According to Raimo Seero, Chief Technology Officer at Uptime, the winners will be companies that implement AI with a conscious approach to risk – thinking in advance about how to limit potential damage and recover from mistakes. In Seero’s view, avoiding AI altogether may leave a company at a competitive disadvantage in terms of decision-making speed and productivity. However, overly careless adoption can also result in a range of negative consequences.
“AI adoption is progressing faster in companies than AI risk management. Just a few years ago, AI was mainly used in isolated pilot projects. Today, artificial intelligence is increasingly being integrated into business-critical systems. While this creates new opportunities to improve operational efficiency, the pursuit of better results often means that risks are not fully considered,” he explained.
AI: A Highly Capable but Unreliable Employee
According to Seero, managing AI-related risks is essentially traditional engineering thinking applied in a new context. “In aviation, manufacturing, banking and software architecture, the same questions have always been asked: what happens if a component fails? How significant is the potential damage, and how can its impact be contained? The same questions must be asked about AI agents,” he said.
The key difference is that an AI agent is a goal-oriented system. “It does not inherently consider the harm it may cause while pursuing an objective. If it is given overly broad access permissions, it may achieve a result in a way that its human operators did not intend. That is why safeguards must be built into the system from the outset,” Seero added.
An AI agent should be treated like a highly capable employee, but not one who is granted unlimited authority. “You would not give a new employee access to every system on their first day, permission to make changes in the production environment, or the ability to make binding commitments to customers. The same principle applies to AI, which can act very quickly and at scale,” he said.
“If AI helps prepare a proposal, it does not need to send it directly to the customer. If it analyses customer-support conversations, it does not need permission to change billing records. For high-impact actions, a human should remain the decision-maker,” he added.
In addition to access controls, organisations should establish logging that makes it possible to review where AI made an error and which data may have been affected. It is also important to set spending limits and define a process for stopping the AI if something goes wrong. “When a company understands the potential scale of damage and the boundaries of its systems, it can use AI more confidently and gain greater value from it. Without clear limits, even a small mistake can quickly escalate into a major problem,” Seero said.
When Something Goes Wrong
Another essential part of AI risk management is assessing the potential impact of errors and preparing a recovery plan. Seero recommends that organisations consider in advance how an AI-related error can be detected quickly and how its actions can be reversed. This may include rapidly restricting the AI’s permissions, notifying customers, or restoring data from change logs or backups.
It is important that such a plan is not created only once a crisis has already begun. “If AI has access to important systems, it is not reasonable to start planning your response only after something goes wrong. By then, the damage may already have been done,” he said.
However, AI-related risks are not a reason to avoid the technology altogether. “Rather, they are a reason to adopt AI with the same discipline used to manage financial, information-security, or operational risks. By breaking risks down into manageable components and deploying AI where it makes sense, organisations can use informed risk-taking to reduce risks overall and benefit from the technology,” Seero concluded.
