Shadow AI has gone agentic. Are organizations ready?
For CIOs and governance leaders, the immediate priority is to establish a sanctioned build-and-run environment before prohibition drives more activity underground. That means approved platforms, identity-based access, logging, agent registration, and clear thresholds for human review. The trade-off is important: too much friction suppresses useful innovation, but too little control creates unmanaged execution risk, hidden dependencies, and weak auditability.
There is also a portfolio and capability question. If employees are encoding workflow logic and operational know-how into unsanctioned agents, the enterprise may be losing reusable process IP and creating key-person risk in a new form. Leaders should ask which agent use cases merit enterprise funding, which should remain local experiments, and how successful automations are documented, transferred, and supported when roles change.
Practical next questions:
- Do we know where agents are acting, not just where AI is being queried?
- Who approves agent access to systems, data, and decision thresholds?
- What minimum controls are required for registration, monitoring, and incident response?
- How will we retain the workflow knowledge embedded in employee-built agents as an institutional asset?
We’re long past the era of shadow AI looking like unsanctioned chatbots. Employees aren’t just searching for answers through tools like ChatGPT and Claude anymore. They’re building and deploying their own AI agents to take actions and complete tasks on their behalf, often without their organizations’ knowledge or oversight.
Nearly three in 10 employees are already using unsanctioned AI agents at work, according to a 2025 Microsoft survey of 1,725 data security leaders. In a recent survey of more than 1,000 workers who use AI weekly, Adobe found that nearlyhalf of the employees who receive agentic AI training go on to build their own custom agents. That’s a significant escalation from employees copying sensitive information into a chatbot.
While a chatbot answers a question, an agent can log in to systems, move data among them and make decisions that used to require a person in the loop. When AI agents can access systems, execute tasks and make decisions independently, the risks to organizations grow accordingly, and most aren’t prepared for them.
These tools are increasingly acting like employees. They complete assigned tasks and develop their own workflows — except no one is managing their performance.
Employee-driven innovation should be encouraged, but it must come with the same basic guardrails organizations apply to any new hire: visibility into what employees are doing, boundaries around what they can access, and accountability when something goes wrong.
Many employees simply aren’t aware of the risks they’re creating when they build and deploy agents. That’s not necessarily because they’re being careless, but because the tools make it remarkably easy to build something powerful before the risks are fully understood.
AI adoption is already outpacing most organizations’ ability to govern it. A recent Deloitte survey of 3,235 IT and business leaders found that only 21% of organizations report having mature governance models in place for agentic AI risk, while more than half of the organizations surveyed by the Cloud Security Alliance this year have experienced agents exceeding their intended permissions. That gap should concern any leadership team. It means most organizations deploying agentic AI, whether they’ve sanctioned it or not, have already seen it behave in ways they didn’t intend.
But the exposure of sensitive data is only part of the story. Agentic shadow AI also creates a longer-term risk: the loss of institutional knowledge.
When an employee builds an agent outside the corporate infrastructure, they may embed valuable organizational knowledge into it, such as workflows, decision-making logic, process expertise and instructions built through their experience. If that agent isn’t visible, documented or governed by the organization, that knowledge can effectively walk out the door when the employee does. The company loses not only the employee, but — potentially — also the systems they built to make their work more efficient.
The answer isn’t to shut that innovation down. It’s to give people a shared environment to build it, one where organizations can understand what agents are being created, put appropriate controls around them and ensure that valuable tools and knowledge remain with the business.
That also requires organizations to rethink the infrastructure beneath AI adoption. Data, content and business activity need to be accessible and legible to AI systems. Otherwise, employees will continue to find their own workarounds and build outside approved environments.
This is the broader AI-native shift for which organizations need to prepare. Moving beyond chat interfaces and embedding AI into how a business operates takes time. My organization expects that transformation to unfold over roughly two years.
That is exactly why organizations need to start now. The longer governance lags behind adoption, the more institutional knowledge and enterprise value organizations stand to lose before they even realize it’s happening. Every company that wants to lead in this new AI economy should prioritize turning that experimentation into lasting organizational capacity, so the knowledge, workflows and value employees create stay with the organization and scale as the company evolves, rather than walking out the door with them.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.