Shadow AI Agents: The Overlooked Risk in AI Governance

As the first wave of excitement settles, another is rising about the risks and consequences of the world’s mass and rapid adoption of AI.

One such example is autonomous AI agents – companies are still rushing to deploy them across their operations, but many are doing so without proper oversight.

This is what one IBM executive describes as a dangerous new form of shadow IT.

Hans Petter Dalen has spent years watching enterprises struggle with emerging technologies.

As IBM’s AI for Business Leader for EMEA, he’s seen firsthand how businesses jump on new tools without thinking through the consequences.

The latest concern? AI agents that don’t just assist employees – they make decisions on their own. 

The AI agents’ problem

It’s easy to see why these AI agent systems are popular. 

They can handle complex tasks through simple conversations, routing support tickets, suggesting pricing changes or even chatting with customers directly. 

Business units often deploy them quickly, sometimes bypassing IT departments entirely.

“These aren’t just unmanaged tools – they’re decision-makers,” Dalen says. 

While shadow IT refers to employees using software or systems without approval from the IT department, shadow AI agents pose a more significant threat – because they are autonomous AI tools that operate without oversight and make decisions independently.

“Agents can interact with business-critical systems, trigger downstream workflows and evolve based on real-time data.”

But this speed creates problems.

When AI agents operate without central oversight, he says companies lose track of what these systems are actually doing.

“If AI agents aren’t centrally registered, inventoried and monitored, they become ‘shadow agents’: operating out of scope, out of oversight and potentially out of alignment with enterprise policy.”

IBM research revealing widespread AI struggles

The scale of this AI agent challenge is becoming clear through IBM’s research. 

The company’s Think 2025 study found that nearly two-thirds of CEOs believe their business success over the next three years depends on adopting advanced AI. 

Yet only a quarter of organisations report that their AI initiatives have delivered what they promised.

Dalen reckons poor governance is part of the problem. 

AI agents can fail in ways that traditional software doesn’t – they can generate false information, develop biases or drift from their original programming over time.

“These agents – like any AI model – can experience inconsistent performance, hallucinations, bias, drift or misalignment with evolving compliance standards,” he explains. 

“Left unchecked, they can introduce reputational risk, decision-making errors and even legal violations.”

History repeating with Cloud computing parallels

For Dalen, this feels like déjà vu. He’s watched companies make similar mistakes with cloud computing adoption. 

Initially, departments deployed cloud services independently, often without involving IT teams. This creates security vulnerabilities and compliance gaps that took years to address. “This was evident in early cloud adoption,” Dalen recalls.

“Initially, cloud platforms were deployed ad hoc by departments, often without IT involvement. Over time, that decentralisation led to security breaches and compliance gaps.”

Eventually, companies learned to govern their cloud infrastructure properly – not to stop innovation, but to make it sustainable. Dalen believes the same approach is needed for AI agents.

“What’s needed now is AI-specific governance for this new generation of tools – not to limit their use, but to ensure they’re deployed responsibly and deliver sustainable value,” he argues. 

But he warns: “Speed shouldn’t come at the expense of security or accountability.”

Building visibility into shadow operations

Dalen’s solution starts with a basic principle: companies must understand what AI agents are operating in their environment. 

This sounds straightforward, but many organisations have no systematic way to discover these systems.

“Enterprises need tooling that can automatically discover AI applications and agents operating in the environment – even those deployed by business users without formal approval,” he explains. “After all, you can’t govern what you can’t see.”

Once companies can see their AI agents, they need to catalogue them properly.

“Agents should be registered, categorised by function and mapped to a relevant owner or business process,” he points out.

“Each agent’s scope – what it can access, decide or trigger – should be clearly defined.”

The next step involves assessing risk. 

Dalen asks the questions that matter: “What data does the agent handle? Is it accessing regulated systems? Could its outputs influence financial or legal decisions? Organisations should apply tiered governance depending on an agent’s level of autonomy and potential business impact.”

Preventing chain reactions from agent failures

One aspect that particularly worries Dalen is how multiple AI agents often work together. 

If one agent in a chain behaves unpredictably, it can cause cascading failures across business processes. “If one agent behaves inconsistently, the entire value chain falls apart,” he explains. 

“Monitoring of individual agents based on a pre-defined set of metrics is essential across the entire value chain.”

This requires building what Dalen calls “explainability and traceability” into AI systems. 

Companies need to track what their agents are doing and why. 

“Outputs should be auditable. Toxicity thresholds should be established and monitored. Agent decisions should be logged with clear attribution – so if something goes wrong, the root cause can be identified and corrected.”

Breaking down organisational silos

Dalen reckons that governing AI agents isn’t just a technical problem – it requires different parts of the organisation to work together in new ways.

“AI product teams, data scientists, security professionals and compliance leaders need to collaborate early and often,” he says.

“AI security and governance shouldn’t sit in siloes; they must be integrated into the entire development and deployment lifecycle.”

This collaboration is becoming more urgent as regulators pay closer attention to AI systems. 

The EU AI Act, for example, places specific requirements on high-risk AI systems around transparency and oversight.

Autonomous agents could easily fall into this category depending on how they’re used.

“Enterprises that fail to map and govern their agents may soon find themselves out of step with global policy trends,” Dalen warns.

But beyond avoiding regulatory trouble, Dalen argues that good governance builds trust with customers, regulators and employees.

“By showing that AI systems are accountable, transparent and fair, organisations can unlock broader adoption and avoid the reputational backlash that has accompanied some early AI missteps.”

“The enterprises that govern them early will lead with confidence, not caution,” Dalen says. 

“Now’s the moment to bring visibility to the shadows and put trust at the centre of AI innovation.”