Industry Insights: The Rise of Agentic AI – Navigating the Next Wave of Artificial Intelligence

Contributed by the Irish Funds Emerging Tech and Innovation Working Group

"The Rise of Agentic AI – Navigating the Next Wave of Artificial Intelligence" explores the transformative potential of Agentic AI, a new generation of autonomous AI systems capable of achieving goals without constant human guidance. Building on the rapid adoption and success of Generative AI technologies, this document delves into the advancements, applications, and governance of Agentic AI. It highlights how these systems can revolutionise business operations, enhance efficiency, and address complex challenges, while also emphasising the importance of robust risk management and human oversight to ensure responsible and effective deployment.

In his Netflix documentary, ‘What’s Next? The Future with Bill Gates’, Gates identifies ChatGPT’s November 2022 launch was “the moment AI finally woke-up”.  The Cambrian explosion of interest and application of Generative AI technologies following the release broke all technology records.  ChatGPT was the quickest application to reach 1 million users, achieving this milestone in just 5 days. By comparison, Instagram took approximately 2.5 months, and Spotify required 5 months to reach the same user base. Generative AI has shown a faster initial adoption rate than smartphones or tablets, with Bloomberg projecting Generative AI revenue to exceed $1 trillion by 2031.  

The remarkable success and growing application of Generative AI have also paved the way for innovative solutions to address complex challenges. While domain-specific and knowledge-critical tasks have posed hurdles for LLMs, advancements such as Retrieval Augmented Generation (RAG) are transforming the landscape. RAG operates by building a specialized corpus of knowledge that allows LLMs to draw from relevant information when responding to queries.  When a question is posed, the system retrieves the relevant information and sends it to an LLM, which then provides a clear and articulate response based on the retrieved data. This approach not only ensures accuracy but also allows for referencing the source of the answer, building trust and reliability in the generated content.   

Agentic AI: The Next Big Trend in 2025

RAG applications will continue to grow through 2025, but what other applications will become more common place this year? One strong candidate is Agentic AI. Gartner noted that Agentic AI is one of the Top 10 strategic trends for 2025. MIT Sloan Review highlight Agentic AI as one of the top 5 trends this technology. In his 6th January blog, Sam Altman stated that “we believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies”. According to research from Citi, references to “agentic AI increased 17x in 2024 – and we expect them to go parabolic in 2025”. Dennis Hassabis, CEO of Google Deepmind, note in 2024 that “I think the next things we're going to see perhaps this year, maybe next year, is more agent like behaviour," indicating a strong shift towards Agentic AI systems taking on more proactive roles in businesses and beyond

So, what are the AI agents Altman refers to, and how will they join the workforce? Agentic AI, as defined in a recent Harvard Business Review article, are “AI systems and models that can act autonomously to achieve goals without the need for constant human guidance. The agentic AI system understands what the goal or vision of the user is and the context to the problem they are trying to solve.” Generative AI, which many of us are familiar with, requires a user to enter a specific command or request through a user interface. This request, commonly referred to as a prompt, forms the basis for the Generative AI model’s response. In contrast, Agentic AI can independently act on specific tasks interpreting user goals and contextual requirements.

Whilst this technology has existed in various forms for many years, Generative AI's natural language capabilities have unlocked new possibilities. Agentic AI leverages the rapid advancement of LLMs, using probabilistic technologies to select optimal outputs based on its training. This key feature differentiates Agentic AI from Robotic Process Automation (RPA) or other previous iterations of this technology, which are rules-based, with systems following predefined instructions. However, Agentic AI should not be considered synonymous with LLMs. While LLMs focus on generating content based on user prompts, Agentic AI is designed to make decisions and execute specific tasks autonomously. For example, Agentic AI systems can independently perform a sequence of actions to achieve their goals. This ability to act decisively and independently marks a significant evolution in AI capabilities, enabling organisations to automate complex workflows and enhance operational efficiency.

How does Agentic AI systems work?

Nvidia define the following workflow for Agentic AI system.

1. Perceive – AI agents receive and process data from various sources pertinent to the task. For example an AI agent might analyse market data, economic

2. Reason – The LLM acts as the orchestrator that ‘understands’ the tasks. The LLM can be supported by RAG to leverage priority data and information. Continuing with our example, an AI agent might use RAG to extract and analyse customer credit histories, economic conditions, and regulatory compliance guidelines to support investment opportunities.

3. Act – By integration with other tools and applications, the AI Agent can quickly execute the tasks based on the plans it has formulated. An AI agent might execute trades in real-time based on pre-defined strategies, adjust portfolio allocations dynamically, or alert clients about critical market movements.

4. Learn – The ‘Data Flywheel’ concept enables the model to continually learn and improve based on data generated from its interactions

As one can appreciate from the Nvidia work flow, there is significant potential for this technology and the use-cases it can be applied to are vast.  McKinsey define a use-case where an Agentic AI system—made up of multiple agents supporting different tasks —supports the credit-risk underwriting process. 

1. The process is initiated by a user providing a high-level work plan of tasks with rules and specifications.  

2. This task would be decomposed into individual discrete tasks supported by specific AI Agents:

a) One agent could support the relationship manager to handle communications between the borrower and financial institutions.   

b) An executor agent could compile the necessary documents 

c) This documentation would be sent to an analyst agent to review the financial details  

d) A reviewer agent would look to identify discrepancies and errors and provide feedback.  

There are many examples of early adopters driving use-cases with this technology.  Moody’s has leveraged a multi-agent AI system to enhance its financial analysis processes. As highlighted in a recent Wall Street Journal article, this approach involves 35 specialized agents working collaboratively on various tasks, from project management to more complex supervisory roles. These agents are equipped with detailed instructions, distinct personalities, and significant data and research, enabling them to operate autonomously while delivering valuable insights. 

Risk 

Whilst the application of this is vast, its adoption presents significant risks, particularly the potential for misuse. Gartner stated that by “By 2028, 25% of enterprise breaches will be traced back to AI agent abuse, from both external and malicious internal actors.” As note earlier, at the core of these systems are Large Language Models (LLMs), which are probabilistic systems. As with any LLM-driven technology, the inherent risk of errors or inaccuracies extends to Agentic AI.  The risks with Agentic AI primarily stem from the following critical factors:  

1. Unclear instruction or training 

2. Agentic AI systems executing on actions that are not aligned to the use-case it was created for.   

3. Hallucinations from LLMs 

4. Potential risk from external actors 

Mitigating these risks requires robust governance frameworks aligned to the specific use case, supported by strategic oversight.  A foundational principle of this governance should be the inclusion of ‘Human Oversight’.  As Pawel Gmyrek emphasised in the World Economic Forums discussion on Agentic AI, “A ‘human above the loop’ approach remains essential, with AI complementing human abilities rather than replacing the judgment and accountability vital to the sector.”   

To implement this approach, visibility is paramount.  Visibility provides the ability to oversee “the  development and operation of these systems would significantly improve the  ability to predict potential harms from the technology and hold the relevant actors accountable.”  Key considerations should include; 

• Use Case Selection and Stakeholder Engagement - Ensure the selected use is appropriate and engage all stakeholder to align on the Agentic AI approach 

• Agent Identifiers - Implement agent identifiers to indicate when an AI agent is involved in an activity, the nature of the activity is and it’s success or failure metrics 

• Continuous Monitoring and Feedback Loops  -Actively monitor, track and analyse the Agentic AI systems and provide actionable feedback to both technical and operational resources 

• Real-time Logging and Analysis - Log all actions taken by Agentic AI system and conduct real-time analysis to detect anomalies or risks early 

• Frequent Operational Maintenance - Conduct routine debugging, patching and updates to address such  issues as model drift, biases, or emerging errors due to outdated data or assumptions. 

By embedding these practices within a comprehensive governance structure, organizations can reduce risks and enhance the strategic value of Agentic AI while maintaining accountability and trust. 

Opportunity

Agentic AI represents a significant evolution in artificial intelligence capabilities, moving beyond the reactive nature of Generative AI toward autonomous, goal-oriented systems. In a recent note, Bank of America identified Agentic AI as the third major evolution in AI, succeeding the Pre-GenAI and GenAI phases in the evolution of AI.  The opportunity is there for organisations to start applying this technology to use cases, through pilot programs with robust governance framework.  Pilots should focus on low-risk use-cases to support building confidence with this technology and establish clear metrics for measuring success and business value will be required to scale beyond pilot programs.  As we move through 2025, the strategic integration of Agentic AI can help drive business transformation, but this transformation must be balanced with appropriate risk management and governance controls.