For years, one of the basic assumptions of enterprise security has been that when a new technology creates unacceptable risk, organizations can restrict access until appropriate controls are in place. Generative AI is challenging that model because employees have discovered something genuinely useful. They are using AI to research, summarize information, write code, analyze data, prepare presentations, and automate repetitive work. Once employees experience those productivity gains, outright prohibition becomes difficult to sustain and may simply push usage beyond IT’s visibility.
That creates a dilemma for IT and security leaders. Organizations have legitimate concerns about what information employees share with AI services, how that information may be retained or used, which applications employees access, and whether those tools comply with corporate policies. But preventing access can create another problem. When the approved path is too restrictive, employees have an incentive to find another path. For IT leaders, the challenge is finding a way for employees to use AI without putting corporate data or systems at unnecessary risk.
See also: How Shadow AI Becomes an Enterprise Security Risk
Restriction Doesn't Eliminate Demand
Enterprise technology has been through versions of this cycle before. Employees adopted personal smartphones before many organizations had mature bring-your-own-device policies. Cloud applications spread through businesses faster than IT could approve them, and remote work accelerated the use of collaboration and file-sharing tools outside traditional corporate boundaries. Organizations eventually adapted by recognizing that effective security doesn't always mean eliminating a behavior. Often, it means creating a controlled environment in which that behavior can safely occur.
AI requires a similar change in thinking. Blocking public generative AI services may reduce their visible use on corporate systems, but it does not necessarily eliminate demand. An employee who believes an AI tool can save hours of work still has access to a personal laptop, smartphone, or unapproved service. In that scenario, the organization may have successfully blocked the application while simultaneously losing visibility into how it is being used.
That risk is not hypothetical. In Parallels’ 2026 Cloud Survey, 33% of respondents said employees frequently use unauthorized SaaS, collaboration, or remote-access tools outside IT’s control, while another 40% said it happens occasionally. The lesson for AI governance is clear: policy alone does not guarantee that demand disappears.
This is particularly problematic because AI interactions can involve corporate information. An employee may paste a document into a chatbot for summarization, use customer information to draft a response, upload code for analysis, or provide internal data to help create a presentation. What feels like a simple productivity shortcut can create questions around data protection, intellectual property, compliance, and governance.
Organizations therefore need to distinguish between controlling AI use and controlling the risks created by AI use. Those are not necessarily the same thing.
Security Needs to Follow the User
This challenge is part of a larger shift that has been underway since work moved beyond the traditional corporate perimeter. Applications, data, and users are distributed across offices, homes, cloud services, SaaS applications, and personal devices. AI now adds another layer because the interaction itself can become a new path through which corporate data is accessed, transformed, or shared.
As organizations move beyond chatbots and copilots toward AI agents that can access applications and perform actions on behalf of users, governance must address not only AI conversations but also AI-driven activities across enterprise systems. That means organizations need to think carefully about what they’re allowing AI to actually do, not just what information employees are sharing with it.
Our survey found that 74% of respondents said they are concerned about AI insecurity or vulnerability. Yet only 7% said their organizations currently use browser isolation technology as their primary approach to web browsing security. That gap illustrates how quickly the nature of the risk is changing compared with the controls many organizations have in place.
Many security architectures still depend heavily on controlling the device or blocking access to particular applications and websites. Those controls remain important, but they are no longer sufficient by themselves. When work increasingly happens through browsers, SaaS applications, and AI services, security also needs to govern the interaction: who is accessing information, what data can move, where it can go, and what actions are permitted.
Depending on the organization and use case, that might mean isolating browser sessions, controlling what information can move between an endpoint and an AI service, restricting downloads or uploads, applying identity and access policies, separating corporate data from personal environments, or delivering applications and desktops from controlled infrastructure.
No single technology solves AI governance. The architectural principle is more important: controls should be applied as close as possible to the data, access, and execution itself, without forcing employees to abandon the tools that make them productive.
Make the Safe Path the Easy Path
There is another lesson enterprise IT has learned repeatedly: employees tend to gravitate toward the technology that helps them get their work done with the least friction. That makes user experience part of the security equation.
If employees encounter blocked websites, lengthy approval processes, and blanket restrictions when they try to use AI, some will inevitably look for alternatives. If approved services are readily available and corporate policies are enforced transparently in the background, employees have far less reason to step outside the managed environment.
This principle should influence how organizations approach AI governance. Rather than beginning with a list of prohibited tools, IT and security teams can start by asking what employees are trying to accomplish. Which AI use cases deliver meaningful productivity gains? Which data should never be exposed to external AI services? Which activities require greater isolation or oversight? Where can policies be applied transparently without interrupting legitimate work?
Answering those questions starts with understanding the data involved. Public, internal, confidential, and regulated information may each require different levels of protection, monitoring, and control when used with AI services.
The appropriate controls will also depend on who is using that data and for what purpose. A developer working with proprietary source code may require very different controls than an employee using AI to improve the wording of a presentation. Someone working with regulated customer information may need a more tightly controlled environment than an employee researching publicly available information.
Effective AI governance therefore needs to be contextual. Identity, role, data sensitivity, application, and intended action should help determine the appropriate level of control rather than applying the same restriction to every employee and every AI interaction.
But putting the right controls in place is only part of the equation. Employees also need practical guidance on responsible AI use, including what information can be shared, how AI-generated outputs should be validated, and when human judgment remains necessary. Effective governance combines enablement, policy and education.
Build for AI Evolution, Not Today's AI Tools
Organizations also need to be careful about designing security policies around the AI applications that happen to dominate today. New models, agents and AI-enabled applications are appearing constantly, while AI capabilities are increasingly being embedded directly into productivity suites, browsers, operating systems and business applications. Soon, using AI may be less a distinct activity than an ordinary part of using software.
Security strategies therefore need to focus less on individual tools and more on the underlying activities: what data users can access, where that data can go, which applications can interact with it, and what actions AI systems are allowed to take.
That becomes even more important as AI shifts from answering prompts to taking actions. An AI assistant that summarizes a document presents one level of risk. An agent that can access multiple applications, retrieve corporate information, and execute a workflow presents another. Governance will increasingly need to address not just AI access, but AI execution.
As those actions become more consequential, human oversight will remain an important safeguard. Organizations may need review and approval checkpoints when AI initiates workflows or interacts with enterprise applications in ways that could have a business, regulatory, or customer impact.
This shift also has implications for the environments in which applications, data, and AI-enabled workloads are delivered and secured. The fact that 66% of organizations in our survey are actively evaluating new DaaS or VDI solutions is noteworthy in this context. The renewed interest isn't necessarily about returning to older models of computing. It points to a broader need for flexibility in how organizations deliver and secure increasingly distributed applications, data, and workloads. AI will make that flexibility even more important as organizations determine which capabilities should run locally, in the cloud, or within more controlled environments.
Enablement Is a Security Strategy
The organizations that handle enterprise AI best will not necessarily be those with the longest list of restrictions. They may be the ones that most effectively align security with the way employees actually want to work.
AI adoption is being driven by utility, and security policies that ignore that value will constantly compete against employee behavior. Security becomes more sustainable when the approved environment enables productivity rather than standing in its way.
A better strategy is to give employees approved AI options, establish clear policies about data use, apply controls around sensitive interactions, and create different levels of access based on role and risk. Just as importantly, organizations should make the secure option convenient enough that employees have little reason to work around it.
That starts with giving employees AI tools they can use with confidence. Approved enterprise AI solutions that include contractual data protections, centralized administration, and corporate governance can significantly reduce the incentive to turn to unmanaged alternatives.
Secure enablement, however, does not mean unrestricted AI. Organizations still need strong boundaries around corporate data, applications and systems. But those boundaries increasingly need to govern data, access, and execution rather than simply deciding whether a particular AI website is allowed.
That balance between access and control will become even more important as AI becomes part of everyday work. As agents begin acting as well as assisting, security leaders have an opportunity to move from a model based primarily on restriction to one built around secure enablement. The safest AI environment may ultimately be the one employees have every reason to stay inside.
