An IBM study reveals significant gaps in the control of autonomous AI agents: 80 per cent of the executives surveyed reported escalated AI incidents. Only 40 per cent of the affected companies have subsequently adjusted their governance frameworks.

Artificial intelligence is increasingly becoming a challenge for companies in terms of information security, risk management and corporate governance. According to a recent IBM study, 80 per cent of executives surveyed across eight countries in Europe and the Middle East report that at least one AI-related incident had been escalated to board or senior management level within a year. In Germany, the figure stands at 82 per cent.

The reported incidents include malfunctions of AI systems, the disclosure of sensitive data, erroneous AI-driven decisions and security breaches.

Particularly striking: only 40 per cent of the affected companies subsequently adjusted their AI governance policies or control mechanisms. In Germany, the figure was just 39 per cent.

The findings come from a survey of 3,221 executives conducted by Censuswide on behalf of IBM. They highlight a growing discrepancy between the operational use of artificial intelligence and companies’ ability to effectively manage the associated risks.

Autonomous AI agents heighten security requirements

A particular challenge arises from so-called agent-based AI. Unlike traditional AI assistants, such systems can independently carry out tasks within their authorised scope, connect applications with one another and influence operational processes.

Already, 35 per cent of the companies surveyed are using AI agents in production. In Germany, the figure is 19 per cent.

At the same time, 75 per cent of executives in the countries surveyed expect AI agents to achieve a certain degree of decision-making autonomy by 2027. In Germany, this figure stands at 62 per cent.

Existing governance structures are only partially prepared for this development. Just 37 per cent of respondents consider their control systems to be sufficiently scalable to safely support the increasing use of autonomous AI agents. In Germany, the figure is 38 per cent.

Consequently, the requirements regarding access rights, logging, monitoring and the ability to immediately interrupt critical system activities are increasing.

Companies are introducing AI despite a lack of control mechanisms

The study also shows that many companies are pressing ahead with their AI projects despite incomplete governance structures.

56 per cent of the executives surveyed state that they are introducing new AI applications within agreed risk limits whilst continuing to develop the necessary control mechanisms in parallel. In Germany, this figure is as high as 64 per cent.

In contrast, 33 per cent of companies in the countries surveyed are postponing implementation until appropriate governance structures are in place. In Germany, the figure is just 23 per cent.

This creates a tension for information security: whilst AI systems are increasingly accessing internal data, applications and business processes, security managers must ensure that access rights are restricted, decisions are traceable and malfunctions are detected in good time.

Third-party AI remains a weak point

The survey reveals particular shortcomings in AI applications from external providers.

54 per cent of respondents report governance gaps or restrictions when using AI systems within the supply chain. In Germany, this figure stands at 47 per cent.

Nevertheless, only 21 per cent of companies continuously monitor the risks associated with third-party AI. In Germany, the figure is as low as 18 per cent. Only 13 per cent of companies require independent audits or certifications; in Germany, the figure is 11 per cent.

This means that securing external AI components is becoming increasingly important. Companies must not only monitor their own systems, but also be able to track which data and applications are accessed by external AI services and identify any security-related changes made by the providers.

Who is liable if an AI agent causes damage?

Although, according to the respondents, responsibility for AI risks is formally enshrined at executive board level in 98 per cent of companies, uncertainties remain regarding the specific allocation of responsibility.

Particularly in the case of autonomous AI agents, the question arises as to who is responsible for erroneous decisions, unauthorised system access or any resulting damage.

Respondents cite contractual liability provisions, defined thresholds for human intervention, real-time monitoring and traceable audit logs as key requirements.

The ability to immediately suspend agent activities is also gaining in importance. Particularly in the case of AI systems with extensive access rights, such a function can be crucial in limiting the impact of erroneous or manipulated actions.

EU AI Act increases pressure on governance

In addition to operational security risks, regulatory requirements are influencing investment in AI governance.

According to IBM, 41 per cent of executives surveyed in the European Union are accelerating their governance investments in light of the requirements of the EU AI Act. In Germany, the figure stands at 50 per cent.

However, only 37 per cent of the companies surveyed have a formal quantification of the financial risks arising from AI governance gaps, reviewed by the board. In Germany, the figure is 38 per cent.

As a result, there is often a lack of a robust basis for systematically assessing potential damage and prioritising investment in protective measures.

AI governance is becoming a corporate security responsibility

The IBM study makes it clear that AI governance is increasingly moving beyond traditional compliance issues.

With the deployment of autonomous AI agents, companies must consider technical authorisations, operational risks and organisational responsibilities as a whole. Information security, risk management, data protection and corporate governance are required to develop binding control and escalation procedures.

The high number of AI incidents at senior management level shows that these risks have already become a reality in business practice. The fact that comparatively few companies subsequently adapt their governance underscores the need for action.

For security managers, one question therefore becomes central: not only what AI agents are capable of, but also what decisions they are permitted to make independently – and how companies can intervene if something goes wrong.

Methodology: On behalf of IBM, Censuswide surveyed a total of 3,221 executives from companies with at least 1,000 employees between 14 and 24 September 2026. The survey covered Germany, France, Italy, Spain, the Netherlands, the United Kingdom, Saudi Arabia and the United Arab Emirates.

Source: IBM, press release dated 15 October 2026.