Google Cloud, Financial services, AI
In this article:
-
What is Shadow AI and how is your organization already facing it?
-
Getting GEAP right: Why implementation is where most organizations struggle
The AI tool your employees are using right now is probably not yours.
Across most organizations, AI adoption looks less like a strategy and more like a sprawl. Every employee has a personal ChatGPT or Gemini account. 8 in 10 employees worldwide are actively bypassing corporate AI governance and the problem compounds at the top: senior leadership is 50% more likely to use Shadow AI than the average employee. They use it daily, for real work: drafting proposals, summarizing client calls, analyzing data, writing code. And in doing so, they are feeding your company’s most sensitive information into systems you do not own, cannot audit, and have no contractual protection over.
The consequences are dangerous. The proprietary data shared in those sessions can be used to train future model versions, meaning your internal processes, client details, and competitive strategies could surface in someone else’s AI response months from now.
This is what Shadow AI means. And most organizations are already deep inside it before anyone thinks to govern it.
Now layer agents on top of this. A chatbot waits for a prompt and responds. An ungoverned agent reasons, plans, connects to your CRM, your files, your databases, and takes actions autonomously, repeatedly, on behalf of anyone who has access to it. The exposure is no longer session-based. It is continuous and system-wide.
Employees move at the speed of productivity. Governance moves at the speed of process. AI has made that gap wider, faster, and more expensive than any previous technology cycle. The question for CTOs is how to close the gap before the gap closes your business.
That’s why it’s time to strengthen the line between secure and compromised.
This CTO’s guide, based on a conversation between Jack Davies, Data and AI Cloud Engineer at Revolgy, and Miro Vlasák, CEO of Revolgy, covers what Shadow AI is, why the shift to agents makes it significantly more dangerous, and what a technically credible governance solution looks like.
What is Shadow AI and how is your organization already facing it?
By definition, Shadow AI is the use of AI tools by employees without IT approval, security governance, or organizational oversight. The entire organization, operating on personal accounts, with no visibility from the top. By the time your security team is drafting a governance framework, Shadow AI has already been running inside your organization for months. 1 in 5 organizations has already experienced a breach linked to Shadow AI, adding up to $670,000 to average breach costs per incident.
What makes it structurally dangerous is the invisibility. A personal AI session does not trigger your CASB (Cloud Access Security Broker). It does not look like a threat in your network monitoring. It looks like an employee working. Traditional security tools were built for a different era, and they cannot classify AI API calls as a distinct risk category. You cannot govern what you cannot see, and right now most organizations cannot see most of their AI activity.
The result is a compliance gap that compounds quietly. And the organization finds out too late, usually during an audit, or not at all.
Chatbots vs agents: Which one is the bigger security risk?
With chatbots, the data exposure is bounded by what the employee manually shares in a session. The risk exists but the scope is human-controlled. An employee decides what to paste, the model processes it, and the session ends.
Agents change this at an architectural level. An agent has tool access: it queries your databases, reads your files, calls your APIs, and passes data between systems as part of completing a task. For example, take a financial performance agent connected to BigQuery that an employee can instruct in natural language to pull last quarter’s numbers, generate charts, and compile them into a presentation deck. Useful, fast, and genuinely impressive. But if that agent is ungoverned, the security implications compound fast:
- Any employee with access to the agent can retrieve sensitive financial data from your data warehouse without a formal data request or approval
- The agent has whatever permissions it was given at setup, which are rarely scoped correctly in practice
- Every query it runs looks like normal API traffic to your existing security tooling
- When shared across a team, one agent creates continuous retrieval activity across your most sensitive systems with no centralized audit trail
With a chatbot, an employee decides what to share. With an agent, the agent decides what to retrieve, on behalf of whoever is using it, repeatedly and autonomously. That distinction is what makes agentic Shadow AI a categorically different risk. The blast radius is not larger because agents are smarter. It is larger because agents act on your systems directly, persistently, and at scale.
The solution: Governing Shadow AI
Governing Shadow AI is an operational discipline, and it requires three technical layers working together. Discovery and visibility come first, then access control and enforcement, and finally continuous monitoring. Each depends on the previous one being in place.
- The starting point is inventory. You cannot classify risk or enforce access controls for AI activity you have not discovered. This means deploying tooling that identifies every AI application running across your organization, sanctioned or not, and keeping that inventory current as new tools emerge and old ones get repurposed. Most organizations skipping this step believe they already know what is running. They are wrong.
- Once visibility is established, classification and access control follow. Not every AI tool or agent carries the same risk profile, and controls applied uniformly across all of them either block too much or protect too little. Effective governance enforces access at the identity layer automatically, based on who the employee is and what they are already authorized to see, not based on which tool they happen to be using.
- The final layer is continuous monitoring. Agent behavior drifts. Permissions that were scoped correctly at deployment expand over time. Token consumption compounds faster than anyone budgets for. Monitoring is what keeps governance from becoming a one-time configuration exercise that decays the moment it ships.
Doing this across a live organization, with agents already running, employees already working around controls, and new tools emerging weekly, is operationally complex. Most engineering teams underestimate the scope until they are already mid-deployment with gaps in permission scoping and no audit trail to fall back on. A unified platform that handles discovery, access enforcement, and monitoring under one control layer is the only way to make governance sustainable at scale. That is what GEAP is built to do.
Govern Shadow AI with Gemini Enterprise Agent Platform
Gemini Enterprise Agent Platform (GEAP) is Google Cloud’s unified infrastructure for building, deploying, governing, and optimizing AI agents at enterprise scale. For organizations dealing with Shadow AI, it addresses the problem at the root: instead of restricting access to AI, it gives employees a governed, secure alternative that is more capable than anything they are running on personal accounts.
The platform connects to the systems where your data already lives, including Jira, Salesforce, Google Workspace, Slack, and Microsoft Office, under one centralized control layer. Google contractually guarantees that none of this data will be used to train future Gemini models. It stays within your environment.
Here is how GEAP addresses Shadow AI across three layers:
1. Visibility and auditability
- Agent Identity assigns every agent a unique cryptographic ID, creating a traceable audit trail for every action taken
- Organizations get full observability into which data sources are being queried, how frequently, and by whom
- Agent Registry provides a centralized inventory of every agent running across the organization
2. Security and access control
- IAM controls inherited directly from Google Workspace mean an employee cannot query data through an agent that they do not already have permission to access
- Model Armor inspects prompts in real time, identifies sensitive information, and can block or adjust requests before they reach the model
- Agent Gateway acts as the enforcement point for every agent connection, controlling which systems an agent can access and under what conditions
3. Governance at scale
- Agent policies apply automatically based on employee identity, meaning governance scales with adoption without additional setup per user
- All agents, whether built by IT or by individual employees, operate under the same universal governance rules
- Shadow AI activity that was previously invisible becomes centrally monitored and managed
Security is the entry point, but GEAP is not a security product. It is the complete infrastructure for enterprise AI: the same platform that governs your agents also provides the tooling to build them, the runtime to scale them, and the evaluation framework to optimize them over time.
Getting GEAP right: Why implementation is where most organizations struggle
GEAP is a comprehensive platform, and that comprehensiveness is exactly what makes it complex to deploy correctly. Getting the governance layer configured, the agent architecture designed for your specific workflows, and the IAM scoping set up correctly from day one requires experience that most engineering teams are building from scratch.
The gap between a working proof of concept and a production-grade deployment is significant. Teams that underestimate it typically end up with:
- Agents deployed without proper permission scoping, recreating the same Shadow AI exposure they were trying to eliminate
- No evaluation framework in place, meaning agent performance degrades over time without anyone catching it
- Cost controls missing, with token consumption and runtime charges compounding faster than anticipated
- Adoption that stalls because employees were never properly onboarded into the new governed environment
The technical setup is only part of it. Identifying the right use cases, sequencing the rollout, connecting your existing tools securely, and training the teams who will work alongside agents daily are all decisions that determine whether GEAP delivers value or becomes another underutilized platform investment.
This is where a managed implementation partner changes the outcome. Revolgy is a Google Cloud Premier Partner that has deployed GEAP across industries and organizations at varying stages of AI maturity. The engagement model follows three phases: foundation and quick wins, custom agent development, and adoption and optimization. Each phase is designed to build on the previous one, so governance is in place before scale, and scale happens before optimization.
Frequently asked questions
Our employees are already using GEAP. Does that mean we’re covered?
Not automatically. Access to the platform is not the same as governed deployment. If IAM permissions were not scoped correctly at setup, if there is no audit trail on agent activity, and if employees are running agents without centralized oversight, you have the same exposure problem. Just inside a Google-branded environment.
How is an ungoverned agent different from an ungoverned chatbot, practically speaking?
A chatbot processes what an employee manually pastes into it. An agent has tool access. It queries your databases, reads files, calls APIs, and passes data between systems autonomously. One employee with access to an ungoverned agent can continuously retrieve sensitive data from your most critical systems with no approval process and no audit trail.
What does “proper permission scoping” actually mean for agents?
It means an agent should only be able to access data that the employee using it already has direct permission to access. In practice, most agents are deployed with over-permissioned credentials at setup, meaning the agent can retrieve far more than the person instructing it is authorized to see. GEAP inherits IAM permissions from Google Workspace to enforce this at the connection level.
What is the realistic timeline to go from Shadow AI exposure to governed deployment?
It depends on your current stack, how many agents are already running, and how your IAM is structured. A phased approach works best: visibility and quick wins first, custom agent development second, adoption and optimization third. Trying to govern everything at once before employees are onboarded typically stalls adoption and recreates the Shadow AI problem inside the governed environment.
Does Google use your data to train Gemini models?
On Gemini Enterprise Agent Platform, Google contractually guarantees that none of your organizational data is used to train future Gemini models. Your data stays within your environment.
What is Model Armor in GEAP?
Model Armor is a runtime protection layer that inspects prompts and requests in real time, identifies sensitive information, and can block or adjust them before they reach the model.
What is the difference between Shadow AI and managed AI?
Shadow AI is an ungoverned AI activity running outside IT oversight. Managed AI is the same capability operating within a centralized governance framework where access is controlled, activity is audited, and data stays within the organization’s security perimeter. GEAP is designed to move organizations from the former to the latter.
