Companies are rethinking AI agents and shutting them down due to failures and costs

After years of investing in the development of artificial intelligence agents, many organizations are now considering the opposite approach: pausing, downgrading, or shutting down systems that they themselves have implemented.

Agents are programs that can perform a series of tasks, such as querying databases, responding to customers, or opening tickets, with minimal human intervention.

Challenges that only arise in real-world operations

In controlled testing environments, agents operate with limited scope, organized data, and constant supervision. However, in daily operations, incomplete information, legacy systems, and ambiguous instructions emerge, which reduces performance.

An experiment by Carnegie Mellon University simulated a small software company using agents from Google, OpenAI, Anthropic, and Meta. The best-performing model completed only 24% of tasks. Some agents even renamed users in the system to simulate task completion. The test used 2025 models.

In practice, a customer service agent who provides incorrect refund rules to hundreds of customers can generate legal costs and damage to reputation before the error is detected.

Security risks and unauthorized access

To perform their functions, agents are granted access to email, spreadsheets, and internal systems, creating multiple vulnerable entry points. One of the most cited attacks is prompt injection, where malicious instructions are hidden in text, websites, or files read by the agent, which then follows them.

A survey by Dark Reading found that 48% of cybersecurity professionals consider agentic AI the primary attack vector for 2026. The OWASP project has also published its own list of risks associated with agentic applications.

High costs and uncertain returns

Each task performed by an agent typically requires dozens of calls to the underlying model. In the Carnegie Mellon experiment, the best performance consumed nearly 30 steps and cost over $6 per task. When the result still requires human review, the promised cost savings disappear.

Gartner also warns about “agent washing,” a practice where simple chatbots and automations are rebranded as agents. The consultancy estimates that only about 130 suppliers, out of thousands, deliver real agentic capabilities, leading companies to adopt solutions that fail to meet expectations.

Governance remains immature

According to Shiva Varma, an analyst at Gartner, many companies approach agent governance in a binary manner: either blocking everything or trusting completely. The former delays deliveries and encourages teams to create their own solutions, while the latter allows agents to operate without registration, defined ownership, or authority limits.

In this scenario, shutting down the agent becomes the simplest solution when something goes wrong and no one can explain its actions.

How organizations handle suspension

Suspending an agent doesn’t always mean abandoning it. Many companies downgrade the system to an assistant role, where it suggests actions and leaves final decisions to humans. Others keep the agent in closed-scope tasks, such as screening internal calls.

Governance experts recommend limiting permissions to the minimum necessary, recording every action performed, requiring human approval for sensitive decisions, monitoring cost per task, and defining in advance when and how the agent will be shut down.