A Red Hat survey found a gap between German organisations’ confidence in their AI governance and their ability to monitor it: 94% already use AI agents in production, but only 43% can trace where data is stored and processed.

An AI agent can summarise documents, retrieve information from a business application and carry out tasks automatically. For employees, this may feel like a single, simple step. Behind the scenes, however, the process can involve multiple applications, cloud services and external AI models. Without a clear view of data flows, software components and providers, organisations may miss emerging risks.

A recent CIO survey by Red Hat reveals a striking gap: most German respondents are confident that they operate AI agents securely. Yet many lack the operational visibility needed to verify that confidence.

AI governance in German organisations: six findings

• 94% already use AI agents in production.
• 80% are confident their AI agents operate securely and comply with regulations.
• 43% can precisely track where data is stored and processed.
• 38% have a complete overview of their software supply chain, including direct and indirect dependencies.
• 33% continuously assess AI provider risks during operation using measurable criteria.
• 77% say they can patch critical software vulnerabilities within 24 hours of a fix becoming available.

Confidence meets gaps in visibility

The survey suggests that safeguards such as guardrails, monitoring and automated intervention are in place at many organisations, but their effectiveness cannot always be fully verified. Forty-five per cent of German respondents say they rely entirely on tested guardrails, continuous monitoring and automated intervention mechanisms. A further 35% are “fairly confident” but acknowledge gaps in safeguards and monitoring.

This matters because AI environments are dynamic. New models and features are introduced, data sources change and applications connect to additional services. A system that meets security and compliance requirements at launch may develop new risks over time. Governance therefore needs to cover the full lifecycle of an AI application.

Data flows remain difficult to trace

Only 43% of German CIOs surveyed say they know their organisations’ data flows precisely and where information is stored and processed. Thirty-five per cent report minor visibility gaps. Twelve per cent have introduced AI features without fully understanding the data flows, while 6% have only very limited visibility into where data is stored.

For organisations facing strict data protection and regulatory requirements, this transparency is essential. They need to know what data an AI application processes, where information is transferred and which services are involved. Without that insight, it becomes harder to identify unauthorised data transfers and demonstrate compliance.

The survey also shows significant differences internationally. In South Africa, 61% of respondents say they can fully trace data flows, compared with 31% in France. For Germany, the findings point to a clear need for better visibility, despite the confidence many respondents express in their existing safeguards.

Rapid patching depends on knowing what is in the software supply chain

German organisations are relatively confident in their ability to address critical software vulnerabilities: 77% say they can apply a patch within 24 hours of a fix becoming available. The rest say they need a few days; no respondents reported taking longer than a week.

However, a quick response is only possible if organisations know which components they use and where those components are deployed. Just 38% have a complete and continuously updated overview of their open-source components, including their origin, versions and direct and indirect dependencies. A further 37% know their components and versions but have gaps in their knowledge of origin or indirect dependencies. Twenty-three per cent lack a complete overview, while 2% do not systematically track their open-source software supply chain.

Indirect dependencies can create particular blind spots. Libraries and frameworks often incorporate components from other projects, allowing a vulnerability to enter a product without appearing in an organisation’s immediate software inventory. The survey found that 62% of respondents consider indirect dependencies a significant risk.

Provider risk needs continuous assessment

The survey also points to a gap between initial assessments and ongoing oversight of AI providers. Only 33% of German organisations continuously assess provider risks during operation using measurable criteria. A further 34% assess new AI technologies before deployment but struggle to monitor them afterwards. Another 21% rely solely on an initial assessment.

Models, data processing practices, threats and regulatory requirements can all change. Risk assessments therefore need to be repeated and supported by criteria that organisations can monitor internally. Five per cent of respondents rely solely on external audits, certifications and contractual assurances. A further 7% rely on provider reputation and past experience, or have no structured approach to assessing provider risk.

“Effective AI governance does not depend solely on policies and review processes; it first requires comprehensive visibility,” says Gregor von Jagow, Country Manager Germany at Red Hat. Organisations need to understand data flows, the software components they use and the risks posed by AI providers throughout the lifecycle of AI agents. Monitoring and governance mechanisms should therefore be in place before AI and agents are scaled.

Censuswide conducted the survey among 1,400 CIOs aged 25 and over in 14 countries, including Germany, between 24 August and 1 September 2026. The results reflect respondents’ own assessments. They also highlight what organisations need to address as they expand AI use: confidence in safeguards cannot replace a verifiable view of data, dependencies and providers.