Nearly half of IT security decision-makers say their organizations have experienced sensitive data exposure through an AI tool. More surprisingly, in 77% of those incidents, a company-approved AI tool was involved.
Those findings, from new research commissioned by BlackFog, highlight a growing problem for organizations embracing AI: Approving an AI tool doesn't necessarily mean controlling what information employees share with it or the associated risks.
Here’s the problem: Nearly four in ten companies believe AI adoption is moving faster than their organizations can govern it. Among those struggling to keep pace, only a quarter report having comprehensive AI controls in place.
Businesses are adopting AI to improve productivity and accelerate operations – that has been well documented. Security teams, meanwhile, are struggling to establish the visibility and controls needed to prevent sensitive information from being exposed. For MSPs, that creates an opportunity to address a growing security challenge that technology approval and written policies alone cannot solve.
Why Approved AI Tools Still Create Risk
One of the most revealing findings in the research may be that sanctioned AI tools were involved in more than three-quarters of reported exposure incidents. It challenges the assumption that selecting and approving an AI platform is sufficient to protect company information.
To be clear, the problem isn't necessarily that organizations are approving the wrong tools. Even a properly vetted AI platform can introduce risks when employees share information they shouldn't, connect AI to sensitive repositories, or use AI for purposes outside established policies. Tool approval addresses only one part of the risk. Controlling how the tool interacts with business information is a separate challenge.
Most organizations recognize the need for governance. The BlackFog research found that 98% have an AI usage policy, but enforcement remains inconsistent. Thirty-five percent have written policies specifying what employees must not enter into AI prompts without technical enforcement behind those restrictions. Only 56% support their broader AI usage policies with comprehensive technical controls.
The problem isn't a lack of written rules, but a gap between policy coverage and the ability to enforce those rules consistently.
Visibility is another issue. While 62% of respondents can identify which users or departments are sending data to AI tools, only 48% can produce any record of the types of data that actually left their organizations. Knowing who is using AI isn't the same as knowing what information they shared.
For security teams, this exposes three distinct challenges:
- Identifying risky activity,
- Preventing unauthorized disclosures, and
- Documenting what happened.
Visibility alone doesn't provide protection, while prevention without adequate reporting can leave organizations struggling to investigate incidents or demonstrate compliance.
The weaknesses extend to specific categories of sensitive information. Across areas such as sales data, proprietary code, and intellectual property, only about half of organizations can technically prevent risky sharing. Controls are particularly weak around proprietary design, creative and marketing content. Specifically, only 32% can detect or monitor sharing of this information but cannot consistently prevent it, while another 20% rely primarily on written policies or training.
These aren't necessarily conventional customer records or financial information, but they can be equally important to a business. Product designs, creative assets, marketing strategies, and proprietary content often represent significant competitive value.
The challenge becomes more complicated as AI evolves beyond employee-operated chatbots.
More than half of respondents (56%) say autonomous AI agents acting on employees' behalf fall outside their organizations' existing controls. Among technology and telecom respondents, that figure climbs to 70%.
Unlike employees manually entering information into a chatbot, autonomous agents may retrieve information from multiple systems, interact with external services, and take actions without individual human approval. That makes identity, access permissions, data handling, and auditability central to AI governance. Controls designed around employees using approved applications may not adequately address software acting on their behalf.
Shadow AI creates additional exposure. Only 54% of respondents say they can block or restrict unsanctioned AI tools, while another 32% can detect their use but cannot consistently stop it. That means nearly have of companies either can’t detect or can’t stop unapproved AI.
The long story short is that organizations tend to be better equipped to establish acceptable-use policies and identify AI activity than to consistently control how sensitive information moves through approved applications, unauthorized services, and AI agents.
Turning AI Governance into a Managed Service
The BlackFog survey focused on organizations with more than 500 employees, so its findings cannot be directly applied to smaller businesses. That said, it stands to reason that if larger SMBs struggle with AI governance and visibility, smaller businesses with less expertise and fewer resources are likely to have even greater challenges.
For MSPs serving those customers, AI governance represents a potential extension of existing cybersecurity, identity management, and compliance services. MSPs already manage many of the systems through which employees access and share business information, including endpoints, identities, cloud applications, and collaboration platforms. That puts them in a position to help customers move from written policies to operational controls.
The opportunity isn't simply to add another AI security product to the portfolio. Rather, it’s to provide an ongoing service that helps customers identify AI-related risks, implement appropriate safeguards, and demonstrate that those protections are working.
For example, a practical managed AI data protection offering might follow four steps.
Discover: Inventory AI usage across the customer's environment, including approved applications, unsanctioned tools, and autonomous agents. Identify where sensitive data is stored, who can access it, and how it may be shared with AI services. An AI risk assessment provides a natural entry point for identifying gaps and establishing priorities.
Govern: Translate acceptable-use policies into specific rules based on data classifications, user permissions, and business requirements. Restrictions involving customer information, source code, financial records, and proprietary content should reflect the organization's actual risks rather than relying on general instructions not to share sensitive information.
Enforce: Implement technical safeguards where sensitive information can be accessed, shared, or transferred. Depending on the environment, that may involve endpoint protection, data loss prevention, identity and access controls, AI application configurations, and restrictions on autonomous agents. Prevention should complement detection and monitoring, not simply generate additional alerts.
Prove: Provide regular reporting on AI usage, policy violations, prevented disclosures, and unresolved risks. Customers need evidence they can use to assess whether their controls are effective, support compliance activities, and communicate security priorities to business leadership.
This framework creates a logical progression from an initial assessment to implementation and ongoing management – and recurring revenue through continuous monitoring, policy updates, access reviews, and governance reporting.
This is not to suggest MSPs assume responsibility for every decision involving AI. Customers must establish their own risk tolerance, determine acceptable uses, and decide which information requires protection. The MSP may provide consultation, but its real role is to help translate those decisions into enforceable controls and maintain them as technologies and business requirements change.
That’s an important piece of the equation because AI governance is not a one-time implementation project. New tools, integrations, and autonomous agents introduce new access paths and potential exposures. Controls that work today may need to be revised as customers adopt additional capabilities or change how existing systems are used.
There is a real service opportunity, but capitalizing on it requires more than adding AI monitoring to an existing security package. MSPs need expertise in data governance, identity management, and AI application security, along with repeatable processes for assessments, implementation and reporting.
Providers must also determine which capabilities they can deliver internally, where partnerships make sense and how to price services profitably. Not every customer will need the same level of protection or ongoing oversight, making flexible service packages valuable for attracting a range of customers.
But, for MSPs that are willing to take on this challenge, AI governance offers a pathway to deeper customer relationships and engagement with business leaders on questions extending beyond traditional cybersecurity. It also creates opportunities to demonstrate value through measurable outcomes rather than just the number of tools deployed. Increasingly, the MSP story is becoming less about the number of tools and more about business outcomes, and that’s where AI governance has its strength.
Edited by
Erik Linask
