Shadow AI Is Shadow IT All Over Again, Just Faster

Gartner has been tracking shadow IT for years, and the estimate I keep coming back to is that 30 to 40 percent of IT spending happens outside the IT department. That number has been stable for a while, which I think tells you something about how deeply entrenched the workaround is in organizational life. But the acceleration with AI is what bothered me when I read it. Employees are using ChatGPT at work without telling anyone. They are copying corporate data into public models. They are building internal workflows on top of APIs their organizations do not know exist. The 80 percent internal policy violation rate that AI security platforms report is not a new problem. It is the same problem shadow IT always was, just faster, because AI tools bypass every friction point that used to slow shadow IT down. No installation. No procurement. No IT request. A browser tab and a free account is all it takes.

Ferneley and Sobreperez (2006) argued that workarounds have been the employee's response to system misfit long before AI existed. They identified three types. Working around means bypassing the system entirely for a specific task, doing the work through other means. Working with means using the system in a way its designers did not intend, bending it to serve a purpose it was not built for. Working through means accepting the misfit and absorbing the cost silently, using the system as designed even when it is inefficient. The third type is the most dangerous for organizations because it produces no complaint, no ticket, no signal. People just do their jobs slightly less efficiently every day, and nobody ever learns that the official system is the bottleneck.

I wrote about this classification before in my post on workarounds and shadow IT, where I argued that shadow IT is more diagnostic than it is dangerous. Each type tells you something different about the gap between what the official system provides and what the work actually requires. Working around is a loud signal: the system cannot do this task at all. Working with is a quieter signal: the system can do this, but not the way the work actually proceeds. Working through is the silent signal that is easiest to miss: the system makes this harder than it should be, and people have given up complaining about it.

Shadow AI maps onto all three types. An employee who drafts a document in ChatGPT because the official content management system requires six approval steps and three different logins is working around. An employee who feeds proprietary data into a public AI model because the approved enterprise AI assistant lacks the domain-specific knowledge to answer the question is working with. An employee who types a prompt into the approved tool, receives a generic output that requires significant editing, rewrites it manually, and never reports the inefficiency is working through. The tool is in use. Compliance metrics look fine. The misfit is completely invisible to the organization.

What changed with AI is the access cost. Traditional shadow IT required significant effort. Installing unauthorized software, procuring cloud services, setting up infrastructure outside IT control, these actions took time, money, and some level of technical skill. AI tools removed almost all of that friction. A free ChatGPT account is created in thirty seconds. It requires no IT approval, no credit card, no installation. The switching cost between the official tool and the unauthorized one is effectively zero. When the official AI tool does not fit the task, the employee does not spend weeks designing a workaround. They open a new browser tab. The workaround is instant, which means it is now the default response to misfit rather than the exception.

Goodhue and Thompson (1995) built task-technology fit to explain exactly this dynamic. Their argument is that technology improves performance when its functionality matches the task requirements. Fit is not a global property of a tool. A tool that fits one task perfectly can be a terrible fit for a different task, even inside the same organization. The same approved AI assistant that works well for drafting marketing copy may be useless for analyzing financial data or summarizing legal documents. When the task does not fit the tool, the employee does not stop needing to do the task. They find a tool that fits. Every unauthorized ChatGPT session is a task-technology fit diagnosis that the organization is not collecting. I covered this more broadly in my post on task-technology fit, but the mechanism is the same: performance depends on the match between what the tool does and what the work needs, not on whether the tool is good in the abstract.

Banning specific AI tools does not solve the misfit. It pushes it underground. I have watched organizations block ChatGPT and observe employees switch to a dozen other models that do essentially the same thing with less name recognition. I have seen IT departments enforce blanket bans while adoption of the approved tool declined at the same rate that shadow AI increased. The approved tool was worse than what employees could find on their own, and nobody asked why. The ban makes the workaround invisible. Invisible workarounds are more dangerous than visible ones because the organization loses the ability to learn from them. The employee who was working with ChatGPT on a task the approved tool could not handle will find another model, and the misfit will remain unaddressed.

This is why I think AI security platforms that only monitor and block will fail at their stated mission. I wrote about why these platforms are better understood as governance boundary resources than security products, but the governance function only works if the organization uses what it learns. A platform that identifies blocked traffic and flags violations without feeding that information back into tool selection and task analysis is treating the symptom. The employees who get blocked from one AI tool will find another. The ones who do not get caught will continue working around the official systems. The platform must be paired with a diagnostic loop: look at what employees were trying to do, understand what task created the need, and ask whether the approved tools support that task. If the answer is no, block nothing. Fix the tool.

The compliance framing of shadow AI is, in my opinion, the wrong framing. It casts employees as a problem, people who need more training, stricter policies, better monitoring. I think the evidence from the IS literature on workarounds and task-technology fit points in a different direction. Employees are not bypassing official AI tools because they misunderstand policy. They are doing it because the official tools do not support the tasks they are actually trying to do. The workaround is the organic response to a system-task misfit. It always has been. The only difference is that AI made the response instantaneous.

The 80 percent internal policy violation rate is not telling you that employees are out of control. It is telling you that the gap between what the official AI tools provide and what employees need is larger than most organizations want to admit. If I were advising a company on AI governance, I would start with that gap, not with the compliance report. The workarounds are already written in the usage logs. You just have to read them as requirements documents instead of violations.