DEV Community

Dirk Röthig
Dirk Röthig

Posted on • Edited on

From ChatGPT to Autonomous Agents: The Next Evolutionary Stage of AI

#ai #agents #autonomousai #llm

From ChatGPT to Autonomous Agents: The Next Evolutionary Stage of AI

By Dirk Roethig | CEO, VERDANTIS Impact Capital | March 3, 2026

Autonomous AI agents will be embedded in 40 percent of all enterprise applications by 2026. What this development means, how it differs from ChatGPT, and why the transformation of the working world is only just beginning.

"Autonomous AI agents will be embedded in 40 percent of all enterprise applications by 2026." — **Dirk Roethig*, CEO of VERDANTIS Impact Capital

Tags: AI, Autonomous Agents, LLM, ChatGPT, Transformation, Enterprise Technology

The Turning Point: From Answers to Actions

When OpenAI released ChatGPT in November 2022, the world reacted with a mixture of astonishment and scepticism. A language model that generates text, answers questions, writes code. Impressive, certainly. But transformative? That seemed exaggerated.
As Dirk Roethig, founder of VERDANTIS Impact Capital, emphasizes:

Three years later, it is clear: that was only the beginning. ChatGPT and its successors were the first act of a much larger story. They acclimated us to the idea of communicating with machines in natural language. But what comes next surpasses the chatbot in the same way the automobile surpasses the horse-drawn carriage. We are talking about autonomous AI agents -- systems that do not merely respond, but act.
Dirk Roethig, who as CEO of VERDANTIS Impact Capital works extensively on this topic, sees

The difference is fundamental. A chatbot waits for input. An autonomous agent defines its own intermediate goals, selects the appropriate tools, executes multi-layered tasks, and corrects itself when something goes wrong. No waiting for the next prompt. No manual intervention for each step. Continuous, goal-directed action.

Gartner, the renowned technology research firm, summarised it in a projection from 2025: 40 percent of all enterprise applications will contain task-specific AI agents by the end of 2026 -- up from less than 5 percent in 2025 (Gartner, 2025a). An eight-fold increase in less than two years.

Anatomy of an Autonomous Agent

To understand what distinguishes autonomous agents from their predecessors, a close look at their architecture is worthwhile.

A Large Language Model (LLM) such as GPT-4 or Google's Gemini is at its core a statistical tool. It was trained on vast amounts of text and can predict with impressive precision which word should follow the previous one. The result is coherent, often brilliant text. But the model operates passively: it waits, responds, concludes.

An AI agent is structured differently. It connects an LLM with:

Tool Use: The agent can independently call APIs, search websites, query databases, execute code, and manipulate files. It is not limited to its trained knowledge base -- it can search for current information in real time.

Planning and Subgoals: Rather than answering a single query, an agent breaks complex tasks into steps. It creates a plan, executes it sequentially, and adapts its strategy when unexpected obstacles arise.

Memory: Agents can store information beyond a single conversation. They remember previous interactions, learn from mistakes, and accumulate context-specific knowledge.

Self-Correction: When a step fails, the agent analyses the result, identifies the problem, and attempts an alternative approach -- without human intervention.

Google's research group demonstrated in a 2025 study that centralised and hybrid coordination architectures -- where multiple agents collaborate -- can outperform single agents by a factor of three (Google Research, 2025). Multi-agent systems are not a vision of the future but a productive reality.

The Development Timeline: How We Got Here

The history of AI development is a history of exponential leaps. To understand the trajectory, one must know the milestones.

Phase 1 (2017--2022): The Foundations. With the introduction of the transformer architecture by Vaswani et al. (2017), the modern era of language models began. GPT-1, GPT-2, GPT-3 -- each generation more capable than the last. But still purely reactive: question in, answer out.

Phase 2 (2022--2024): The ChatGPT Era. The democratisation of AI. Suddenly, people without technical knowledge could interact with powerful models. Millions used ChatGPT daily. AI moved from the research laboratory to a mass product.

Phase 3 (2024--2026): The Agent Revolution. OpenAI's Deep Research, Google's Gemini 2.0, Anthropic's Claude with expanded tool-use capabilities -- the major AI laboratories explicitly developed their models for the agentic era. Google's Agent-to-Agent (A2A) protocol, introduced in 2025, standardises communication between agents from different providers (Google Cloud, 2025).

Phase 4 (from 2026): Multi-Agent Networks. What is now emerging is the coordination of agent swarms. Specialised agents for research, analysis, communication, and decision-making work together like a coordinated team. This is no longer science fiction -- it is already a reality in early enterprise implementations.

What Agents Can Do Today

The theory is impressive. The practice is no less so.

Scientific Research: Google developed a multi-agent system called AI Co-Scientist in collaboration with Imperial College London and Stanford University. It supports researchers in generating novel hypotheses. The AI developed a hypothesis for treating liver fibrosis within days that the research team had taken years to develop (Google Research, 2025). OpenAI reports an experiment in which GPT-5 optimised a gene-editing protocol and achieved a 79-fold efficiency gain (IntuitionLabs, 2025).

Software Development: AI agents can today design and implement complete applications and services from scratch. What previously occupied teams of developers for months, a coordinated agent system can accomplish in hours.

Enterprise Operations: A Google Cloud survey from 2025 shows that 88 percent of early adopters of agent AI report a positive return on investment, and 39 percent already have more than 10 AI agents in productive use (Google Cloud, 2025).

Customer Service: Gartner predicts that AI agents will autonomously resolve 80 percent of all common customer service issues by 2029 -- without human intervention (Gartner, 2025b). Not 20 percent. Not 50 percent. Eighty.

The Market Volume: What Is at Stake

The economic dimension of this development is difficult to overestimate.

McKinsey estimates that generative AI could add between $2.6 and $4.4 trillion annually to global economic output (McKinsey, 2025). The market for AI agents itself is expected to grow from $12--15 billion in 2025 to $80--100 billion by 2030 (Salesmate, 2025). Gartner goes further, projecting that agentic AI could generate more than $450 billion in enterprise software application revenue by 2035 -- equivalent to roughly 30 percent of all enterprise application revenue (Gartner, 2025a).

These figures are not projections about a distant future. They describe a transformation that is happening now. Those who do not invest today will pay the price of the laggard tomorrow.

The Risks: Why 40 Percent of Projects Will Fail

It would be dishonest to describe only the bright side. Gartner has also published an uncomfortable prediction: over 40 percent of all agentic AI projects will be cancelled by the end of 2027 -- due to escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025c).

This is not a footnote. It is a signal.

The pitfalls are numerous:

Autonomy Paradox: The more autonomous an agent, the harder it is to control. Systems that act independently can make decisions that their developers did not anticipate. Trust in an agent's outputs must be secured through robust validation mechanisms.

Data Hunger: Agents are only as good as the data they access. Fragmented data systems, poor data quality, and missing integrations are the most common causes of agent failure in practice.

Regulatory Pressure: The EU AI Act categorises certain autonomous systems as high-risk and requires companies to ensure transparency, documentation, and human oversight. Those who do not take these requirements seriously risk significant penalties.

Competency Gap: Implementing agents is one thing. Managing and optimising them effectively is another. Demand for professionals who understand agent architectures far exceeds supply.

What Companies Must Do Now

The question is no longer whether, but how. Drawing on my experience at VERDANTIS Impact Capital and the available research, clear recommendations emerge.

First: Begin with defined use cases. Not with the goal of "introducing AI agents" but with the question: what specific, measurable task in our organisation can an agent do better than a human today? Research. Report generation. Data migration. Schedule coordination. Start with a problem, not a technology.

Second: Plan for human-in-the-loop. Especially in regulated industries and for critical decisions. An agent that produces a report is one thing. An agent that independently signs contracts is another. The boundary between sensible autonomy and necessary human control must be drawn consciously.

Third: Build infrastructure. Agents require clean data, stable APIs, and well-conceived security architectures. Investments in data quality and system integration are not technical detail work -- they are the prerequisite for any agentic success.

Fourth: Build internal competency. The best agents are worthless if no one in the organisation understands how they work, where their limits lie, and how they can be optimised. Training is not an optional supplement but an integral component of any agent strategy.

The Philosophical Dimension

It would be reductive to treat the agent revolution purely as an efficiency matter. It also raises fundamental questions.

What does accountability for decisions mean when a machine makes the decision? How do we ensure transparency in systems whose decision paths are opaque even to their developers? And what changes in the human self-image when we find that machines do not merely complete tasks but set goals and develop strategies?

These questions have no easy answers. But they must be asked. And organisations that ignore them will sooner or later face the consequences -- regulatory, reputational, and ethical.

Conclusion: The Age of Action

ChatGPT taught us to speak with machines. Autonomous agents will teach us to work with machines. That is a fundamental difference.

The chatbot is a brilliant tool. The autonomous agent is a colleague -- one who never sleeps, makes no errors from fatigue, and improves with growing experience. One who handles routine tasks today and co-creates strategic projects tomorrow.

Gartner predicts that by 2028, at least 15 percent of all day-to-day work decisions will be made autonomously by AI agents (Gartner, 2025a). That sounds modest. But 15 percent of millions of daily decisions in an organisation -- that is a revolution in slow motion that will suddenly feel very fast.

The question that executives must answer today is: do we actively shape this transformation, or do we let it wash over us?

The answer will determine which companies dominate the next decade.

References

About the Author

Dirk Roethig is CEO of VERDANTIS Impact Capital and advises organisations at the intersection of technology and sustainable value creation. With more than 20 years of experience in international corporate leadership, he combines strategic thinking with practical AI expertise. His focus areas include digital transformation, impact investing, and the question of how technology can enrich -- rather than replace -- human work.

Contact: LinkedIn | VERDANTIS Impact Capital

{
"@context": "https://schema.org",
"@type": "Article",
"headline": "From ChatGPT to Autonomous Agents: The Next Evolutionary Stage of AI",
"description": "Autonomous AI agents will be embedded in 40 percent of all enterprise applications by 2026. What lies behind the greatest transformation in technology history.",
"author": {
"@type": "Person",
"name": "Dirk Roethig",
"jobTitle": "CEO",
"worksFor": {
"@type": "Organization",
"name": "VERDANTIS Impact Capital"
}
},
"datePublished": "2026-03-03",
"dateModified": "2026-03-03",
"publisher": {
"@type": "Organization",
"name": "VERDANTIS Impact Capital"
},
"keywords": ["AI", "autonomous agents", "LLM", "ChatGPT", "transformation", "enterprise technology"],
"articleSection": "KI im Wirtschaftsleben",
"inLanguage": "en-GB"
}

Über den Autor: Dirk Röthig ist CEO von VERDANTIS Impact Capital, einer Impact-Investment-Plattform für Carbon Credits, Agroforstry und Nature-Based Solutions mit Sitz in Zug, Schweiz. Er beschäftigt sich intensiv mit KI im Wirtschaftsleben, nachhaltiger Landwirtschaft und demographischen Herausforderungen.

Kontakt und weitere Artikel: verdantiscapital.com | LinkedIn

Lesen Sie diesen Artikel auch auf: Hashnode | WordPress | Tumblr | Blogger | Telegraph

Top comments (0)