Businesses have spent years investing in identity management, access controls and policies intended to govern how employees use corporate information. Those investments have improved security in many organizations, but they haven't always addressed the authority employees possess to introduce new applications and extend access to information outside the company's direct control. The rapid adoption of AI is bringing that issue into focus because employees can now connect external applications to corporate systems with very little effort or technical knowledge.
The concern isn't necessarily that employees are deliberately circumventing security requirements. Many are using legitimate capabilities provided by the applications their organizations already support. An employee can authenticate through Microsoft 365 or Google Workspace and authorize an AI application to access information associated with that account. The process may be technically permitted and function exactly as designed, even though nobody responsible for managing the business's exposure has evaluated the application or considered what the authorization means.
I've spent much of my career looking at how organizations make decisions about technology and risk. What concerns me about this development is how easily a decision made for a relatively narrow business purpose can create obligations and dependencies that extend beyond the person making it. AI is making these decisions more frequent, while the responsibility for understanding their consequences remains distributed across functions that don't always work from the same information.
What the Research Reveals
A new DNSFilter report, The Cost of Shadow AI, provides useful evidence of the problem. The company surveyed 400 IT and cybersecurity professionals at North American organizations with between 150 and 5,000 employees and examined AI-related network activity across more than 43,000 organizations. According to the survey, 46.1% of respondents reported that employees had connected unapproved AI applications to corporate systems, including Microsoft 365, Google Workspace and single sign-on environments. Another 20.1% reported that one of those connections had resulted in a security incident.
The findings concerning authorization are particularly interesting. In 36.8% of surveyed organizations, employees could grant an AI application access to corporate systems and notify IT afterward. Another 6.5% didn't require notification. These arrangements allow an external application to receive access before the people responsible for evaluating the associated risks necessarily know the relationship exists. The employee may have a legitimate business reason for using the application, but that doesn't establish whether the access granted is appropriate for the organization.
DNSFilter also compared what IT professionals believed employees were using with activity observed on corporate networks. Only 0.5% of survey respondents estimated that employees used more than ten AI tools, while the company's network observations found that 15.2% of organizations connected to more than ten distinct AI services during a typical business day. These measurements describe different things, and a network connection doesn't establish that an application has been granted access to corporate information. Nevertheless, the difference suggests that organizations may have less visibility into AI adoption than they believe.
The research comes from a security vendor and should be considered in that context. Its survey findings describe reported experiences rather than independently verified incidents, and the network observations cannot establish whether every connection represented approved business use. Even with those limitations, the findings raise a reasonable question about whether organizations understand the authority employees are exercising when they connect AI applications to existing systems.
The Problem Is in the Authority Being Delegated
Organizations have invested heavily in authenticating employees and controlling access to corporate resources. Multifactor authentication, conditional access and identity management platforms have become established parts of many security programs. These controls are important, but the ability to authenticate a legitimate user doesn't necessarily establish whether every action that user is permitted to perform is appropriate for the business.
An employee may have legitimate access to customer correspondence because that information is necessary for their job. When the same employee authorizes an external AI application to access their mailbox, the organization has introduced another party into the handling of that information. The technical authorization may be valid, but the business may never have intended the employee's access to include the ability to establish that relationship.
OAuth and similar authorization mechanisms make these integrations possible without requiring employees to disclose their passwords. They also provide ways to limit permissions and manage access when configured appropriately. The difficulty arises when organizations allow users to grant permissions without adequately considering the business authority associated with those decisions.
This is where I believe the issue becomes more significant than shadow AI itself. Businesses routinely delegate operational authority to employees so they can perform their responsibilities without unnecessary delays. That authority has limits, even when those limits aren't reflected in the permissions available through a technology platform. A department manager may be authorized to purchase a productivity application within an approved budget while having no authority to engage an external provider to process confidential customer information.
When those decisions can be made through the same application interface, the distinction may never receive appropriate consideration. The organization has effectively allowed a technical permission to determine the scope of a business decision, even though the consequences of that decision may involve obligations the employee was never expected to manage.
How These Decisions Become Decision Debt
I've used the term decision debt to describe what happens when organizations continue operating under decisions whose assumptions have changed or were never fully understood. The consequences tend to develop gradually because the original decisions often appear reasonable and produce an immediate benefit. Problems become apparent later, when the business has changed and the arrangements established through those decisions are no longer appropriate.
AI provides a particularly clear example. An employee introduces an application to help prepare reports or organize customer information. The application becomes useful, and other employees begin relying on it. Eventually, the department incorporates it into its normal activities without anyone formally deciding that the organization should depend on the provider. The original approval, if there was one, may have concerned a relatively minor productivity improvement, while the resulting dependency has become considerably more important.
The circumstances surrounding that decision can continue changing. The employee who authorized the application may move to another position, the provider may introduce additional capabilities or the company may change how the information being accessed is classified. Unless the organization periodically reviews the arrangement, the original permissions may remain in place without anyone considering whether they are still necessary or appropriate.
This pattern isn't unique to AI, and I've seen variations of it throughout enterprise technology environments. Organizations accumulate applications and dependencies through decisions made at different times by people with different responsibilities. The combined effect can be difficult to understand because the business rarely examines those decisions together.
What makes AI different is the speed at which applications can be introduced and the extent to which they can interact with existing information and business processes. A tool initially used to retrieve information may later acquire capabilities to perform actions across connected systems. If the organization hasn't maintained an understanding of the original authorization, it may also struggle to recognize when the nature of the exposure has changed.
Why More Security Spending May Not Resolve the Problem
One of the more revealing findings in the DNSFilter research is that 74.5% of respondents reported increased cybersecurity budgets during the previous year. AI security and governance was the leading priority for additional investment, yet many respondents remained uncertain about the effectiveness of controls already in place. The report found that 61.5% couldn't confirm that all their security tools were properly configured, while 61.8% couldn't confirm that all were actively monitored.
I find those figures more concerning than the number of AI applications employees are using. Organizations appear willing to invest in managing the problem, but additional technology doesn't necessarily improve their understanding of how existing controls are operating. A business can purchase tools intended to identify unauthorized applications while continuing to permit access decisions that bypass the intended approval process.
This is a familiar problem in cybersecurity. Controls are implemented to satisfy a requirement, address an identified weakness or demonstrate progress against a framework. The organization may subsequently treat implementation as evidence that the risk has been adequately managed, particularly when responsibility for maintaining the control is assigned to another function or service provider.
The distinction becomes important when the control depends on several parts of the organization working together. Security may establish requirements for third-party access, IT may administer the identity platform and business departments may decide which applications employees need. Each function can fulfill its assigned responsibilities while the organization continues allowing connections that nobody has evaluated in their entirety.
I've encountered this throughout my work in enterprise security architecture. The underlying issue is often the way responsibility has been divided, leaving nobody with a complete understanding of how a business decision is authorized and what happens after it has been implemented. Addressing that problem requires examining how decisions move through the organization and whether the available evidence demonstrates that the intended controls are effective.
The Insurance Implications Are Worth Examining
There is a cyber insurance consideration here, although I would be careful about suggesting that an unauthorized AI integration automatically creates a coverage problem. The circumstances of an incident and the wording of the applicable policy determine how an insurer responds. An unauthorized connection that exposes customer information may raise different questions from an application that contributes to an improper financial transaction or an error in a professional service.
The concern is whether the business understands its exposure well enough to evaluate those possibilities before a loss occurs. Cyber insurance applications often ask about access management, third-party relationships and the security controls used to protect sensitive information. A company may have implemented the controls described in its application while remaining unaware that employees can authorize external applications through their existing accounts.
That doesn't establish that the company's insurance representations are inaccurate. It does suggest that organizations should understand how their controls operate in practice before relying on those controls to describe their exposure. The distinction matters because insurance decisions depend on the information provided, and the consequences of a material misrepresentation can become significant when a claim is reviewed.
I've argued that cyber insurers are increasingly influencing cybersecurity expectations for small and medium-sized businesses. Underwriting requirements can encourage organizations to implement controls they might otherwise postpone, and that influence can be beneficial. The limitation is that an insurance application provides only a particular view of the organization at a particular point in time.
If businesses are introducing AI applications and changing access relationships throughout the year, their understanding of risk needs to keep pace. Waiting until the next insurance renewal to examine those changes leaves considerable room for the organization's actual exposure to diverge from the assumptions underlying its coverage.
Where Integrated Assurance Fits
This is the kind of problem that led me to develop Integrated Assurance. Organizations need a way to connect decisions made across different business functions so they can understand the resulting exposure and determine whether it remains acceptable. The objective is to establish enough continuity between the original decision and its ongoing consequences that the business can recognize when circumstances require reconsideration.
In the case of AI integrations, the business should understand which applications have access to corporate systems and the purpose for which that access was granted. The permissions should reflect the authority the organization intended to delegate, and there should be a way to determine whether those permissions remain appropriate as the application and the business change.
Much of the technical capability needed to support this already exists within modern identity platforms and security tools. The challenge is connecting those capabilities with the decisions being made by employees and business managers. Organizations need to determine who has authority to approve external access and what evidence is required to demonstrate that the resulting arrangements remain within acceptable limits.
This doesn't mean every application decision needs executive approval. Businesses need employees to make routine decisions efficiently, and unnecessary restrictions can encourage people to work around established processes. The more useful approach is to establish appropriate boundaries for delegated authority and make those boundaries enforceable through the systems employees already use.
Integrated Assurance provides a way to maintain that relationship between business decisions and the controls intended to support them. The evidence produced should help leadership understand whether the organization is managing its exposure and where additional attention is necessary. That same information can support discussions with insurers, auditors and other parties who need confidence in how the business operates.
The Dependency May Eventually Matter More Than the Access
The DNSFilter findings focus primarily on unauthorized access, but I believe another issue deserves consideration as AI becomes more deeply embedded in business operations. Applications introduced for relatively minor productivity improvements can gradually become essential to activities the organization depends on to generate revenue or serve customers.
The dependency may develop without a formal decision to make the application business-critical. Employees adapt their work around the technology, processes change and alternatives become less familiar. Eventually, the organization may discover that an application originally introduced for convenience is difficult to remove without interrupting normal operations.
This becomes important when the application must be disconnected because of a security concern or becomes unavailable through a provider outage. The organization may have adequate controls for protecting its own systems while having limited ability to continue an important business activity without the external service.
Business continuity arrangements don't always account for dependencies that develop in this way. An application may never have appeared in the organization's original recovery planning because it wasn't considered critical when it was introduced. As its role changes, the assumptions supporting those plans need to be reconsidered.
This is where the discussion connects to Business Survivability. The organization needs to understand the activities it must be able to sustain and the dependencies that determine whether those activities can continue during a disruption. Cyber insurance may provide financial protection for qualifying losses, but the ability to maintain operations depends on arrangements that must exist before an incident occurs.
Bringing the Decisions Back Into View
I don't believe the appropriate response to the DNSFilter findings is to discourage AI adoption. Businesses are finding legitimate value in these technologies, and the ability to connect applications with existing information is often central to that value. The issue is whether organizations have maintained enough control over those connections to understand the exposure they are accepting.
For companies using Microsoft 365, Google Workspace or similar platforms, reviewing existing third-party application permissions would be a reasonable place to begin. That exercise can identify applications that were never formally evaluated and connections that may no longer serve a business purpose. It can also reveal whether the organization's approval process reflects the authority employees actually possess within its technology environment.
The findings should inform how the business manages future integrations, particularly as AI applications acquire greater capabilities to perform actions across connected systems. Organizations will need to understand how their existing identity controls and approval processes accommodate those changes and whether the evidence available is sufficient to demonstrate that access remains appropriate.
My concern is that many businesses will respond to AI risk by adding another layer of policy or purchasing another security product without examining how the underlying decisions are being made. Those measures may help, but their value depends on whether they address the conditions that allowed the exposure to develop.
AI is making a longstanding problem more visible. Organizations have allowed technology permissions and business authority to become disconnected, often because responsibility for the resulting decisions is distributed across functions that don't share a complete understanding of the exposure.
Integrated Assurance is intended to address that disconnect by making the consequences of business decisions visible and keeping the assumptions behind those decisions current. As AI becomes part of normal operations, that understanding will become increasingly important to the organization's ability to manage its risks and maintain essential activities.
Business Survivability ultimately depends on leadership having a realistic understanding of how the organization operates and what could prevent it from continuing. The growing use of AI is changing that operating environment, sometimes through decisions that receive very little attention when they are made. Businesses need to recognize those changes while they still have the opportunity to influence their consequences.
