Unmanaged AI agents are proliferating across enterprise platforms, creating new risks for Kenyan and East African businesses. Here’s how to detect, assess and govern these autonomous tools before they undermine security.
A finance manager connects a new AI-driven workflow to the company’s accounting system. Within hours, invoices are being processed automatically: faster than ever, but with no record in the IT team’s inventory. No one in security reviews the permissions, and the agent’s access persists long after the manager changes roles. This pattern is repeating across Kenyan and East African businesses as shadow AI agents proliferate, moving faster than formal governance can keep up.
Unlike traditional shadow IT, these agents are not just unsanctioned apps or SaaS accounts. Shadow AI agents: autonomous tools built or enabled by employees on platforms like Salesforce, Microsoft Copilot Studio, Zapier, or Retool: connect directly to critical business systems. They hold persistent permissions, act independently, and often operate without any oversight from IT or security teams. The result is a rapidly expanding attack surface that is invisible to most organisations’ existing controls.
Recent security advisories highlight the scale and urgency of the problem. As AI-driven automation becomes more accessible, the risk of unsanctioned agents acting with broad, unchecked permissions grows. For Kenyan and East African business leaders and technical teams, the challenge is clear: unmanaged automation can undermine security, compliance, and operational integrity: unless discovered and governed proactively.
Why Shadow AI Agents Are a Bigger Risk Than Shadow Apps
Shadow AI agents differ fundamentally from unsanctioned apps or chatbots. Where a shadow app might simply access data, an agent can take autonomous actions: issuing payments, updating records, or triggering workflows: based on its programming or AI-driven logic. Critically, these agents often retain persistent, high-level permissions. If misconfigured or compromised, they can cause real-world changes to systems or data, far beyond the consequences of a poorly configured chatbot.
This distinction matters. A shadow AI agent with access to a customer database or financial system presents a direct risk to data integrity and privacy. The agent’s actions are often opaque, logged poorly (if at all), and may continue even after the original creator leaves the organisation or changes roles. This creates both technical and operational blind spots for IT and compliance teams.
How Shadow AI Agents Evade Traditional Discovery
Most organisations attempt to discover shadow IT using platform APIs or network monitoring. This approach is increasingly ineffective for AI agents. Many platforms do not expose detailed agent activity through their APIs, or restrict what can be queried. As a result, agents created via no-code or low-code tools, or embedded within SaaS platforms, often remain invisible to central IT inventories.
Even diligent IT teams can miss agents that operate entirely within user accounts or are created through integrations with third-party services. The lack of standardised logging and discovery APIs across platforms like Salesforce Agentforce, Microsoft Copilot Studio, Zapier, and Retool leaves significant gaps in visibility. The net effect: a growing population of autonomous agents with unknown reach and risk.
Multi-Layered Discovery: Combining API and Browser-Based Methods
To close these visibility gaps, security researchers and vendors recommend a multi-layered approach to discovery. While API-based inventory remains useful where available, it must be supplemented with browser-based and client-side monitoring. For example, browser extensions can passively observe when employees interact with or create agents on supported platforms, capturing details that APIs miss.
Continuous client-side behavioural monitoring adds another layer. By collecting signals throughout a user session: such as navigation patterns, interaction timings, and workflow triggers: organisations can distinguish between human and automated (agentic) activity. Systems like Cloudflare’s Precursor demonstrate how these techniques can help identify unauthorised automation, even when agents attempt to mimic legitimate users.
Assessing and Governing Shadow AI Agents
Discovery is only the first step. Once shadow AI agents are identified, organisations must assess the risk they pose. Key questions include: What permissions does the agent hold? Is it accessible from outside the organisation? Is it dormant or actively in use? Who is the technical owner, and is the agent’s purpose still valid?
- Catalogue each agent, recording platform, creator, permissions, and integration points.
- Assess public accessibility and exposure to external systems.
- Identify dormant agents: those with persistent permissions but no recent activity.
- Assign a technical owner for each agent, ideally the current business process owner.
- Set approval status and require periodic reviews, especially after role changes or departures.
- Prompt owners to remediate or decommission agents that present excessive risk or are no longer needed.
This governance process is ongoing, not a one-time audit. As new platforms and automation tools emerge, so too will new shadow agents. Regular reviews and clear assignment of ownership are essential to maintain control.
Limits, Exceptions, and Unknowns
No detection method is foolproof. API-based discovery is limited by what platforms expose; browser-based monitoring requires user consent and may not cover all environments. Behavioural analysis can generate false positives, especially in complex workflows. There is no direct evidence yet of the scale of shadow AI agents in Kenyan or East African businesses, and regulatory requirements remain undefined. These steps are documentation-derived and should be tested in your environment before large-scale deployment.
The Next Action: Secure Your Automation Landscape Now
Businesses in Kenya and East Africa cannot afford to ignore the rise of shadow AI agents. Begin by mapping your automation landscape using both API and client-side methods. Establish clear governance for every agent discovered, with defined ownership and approval. Regularly review agent inventories and permissions. As automation accelerates, proactive discovery and governance are the only way to prevent unmanaged AI agents from undermining your security and operations.
