The year of AI agents

 

"Generative artificial intelligence has not yet led to the hoped-for productivity increases. So-called AI agents are now expected to bring about a breakthrough as real all-rounders.

 

In a broader sense, says Conrad Caine, he runs a mechanical engineering company. Only it is not about classic manufacturing technology, but about software robots. Caine is the founder of the Munich-based software company Machines Like Me with almost 80 employees.

"Our AI agents imitate what people do on the computer every day," he says. They compare orders with offers, analyze rental contracts, prepare reports and take on tasks for municipal utilities such as receiving meter readings over the phone. "We automate routines," says Caine. "It's mainly about repetitive tasks that usually do not add any value and that many employees don't feel like doing anyway." Because: "Nobody was born to maintain Excel lists."

 

What Caine describes so figuratively as software robots is commonly known as an AI agent - and also one of the key topics at the DLD (Digital Life Design) tech conference in Munich. "AI agents will integrate artificial intelligence much more strongly into our everyday lives," said German AI pioneer Björn Ommer during a lecture there. Ommer heads the chair for computer vision and learning at the Ludwig Maximilian University in Munich (LMU), where the AI ​​image generator Stable Diffusion was developed, among other things. Artificial intelligence understands even better what users need from it. AI agents would thus become real companions instead of AI applications such as ChatGPT, which users used more occasionally. Sam Altman, head of ChatGPT developer Open AI, recently declared 2025 to be the "year of AI agents". His chief developer Colin Jarvis said at the DLD that up to now it had always been about asking AI models questions.

In the future it will be about asking them to complete tasks. This requires the necessary reliability of the models and transparency about the basis on which the AI ​​makes its decisions.

 

The technology is intended to fulfill the promise that it made to many companies when ChatGPT was released: to significantly increase productivity in companies. Many companies are experimenting with the technology. But despite promising prototypes, scaling it up across the company has proven to be complicated. Many office employees are now profitably using the company's internal AI chatbot as a personal assistant, but major leaps in productivity are rare.

 

According to a recent survey by the management consultancy BCG, only a quarter of German companies report significant benefits from their previous AI initiatives.

 

AI agents are intended to change that. What has been common practice in the mechanical world for decades is now also to be achieved in the office by artificial intelligence: the automation of processes.

 

AI agents should not only be able to create and process texts, images, videos or other data on command like previous generative AI, but should also be able to carry out entire chains of tasks autonomously. They collect and process company data, draw their own conclusions from it and initiate the necessary actions themselves. The AI ​​assistants thus become AI colleagues who work independently. This could increase productivity by a factor of three, and in some cases even by a factor of five, believe the management consultants at BCG. Market researcher Gartner predicts that in three years a third of all applications in companies will contain AI agents. Last year the proportion was less than one percent.

 

Such figures are also known to major technology providers such as Microsoft, which itself offers AI agents for certain tasks. "In the future, managers will not only manage human teams, but also AI agent teams," says Jared Spataro, head of marketing for Microsoft's AI business with companies. He describes AI agents as the "apps of the AI ​​era."

 

Spataro's personal favorite is a sales AI agent that reads out interested potential customers from inquiries via social media or emails and contacts them independently. The AI ​​agent is supposed to find out whether the person could actually become a customer and what exactly they are looking for, in order to then hand them over to a human employee. Large companies today employ entire teams of people for this early phase of customer acquisition.

 

All of this is primarily intended to save resources. His software machines are ten times cheaper than the human alternative, promises Conrad Caine of Machines Like Me. His team looks at the processes to be automated and breaks them down into their individual parts - a kind of modern form of Taylorism, says Caine, referring to the American management consultant Frederick Taylor, who in 1911 developed a concept for refining and standardizing work processes. The individual subtasks are then carried out by dozens of specially adapted AI software modules, such as those that specialize in reading unstructured documents. These modules are orchestrated by another AI. His company produces digital employees and trains them in the company. Machines Like Me does not automate the processes 100 percent; there is always a human involved. They continue to take on a small part of the activities - for example final checks, approvals or discretionary decisions. Caine is convinced that more than a third of all administrative activities in Europe can be automated.

 

This raises questions, especially about jobs. BCG's current survey of over 1,800 companies around the world concludes that a good two-thirds of decision-makers see AI and people as complementary. Only 7 percent expect to cut jobs because of AI.

 

But the truth is also that if entire fields of activity are taken over by AI agents - for example in customer service - then employees have to be retrained on a large scale.

 

First of all, companies have to use the technology in the first place. In Germany, 30 percent of the managers surveyed by BCG believe that AI agents will be central or at least complementary to their business in the future; in the United States, the proportion is higher at 37 percent." [1]

 

1.  Das Jahr der KI-Agenten. Frankfurter Allgemeine Zeitung; Frankfurt. 20 Jan 2025: 26.   Von Maximilian Sachse, München