The belief of a non -verified AI agent is like handing over their keys to a drunken graduate

AI agents are now embedded in the core business functions worldwide. Soon these agents could plan our lives, make important decisions and negotiate business on our name. The view is exciting and ambitious, but also asks the question: Who actually supervises it?

Over half (51%) of the companies have deployed AI agentsAnd Salesforce CEO Marc Benioff have taken a billion agents by the end of the year. Despite its growing influence, the verification test is particularly lacking. These agents are commissioned with critical responsibility in sensitive sectors such as banks and health care without proper supervision.

AI agents need clear programming, high-quality training and real-time knowledge in order to carry out goal-oriented actions efficiently and precisely. However, not all agents are created equally. Some agents may receive more advanced data and training, which leads to an imbalance between tailor -made, well -trained active ingredients and mass producers.

This could be a systemic risk if more advanced agents manipulate and deceive less advanced agents. Over time, this gap between agents could lead to a gap in the results. Let us assume that an agent has more experience in legal processes and uses this knowledge to use or exceed another agent with less understanding. The use of AI agents by companies is inevitable, as is the development of new power structures and manipulation risks. The underlying models are the same for all users, but this possibility of deviation must be monitored.

In contrast to conventional software, AI agents work self -containing, complex settings. Her adaptability makes it powerful, but also more susceptible to unexpected and potentially catastrophic mistakes.

For example, a AI representative could diagnose a critical illness in a child incorrectly, since it was mainly trained in adult patients. Or an AI agent -chat bot could escalate a harmless customer complaint because it misinterprets sarcasm as an aggression and slowly loses customers and income due to misinterpretations.

According to industry research, 80% of the companies have announced that their AI agents have made “villain” decisions. Orientation and security problems can already be seen in examples in the real world, such as: B. autonomous agents who exceed clear instructions and delete important work.

As a rule, the employee must be exposed to the HR department if a big human error occurs and a formal examination can be carried out. These guard rails are not available for AI agents. We give them to sensitive materials on a human level without being almost human supervision at the human level.

So do we make our systems through the use of AI agents or do we give up the agency before the right protocols are available?

The truth is that these agents may learn quickly and adapt them according to their respective environments, but they are not yet responsible adults. For years they have not experienced years of learning, experimental and failure and interact with other business people. They lack the maturity that was acquired from the experience lived. If you give you autonomy with minimal checks, it is like the handover of the company keys to a drunk graduate. They are enthusiastic, intelligent and formable, but also unpredictable and need an supervision.

And yet what large companies do not recognize is that this is exactly what they are doing. AI agents are “seamlessly” connected to the operations with little more than a demo and a disclaimer. No continuous and standardized tests. Not a clear exit strategy when something goes wrong.

What is missing is a structured, multi-layered verification framework that regularly tests the agent behavior in simulations of real and high scenarios. If the adoption accelerates, the review becomes a prerequisite to ensure that AI agents are suitable for purposes.

Various review degrees are required according to the complexity of the agent. Simple knowledge of knowledge or those who are trained for the use of tools such as Excel or E -Mail may not require the same strict test as demanding active ingredients that replicate a variety of tasks that perform people. However, we must be appropriate guidelines, in particular in demanding environments in which agents work in cooperation with humans and other agents.

When agents start making decisions on a scale, the edge shrinks quickly for errors. If the AI ​​agents we have checked, critical operations are not tested for integrity, accuracy and security, we risk enabling AI agents to devastate society. The consequences will be very real – and the costs for the damage control could be amazing.