.
The risks hiding in that gap aren’t hypothetical. An unvetted AI tool processing resume data may be storing candidate PII on servers your legal team has never reviewed. A browser plugin summarizing customer support tickets may be sending conversation data to a third-party model provider under terms your DPA doesn’t cover. A free video generation tool may be training on uploaded content under a license agreement your marketing team clicked through without reading. None of these are exotic threat scenarios. They’re the ordinary consequences of adoption outpacing oversight.
Approach
The discovery phase took 48 hours. That’s not a marketing number—it reflects how quickly a structured audit moves when you know what you’re looking for and you have direct access to department leads.
We ran the work in three parallel tracks.
The first track was technical: a review of browser extension inventories, SaaS application logs, and network traffic patterns. Most companies at this size have some form of endpoint management or SSO that surfaces application usage even when the applications weren’t formally requested. This isn’t forensic analysis—it’s reading the data that’s already there.
The second track was departmental interviews. We spent thirty to forty-five minutes with a lead from each major function: marketing, HR, customer support, engineering, and operations. The questions were direct and non-accusatory. What tools are you using to get your work done faster? What did you try last month that you liked? What do you use that IT probably doesn’t know about? People answer honestly when the conversation is framed as inventory, not audit.
The third track was a policy gap review. We pulled the existing employee handbook language, the vendor agreements for the two approved tools, and whatever IT had on file for data classification. This told us what the written rules actually said versus what employees could reasonably have understood them to mean.
What came back from those three tracks in 48 hours was 43 distinct AI tools or AI-assisted features in active use across the organization.
Marketing was using three different AI video generation platforms—one paid, two free tiers—to produce social content and internal training clips. None had been reviewed for data retention or content licensing terms. HR had a recruiter who had been pasting candidate resumes into a free AI sourcing tool to generate outreach templates, a workflow she’d been running for several months without realizing it raised any compliance questions. Customer support had a browser extension installed on most rep machines that summarized long support tickets before routing; the extension had been recommended in a support community forum and installed by a team lead without an IT ticket.
Those three examples got the most attention in the debrief because they were the most concrete. But the other 40 tools mattered too—AI writing assistants, AI-powered grammar checkers, AI scheduling tools, AI meeting transcription services, AI code review plugins. Most were low-risk individually. The problem was that nobody had made that determination. They were just running.
Outcome
By the end of the engagement, the company had a complete AI tool inventory for the first time. More importantly, they had a triage framework that let them prioritize which of those 43 tools needed immediate action versus which ones could go through a standard review cycle.
The video generation tools went to legal for terms of service review before marketing used them again. The HR sourcing workflow was suspended until the team could evaluate a vetted alternative that had a proper DPA and didn’t train on uploaded content. The customer support browser extension was assessed against the company’s existing vendor security requirements; it didn’t pass, and the team found a supported alternative through their helpdesk platform that did the same job under a proper agreement.
The remaining 40 tools were sorted into three buckets: approved as-is, approved with usage restrictions documented, and pending review. That review backlog became a standing agenda item for a biweekly IT/ops sync, so it would actually get worked down rather than sitting in a document.
The IT manager’s comment at the end of the engagement stuck with me. He said he wasn’t surprised that employees were using AI tools he didn’t know about—he was surprised by how easy it was to find them once someone actually looked.
Lessons for IT Managers
If this case study sounds familiar, here’s what you can take from it.
Your approved tool list and your actual AI inventory are probably not the same thing. The gap between them isn’t a failure of IT—it’s a predictable consequence of how fast individual AI tool adoption is moving right now. Acknowledging the gap is the prerequisite to closing it.
Departmental interviews find what logs miss. Technical discovery catches tools that touch the network. It doesn’t catch the employee who’s been pasting data into a browser-based AI tool over HTTPS, which looks like ordinary web traffic. A thirty-minute conversation with a department lead will surface things that a month of log review won’t.
Free tools carry hidden costs in data terms. The sourcing AI and the video generators in this case weren’t costing anyone a line-item budget. They were costing the company data rights and compliance exposure they hadn’t evaluated. Pricing tier is not a reliable proxy for risk.
Employees aren’t the problem. In this engagement, not one person interviewed was trying to create a governance gap. They were trying to do their jobs faster. The fix isn’t a crackdown—it’s a process that makes it easy for employees to flag new tools before they become a workflow dependency, and fast enough that people don’t route around it.
Do this week: Block two hours and run your own version of the three-track approach above at a smaller scale. Pull your endpoint management or SSO dashboard and look for AI-adjacent application names you don’t recognize. Send a five-question email to your marketing, HR, and support leads asking what AI tools their teams are currently using. Then compare what you find against your approved tool list. You don’t need a full audit to find out whether you have a gap—you just need to look.
Once you know what’s running, you need a place to put it. The AI Risk Assessor gives you a structured way to work through vendor risk assessments for tools you’ve found in the wild—so you’re not making triage decisions in a spreadsheet with no criteria. At the bottom-of-funnel stage, most IT managers we work with want to see the tool before they commit; the platform is built for exactly this kind of reactive catch-up work, not just forward-looking governance programs.
The 43-tool discovery in this engagement wasn’t unusual. It’s what a realistic mid-market AI footprint looks like when nobody has looked at it yet. The companies that are ahead of this problem aren’t necessarily the ones with bigger teams or bigger budgets—they’re the ones that looked first.
