The AI You Didn’t Approve
Cyber Governance in the AI Era, Edition 2. First published on LinkedIn on 2 April 2026.
In Brief
AI adoption is already happening beyond formal approval paths. LayerX's 2025 Enterprise AI & SaaS Data Security Report found that 77% of employees paste company data into generative-AI prompts, and 82% of those interactions occur through unmanaged personal accounts. The board issue is no longer whether AI will be used. It is whether the organisation has governed the data movement, judgment, and operational influence that AI use is already creating.
The problem is not future adoption. It is present exposure.
Last month, this newsletter examined what happens when organisations deliberately give AI systems authority. This month addresses the harder question: what happens when AI acquires operational influence before leadership has consciously approved it.
That is the reality of shadow AI. The analyst using a personal AI account to summarise board papers. The developer connecting an unvetted coding assistant to internal code. The manager experimenting with an autonomous workflow that touches enterprise data without any formal review.
In institutional terms, this is unmanaged delegation. Decisions, data handling, and operational judgment shift outside the control environment before leadership has sanctioned the shift.
The numbers make the exposure concrete. Cyberhaven's 2026 AI Adoption and Risk Report, based on actual usage patterns of three million workers across 222 companies, found that 39.7% of all AI interactions involve sensitive data. In leading organisations, 71.4% of employees now use generative AI. A significant share runs through personal accounts: 32.3% for ChatGPT and 24.9% for Gemini, bypassing SSO, enterprise logging, and data retention controls entirely.
IBM's 2025 Cost of a Data Breach Report, conducted by the Ponemon Institute across 600 organisations, found that 63% have no AI governance policies in place. Among those that experienced AI-related breaches, 97% lacked proper AI access controls. The gap between adoption speed and governance readiness is where the exposure lives.
Why this matters at board level
Traditional shadow IT created operational inconsistency. Shadow AI creates something more serious: outward data movement that does not come back, invisible personal-versus-corporate boundaries, and autonomous tools that shape workflows at machine speed.
When an employee pastes a contract into a consumer AI chatbot, proprietary data leaves the organisation and may be retained, exposed, or processed outside approved enterprise controls. IBM's 2025 report found that shadow AI added an average of USD 670,000 to breach costs. One in five studied organisations experienced breaches directly linked to shadow AI.
Nearly a third of ChatGPT usage runs through personal accounts (Cyberhaven, 2026). Most organisations cannot distinguish between corporate and personal AI instances on the same device. The governance perimeter your board assumes exists may not actually exist.
And shadow AI is no longer limited to chatbot queries. Employees now deploy low-code AI agents, browser extensions with AI capabilities, and autonomous workflows that connect to enterprise systems. CrowdStrike's Falcon sensors now detect more than 1,800 distinct AI applications running across enterprise endpoints. These tools operate at machine speed, often with persistent access, and without any oversight framework.
The risk is not defined by how advanced the model is. It is defined by how much institutional activity is occurring outside approved guardrails.
A note on institutional context
My career has been built at the intersection of technology governance and institutional capacity, first in public-sector digital transformation across the Caribbean, and now through DAG's advisory work and my research on institutional development in Small Island Developing States.
Shadow AI is not only a corporate security problem. It is an institutional control problem. From regulated public institutions to financial services and critical infrastructure environments, the common challenge is the same: AI use is expanding faster than institutional control models were built to absorb. Institutional judgments, procurement assessments, policy drafts, legal analyses, are being shaped by models that no one vetted, no one approved, and no one can audit.
Most organisations still discuss AI as though it were a managed transformation programme. In practice, much of the adoption looks more like institutional seepage. Capability enters through convenience, spreads through experimentation, and becomes operationally relevant before governance catches up.
For smaller economies and developing states, where institutional capacity is already constrained and public trust is hard-won, the consequences of unmanaged AI delegation are amplified. A government ministry using consumer AI to draft policy advice is not just a security risk. It is a sovereignty question about who and what is informing state decisions.
This is why I frame shadow AI as unmanaged delegation of institutional judgment, not merely as unauthorised tool use. The distinction matters for how boards and leaders respond.
Signal of the month: RSAC 2026
At RSAC 2026 in late March, shadow AI moved from theoretical concern to product category. Microsoft introduced shadow AI protection in Edge for Business, detecting sensitive data in AI prompts in real time and redirecting users to sanctioned enterprise AI when a policy triggers. CrowdStrike expanded its Falcon platform with shadow AI discovery for endpoints and cloud infrastructure, covering Copilot Studio agents, Salesforce Agentforce, and ChatGPT Enterprise.
The message from the industry's two largest security platforms is direct: shadow AI now has a detection-and-response architecture. Boards should ask whether their organisation is deploying it.
Five governance mandates
Banning AI outright is rarely effective. People route around restrictions when productivity tools are easier to access outside the enterprise than within it. The better governance strategy is to make the managed path easier, safer, and more credible than the unmanaged one.
1. Establish visibility as a governance baseline, not a security project. Management should be able to show which AI services are in use, how they are accessed, and what categories of data are moving into them. Treat AI visibility with the same seriousness you apply to privileged access monitoring.
2. Provide a sanctioned enterprise pathway that employees will actually use. Employees will default to consumer tools when the organisation offers no credible alternative with approved controls. This is predictable and preventable.
3. Classify institutional data before it leaves. Brainstorming prompts, customer records, board papers, source code, and regulated datasets should not be treated as equivalent uses. Establish clear data classification tiers and map them to AI usage policies. Pay particular attention to data that carries regulatory, fiduciary, or public-trust obligations.
4. Treat AI tools and agents as privileged actors. This connects directly to Edition 1. Any AI capability with access to enterprise systems, repositories, or decision workflows should be governed within the identity and access model. Browser extensions with AI capabilities deserve particular scrutiny.
5. Require a quarterly AI exposure review with the same rigour applied to privileged access, material vendors, and cyber incidents. AI adoption is evolving faster than annual policy cycles. Boards should expect periodic reporting on usage patterns, exposure metrics, new tools, and control effectiveness. This should be a standing board agenda item, not an ad hoc CISO report.
What boards and executives should ask this month
Do we know which AI tools our people are using, and on which accounts?
Can management show the top categories of data currently moving into AI tools?
Have we designated an approved enterprise AI pathway with documented controls?
Are browser extensions, plug-ins, and low-code agents included in our monitoring model?
Does our governance framework distinguish between sanctioned AI, tolerated AI, and prohibited AI, and do employees know the difference?
How often does this board receive an AI exposure review, and what metrics are included?
The mandate
Edition 1 asked whether your organisation had governed the authority it deliberately delegated to AI. Edition 2 asks a harder question: have you governed the authority your employees may already be delegating on your behalf, without approval?
Shadow AI is not merely unauthorised tool use. It is unmanaged delegation of institutional judgment. And for most organisations, that delegation is happening right now, at scale, without oversight.
Most organisations do not have an AI problem. They have a control problem disguised as innovation.
Executive question of the month
If AI usage is already happening through unmanaged accounts, what is your organisation's actual data perimeter?
Dr. Inshan Meahjohn is Founder and CEO of DAG (Digital Alliance Global Group) and former CEO of iGovTT, Trinidad and Tobago's national ICT agency. Subscribe to Cyber Governance in the AI Era for monthly, board-level analysis of AI governance, cyber risk and operational resilience.
Sources cited in this edition:
LayerX, Enterprise AI & SaaS Data Security Report, October 2025
Cyberhaven Labs, 2026 AI Adoption & Risk Report (3 million workers, 222 companies)
IBM / Ponemon Institute, 2025 Cost of a Data Breach Report (600 organisations)
Microsoft Edge Blog, "Protect your enterprise from shadow AI," RSAC 2026, March 2026
CrowdStrike, "CrowdStrike Establishes the Endpoint as the Epicenter for AI Security," March 2026
