AI adoption breaks the mold of conventional enterprise software rollouts. Models are becoming commodities, new capabilities are outpacing procurement cycles, and the gap between an employee and an AI system is simply a browser tab or desktop app. For security teams, this means the challenge has now changed too. The task for them is no longer simply permitting and blocking AI applications, but instead controlling an ecosystem that is evolving faster than its software inventory.
This is why the term ‘shadow AI’ misses the mark. It suggests a defined category of unauthorised applications which can be discovered and removed, when the more important issue to examine is the behaviour taking place around them.
AI has broken the software perimeter
The enterprise security model was built around identifiable assets. Applications were installed and managed devices and traffic passed through known infrastructure. Generative AI doesn’t neatly fit into this model. A new capability can arrive as a web application, API or a browser extension inside software employees already use.
The distinction between ‘application’ and ‘service’ is becoming increasingly blurred. It’s an important shift, because it means the security decision is no longer binary. An organisation doesn’t necessarily need to decide whether an AI provider is sanctioned or unsanctioned, good or bad. Instead, it must understand what a user is doing with that service, which information is involved and what controls are appropriate for this interaction.
Consider an employee using an external model to analyse a publicly available market report. There’s no privacy or security risk here, so there is no need to prevent the action. However, the risk profile changes hugely if the same employee submits an internal report full of customer data or commercially sensitive material.
In this case, the model isn’t the problem. The employee’s interaction is. Organisations can blacklist AI services within the organisation, but they will struggle to keep pace with the number of models and interfaces employees can access. An alternative is to move security closer to where data and AI intersect.
The browser and desktop are now parts of the AI control plane
The browser used to be treated primarily as a way to access applications. In an AI-enabled workplace, it is becoming far more significant: a common layer for interacting with AI-powered models and agents. That creates an important security consideration. Network controls can establish where traffic is routed and endpoint controls can establish which device is being used. However, neither necessarily provides enough context about what is being input into an AI system. The prompt itself is the security boundary. If an employee inputs sensitive information into an AI model, it’s not so much that the user accessed an AI provider. It’s that a particular piece of data was about to cross a specific boundary in a specific context.
The browser is not the only place this happens. However, a growing share of AI use now sits in desktop applications, such as standalone chat clients, coding assistants and agents installed on the device, which fall outside what browser-based controls can see. Visibility is therefore the first requirement: security teams need to know which desktop AI apps are in use, who is using them, which accounts are signed in and what data is being attached or sent. As with browser use, the aim isn’t to block by default, but to understand the behaviour first and apply proportionate controls.
Those controls must operate with sufficient context to distinguish between harmless content generation and potentially harmful data movement. This may mean applying different policies to corporate and personal accounts. The risk isn’t the model itself but in how employees use it, and whether this use falls within the organisation’s existing controls.
Treat shadow AI like free company market research
There is another reason not to view unsanctioned AI usage purely through a risk lens. Employee behaviour gives a real-time indication of where AI is proving useful. If hundreds of people independently adopt the same model or extension without guidance from IT, that’s useful data to investigate before simply blocking it.
Traditionally, technology adoption is engineered from the top down; a platform is selected, deployed and then measured against expected usage. AI often reverses this sequence. Employees experiment first, identify capabilities that improve their workflows and create demand for the organisation to respond to. That makes shadow AI usage a valuable source of intelligence for CIOs and CISOs.
It can reveal where sanctioned tools fall short, which AI capabilities are generating genuine productivity gains and where demand is emerging before it becomes visible through formal processes. The security function therefore has an opportunity to become more than the team that decides which AI tools employees cannot use. By analysing how people are actively using AI, the wider organisation can determine where investment is justified and where controls need to evolve. Of course, not every unsanctioned AI agent should be permitted based on end-user popularity alone.
Sensitive data should still be protected, identity still matters and organisations still need clear governance around high-risk use cases. Yet blocking all behaviour without understanding it throws away useful information. With frontier AI models and applications changing faster than enterprise technology strategies, the most valuable signal comes from the tools employees choose when nobody is instructing them what to use.
Organisations can govern AI on its own terms. Rather than trying to predict every application employees may use, focus controls on the interactions themselves to understand how AI is changing the way work gets done. If organisations view shadow AI as only bad behaviour, the answer will always be another policy or another block. IT and security leaders have an opportunity to address the cause rather than chasing the symptom.



