For years, organisations have treated shadow IT mainly as a policy and control problem. Employees have always found ways to use technology outside the approved IT stack, from unauthorised endpoint applications and browser extensions to personal cloud services.
What makes Shadow AI different is the speed and ease with which employees can adopt new tools. In many cases, there is no software installation, no support ticket, and no obvious indication that a new service is accessing corporate data. Employees can begin using powerful AI services within minutes, often before security teams even know those applications exist.
And while this this created security challenges, it usually left a trail. An application had to be installed, a device appeared on the network, or an endpoint agent could identify something that should not be there. From there, an analyst was able to investigate the finding, assess the risk and decide how to respond.
But first SaaS, and now AI, fundamentally change the scale of the problem. Employees are able to add new applications in seconds, often without installing software, or involving IT. They can connect corporate identities, grant access to business data, upload sensitive information and start using powerful AI tools before security teams even know those applications exist.
According to the WatchGuard Cybersecurity Hygiene Report 2026, 64% of employees report using unsanctioned AI tools for work tasks. For organisations and the MSPs supporting them, it is important to discover which applications are being used, identifying when those applications pose a risk and responding before unauthorised access or data exposure becomes a security incident.
Most unauthorised AI and SaaS applications get into organisations through one of three common entry points. Each leaves evidence in a different place and need a different approach to detection.
Employees grant access to company data
One of the most common entry points is when an employee selects “Sign in with Microsoft” or “Continue with Google” to connect an AI assistant, productivity tool or SaaS application to company resources. Within seconds, that application may gain access to email, calendars, files, contacts or other business data. Those permissions can stay in place long after the original reason it was needed.
The same process can also be exploited through consent phishing. Rather than stealing credentials, attackers trick users into authorizing a malicious application directly, potentially granting access to corporate data without ever compromising the user’s account
Attackers can also exploit this process through consent phishing. Instead of than stealing passwords, they trick users into granting an application access directly. This can provide access to company resources without the attacker needing to compromise the user’s credentials.
Organisations cannot manage application access they do not know exists, so identifying which applications are already connected to company data is an essential first step in managing the risk.
Employees use personal accounts
Not every AI tool connects through an organisation’s identity platform. Many employees simply open a browser, visit an AI service and sign in using a personal account. They can then use that service to perform work-related tasks without connecting it to their corporate identity.
This is a different visibility challenge because the activity doesn’t appear in Microsoft 365 or Google Workspace. This means organisations may have no clear picture of which services employees are using or what information they are uploading, sharing or processing through them.
The fact that an employee is using a personal account does not remove risk to the organisation. Company information may still be shared with an unauthorised service and without visibility into that activity, security teams cannot accurately assess the risk or enforce governance policies.
The only evidence is network activity
The third entry point is particularly difficult to identify because it leaves almost no obvious trail. An employee accesses an AI or SaaS service directly from a company-managed device, but the activity does not create an application permission in the organisation’s tenant or involve authentication through a corporate identity provider.
In many cases, the only evidence of this activity will appear in network or firewall logs. These connections often blend into ordinary web traffic, making it difficult to distinguish potentially risky AI use from the thousands of legitimate network connections generated during a normal working day.
Without the ability to identify and interpret this activity, organisations may be unaware that employees are accessing unauthorised AI services. This creates another potential gap in their understanding of where company information is being accessed, shared or processed.
Bringing Shadow AI into view
The challenge with Shadow AI is that there is no single way to identify every unauthorised application. Each of these three entry points leaves different evidence, so relying on just one source of information leaves significant gaps in an organisation’s visibility.
Identity platforms can reveal OAuth permissions and third-party access requests. User and application activity can show how AI services are being used, while network and firewall telemetry can uncover AI applications that leave no trace in corporate identity systems. Organizations need visibility across all three areas to build a complete picture of Shadow AI activity.
Application permissions can reveal which services have been granted access to corporate data. Identity and account activity can help organisations understand how applications are being accessed, while network and firewall activity can identify connections that may otherwise go unnoticed.
For MSPs, it is important that they bring these different sources of information together. This will allow them to build a clearer picture of the AI and SaaS applications being used across their clients’ environments, identify potentially risky activity and determine where action is needed.
As AI adoption continues to grow, organisations must move beyond just having policies that define which tools employees are permitted to use.
Effective governance requires answering three fundamental questions: What AI applications are employees using? What corporate data can those applications access? And which activities create enough risk to require action? Visibility into those answers allows security teams to distinguish legitimate AI adoption from behaviour that creates unnecessary exposure.
They need the ability to discover what is actually being used, understand the associated risks and respond before unauthorised access or data exposure develops into a security incident.
