Half of companies lack control over their AI agents

Investment in AI agents is expected to quadruple in two years, but only 12 percent of companies believe they are fully ready to manage the technology, according to a global survey. The AI Value Report 2026 where SAP and Oxford Economics asked 2,600 business leaders in 13 countries about their AI investments.

The average company in the survey expects a 21 percent return on its AI investments this year, equivalent to $6.3 million (approximately SEK 60 million). This is an increase from 16 percent last year.

Expectations for agent-based AI – where AI systems plan and perform tasks themselves within given frameworks – have increased significantly. Companies are investing an average of SEK 265 million in the technology, but only 3 percent believe they are fully prepared for the technology.

Steering and control lag behind

The report also shows that many organizations lack basic mechanisms for governance and control of AI.

– Oxford Economics notes that only 12 percent of companies have full control over their AI. Less than half have a register of the AI agents used in the organization. Imagine if the same were true for employees or external consultants, then you wouldn't sleep very well at night, says David Pontoppidan, Head of Business AI Nordic & Baltic at SAP.

“In addition, half lack logs of agent decisions, and over 40 percent have no access or permission controls for their agents. Perhaps that is why almost half have already experienced data breaches as a result of unauthorized use of AI. These are very serious figures,” he continues.

Few companies are ready for AI agents

David sees no clear difference between the Nordic region, which is often highlighted as a digital pioneer region, and the rest of Europe when it comes to the ability to integrate AI into business processes.

According to him, the growing expectations are due, among other things, to the fact that the AI debate focuses too much on models rather than on the foundation required to create real value.

– The fascination with AI models has become unhealthy and it is serious. The public discussion revolves around benchmark tests and launches of the latest models that Claude Fable 5 and GPT-5.6 Sol, while 73 percent of companies say data quality is the most important reason they are getting less value from AI than they should. Additionally, 79 percent say they regularly receive low-quality AI-generated content that creates additional work, says David Pontoppidan.

In his role as AI Manager for SAP in the Nordics and Baltics, he works daily with some of the largest and most complex organizations to increase the value of their investments in enterprise-driven AI. He sees a huge gap between demonstrations in controlled environments and how companies are actually prepared to drive AI in an industrial way.

“Two companies can use agent-based workflows that are based on the same AI model and still get completely different results. It has very little to do with how skilled the developers are and much more to do with how the companies approach the task. In my opinion, it is a strategic discussion that needs to start in the management team. Where should business value be created, how should it be measured and which decisions should we never leave to an agent? The choice of AI model comes much later,” he says.

Expert advice – models can be rented, but the foundation must be owned

According to David Pontoppidan, in order for companies to be able to derive clear value from their AI agents, the technology must be connected to the processes, data and control mechanisms that already drive the business.

“My best advice is to let agents move in where the business already works, in systems that know the company’s data, processes, roles and controls. That way, agents learn the basics from day one and resources can be spent on what only humans can do – shaping where judgment should lie in the processes,” he says.

If the ambition is rather to avoid falling behind, the advice is simple. Don't build a parallel AI landscape alongside the business, but integrate the intelligence into the platforms that already drive core processes. Then records, permissions and audit trails will follow automatically.

In most companies, the foundation is already in place, but the intelligence has not yet been connected. That is really the report's conclusion in a nutshell – models can be rented, but the foundation must be owned.

Read the full report here: SAP Value of AI: Oxford Economics 2026