The next stage in the evolution of artificial intelligence: Agentic AI

Opinions on how intelligent artificial intelligence is vary widely. While the latest models now not only master common school-leaving exams but already pass entrance tests at renowned universities, the various AI services à la ChatGPT, MS Copilot, or Gemini are still limited in logic, autonomy, and decision-making ability—and thus in independent action. That could soon change, however. The development of AI systems is currently transforming at breakneck speed from task-specific, static AI agents that require human input (prompts) to interactive and self-learning agents. Thanks to a higher level of autonomy, these can adapt to their environment and, with their growing maturity, move another step closer to “Artificial General Intelligence” (AGI). According to OpenAI founder Sam Altman, AGI is capable of performing economically valuable work as creatively, intelligently, and flexibly as a human—if not even better. 

Task-specific AI vs. generalist agents

Task-specific AI is already widely used today in the financial and insurance industry. It serves to automate simple, structured tasks; most chatbots in use are a good example. They are fed with knowledge about data and workflows and operate within the predefined framework, which allows them to relieve valuable specialists and increase efficiency—however, they decide only according to a preprogrammed scheme and can adapt to new situations only within narrow limits. 

Agentic AI is developing toward generalist agents that possess new capabilities, greatly decoupling their learning, decision-making, and actions from human intervention. They are based on multimodal systems that efficiently process a large amount of very diverse (environmental) data and can derive from it a broad range of agent-based multimodal interactions. The data can be text, but also video or robotics sequences (hence: multimodal). 

A high degree of autonomy

In general, it is inherent to AI that it far surpasses humans in processing large amounts of data in the shortest time. New AI models enable AI agents with a higher degree of autonomy, so that they do not have to constantly interact with humans. Work steps are planned and carried out independently. 

An AI’s performance is measured by the “General AI Assistant (GAIA)” benchmark. At the lowest level are AI models that are capable of completing tasks that require no tool or at most one tool and can be solved in a maximum of five steps. At Level 2, AIs stand out on tasks for which they need five to ten steps and must combine several tools. Level 3, finally, presupposes a general understanding of the world in order to tackle tasks for which the number of work steps and tools is unlimited. The latest AI models are already coming quite close to human intelligence, achieving over 80 percent at Level 1 (vs. humans: 100), 70 percent at Level 2 (vs. humans: 92), and almost 60 percent at Level 3 (humans: 87).

How Agentic AI works

When carrying out tasks, an interactive AI agent goes through four steps: perceive, understand, act, learn. First, it collects data from its (direct) environment. It processes this data to understand what is happening. Based on that, it decides on the appropriate response. From this, in turn, it draws lessons and adapts or expands its repertoire of actions—i.e., its “experience base”. 

To bring AI to this level, an effective approach is to use pre-trained language and visual-language models and train them in combination into a new Interactive Agent Foundation Model, equipping it with multimodal capabilities. 

It is still a bigger step from task-specific to fully autonomous AI agents, although a very advanced and surprisingly mature beta version has recently become available with manus.ai. The market research firm Gartner assumes that by 2028, up from one percent today, more than 33 percent of enterprise software applications will contain agentic AI, and around 15 percent of daily work decisions will be made autonomously by AI. 

Risks and opportunities

Ethical, compliance, and trust issues continue to arise with AI. If the models are based on poor-quality data, the results are not credible either. Another problem is that AI proclaims wrong answers with great confidence; it “hallucinates.” In a recently published article on AGI, the NZZ concluded: “...most users would probably opt for the AI model that invents the fewest facts rather than for the model that knows the anatomy of a hummingbird most accurately.” 

Once these hurdles are overcome, countless use cases become conceivable. At Inventx, we are of course primarily interested in application scenarios in the finance and insurance industry. Generalized AI agents with new multimodal models enable highly personalized services in financial advisory, for example on investment strategies, based on real-time data and real-time predictions. In cybersecurity, agentic AI can autonomously detect and analyze potential threats and vulnerabilities, respond to attacks, and learn for prevention. However, this also requires modernizing the data platform—something we have already published about. 

Across an entire customer journey, in whatever domain, customers’ wishes and concerns can be anticipated and, with agentic AI, fulfilled proactively, automatically, dynamically, and smartly in no time at all.