Caios is working so that AI agents are properly implemented

While they supervise the experimentation and deployments of AI agents and guide other leaders on the trip, Caios also sails through media threshing, concerns concerning security and confidence, and interconnectivity (or its absence) of these systems. Not to mention having to struggle with the question: what even East An AI agent?

The quarrel in media threw causes companies to concentrate

No one may agree on what exactly means the term “AI agent” Fortune And others reported. Companies define the term differently and often use it to describe varied features and capacities, including many that have been described above with other terms such as “AI assistants”. For the AI ​​chief officer of ACCENTURE, Lan Guan, who directed the construction of an AI agent solution called Rafinery IA for customers and also works directly with them on their own deployments of AI and AI agent, this led him to devote a lot of time to help customers sorting contradictions.

“A year ago, everyone said:” I have to do the AI ​​generation “. Now everyone says: “I need to make agents of origin AI or IA. And it’s like, at the end of the day, many of these things are always the same thing. “And so there is a ton of confusion on the market with our customers on:” What is an AI agent? ” What am I deploying? And so we spend a lot of time in education. »»

An furious effect of this was companies that quickly run the so -called AI agents “just for the press release,” said Michelle Bonat, AI Squared AI head, who also works with companies in regulated industries on their AI development. The pressure to have an answer at the moment of the agentic AI means that certain companies rename the functionality or continue the AI ​​agents to stay in fashion, often simply creating fine layers of agents above the foundation models.

“I totally see that. I see this every day, ”says Bonat. “This is why this space is full of noise.”

Security, errors and confidence dominate the risk analysis

Despite media threshing and confused terminology, the main idea of ​​AI agents – systems designed to act independently to perform specific tasks – always generates a lot of justifiable excitement. It is also essential to create the types of systems technologists and science fiction enthusiasts have always dreamed, capable of performing sequences of complex tasks on several platforms in our name. But there are real roadblocks.

Uri Yerushalmi, co -founder and head of AI at Fetcherr, who uses AI for predictive prices in the air transport industry, believes that the opportunities concerning AI agents are “enormous”, but that the unlocking of this value depends on raising the real challenges around confidence and integration and avoiding failure. For example, agents must integrate into inherited systems and align with the constraints of the real world without disturbing existing workflows. And while we give more autonomy agents, we must build railing, surveillance, replacement systems and human alignment mechanisms, he said.

“Companies must trust the agent’s decisions,” he added. “This requires transparency, consistency and demonstrable return on investment.”

One of the most worrying failures is composition errors. The CEO of Google Deepmind Demis Hassabis compared this problem to the interest composed for finances, explaining that even if an agent model has only 1%error, this would lead to a chain reaction of errors which, after a few thousand steps, would ultimately make the probability of a completely “random” result. Bonat underlines this problem of composition of errors as a serious challenge in terms of confidence of AI agents, saying that this potential to worsen a false step without humans being even aware of this could “create ravages”.

This is particularly true for the type of multi-agent systems that many companies are considering, which, according to Guan, can cause dead angles and cause you problems very quickly.

“This may not work for you and can actually present a lot of risks,” she said. “Think about it – a lot of workflows and transactions or commercial interactions are high issues. You do not want the agents to make only a refund for each customer, right? ” She said, adding that even if her customers have a strong appetite to see the impact of AI agents, they also be wary of high cloud bills and security risks.

Safety problems are certainly in the lead in the landscape of the AI ​​agent. By 2028, Gartner predicts, 25% of corporate violations will be traced to AI agents, including the abuses of internal and external malicious actors. The dominant factor contributing to security risks is the combination of autonomy and interoperability of agent systems, which would make them connect, exchange data with and act independently on a wide strip of platforms and systems. In other words, the exact nature of the functioning of these systems and what they are intended to do is what makes them so risky.

Dreams of interoperability have trouble freeing itself from closed gardens

Like all Caios, Ali Alkhafaji, AI chief and technology manager of Omnicom Precision Marketing Group, is concerned with data leak and other security risks. He is also concerned about the fact that many companies marketed agent systems use security as a practical excuse to lock their customers more in their ecosystems, going against collaborative and decentralized vision that many consider as intrinsic for an aged future: “Not because it cannot be resolved, but because it is not in the supplier of the supplier to resolve it.”

“Each supplier builds its own` `agent framework ”, but no one solves for interoperability at the level of the company. Without open executives and semantic standards, we simply build smarter silos,” he said, adding that the protocols of collaboration of agents remain immature and that it is frustrating to see suppliers and hyperscalers of agents and reinforcement gardens.

Deloitte US, the head of Ai Jim Rowan sees this game among its customers, noting that they are mainly going to their current suppliers and use their agent capacities when published. This is another iteration of the advantage of the platform that stimulates growth for suppliers like Openai, Google and Microsoft as they embark their current customers in their new pipelines and AI products.

“There is a final tension on the market between who wants to have the registration system for agents. As, who will owner the register, which will orchestrate all the orchestration around the agents,” said Rowan. “We see this with hyperscalers and SaaS suppliers and third tools that are also in space. I think the jury is still on whom has this.”

Correction, September 3, 2025: An earlier version of this story won the name of the AI ​​agent solution of Accenture.

Global Forum fortune returns on October 26 to 27, 2025 in Riyadh. CEOs and world leaders will meet for a dynamic event only invitation that shapes the future of business. Request an invitation.