The excitement around “agentic AI”—AI that can act on its own rather than just answering questions—is hitting a wall of reality. According to a recent report from Camunda, there is a massive disconnect between what companies want to do with AI and what they are actually achieving. While over 70% of organizations say they are using AI agents, only 11% of those projects successfully moved into a production environment last year. It turns out that giving a piece of software the keys to your business processes is a lot harder than just setting up a chatbot.
The main reason for this stall is a fundamental lack of trust. Businesses are worried about the risks of letting an autonomous agent handle tasks without a human constantly watching over its shoulder. Specifically, 84% of organizations cited business risks as a top hurdle, while 80% pointed to a lack of transparency in how these agents make decisions. Because many AI systems act like a “black box,” it is difficult for managers to explain why an agent took a certain action, which creates a nightmare for regulatory compliance. As a result, many companies are keeping their agents in “silos,” limiting them to basic assistant roles instead of letting them automate complex workflows.
This caution is leading to what some experts call an “appallingly low” return on investment. Organizations are spending millions on the technology but are too afraid to let it do the work it was designed for. When an AI agent still requires a human to click “approve” at every step, the promised gains in speed and efficiency disappear. The report suggests that until companies can build a foundation of governed processes—where they can see, track, and control exactly what an agent is doing—most of these projects will remain stuck in the pilot phase.
For those who have managed to bridge this trust gap, the results are promising. About 95% of companies that fully deployed agentic AI reported measurable business growth. The takeaway is that the technology works, but the infrastructure to manage it hasn’t caught up. If you are looking to bring AI agents into your own workflow, the focus should probably be on “boringly narrow” tasks with clear guardrails rather than trying to build a digital employee that can do everything at once.
