The newest AI revolution has arrived

Large-language models (LLMs) and other forms of generative AI are revolutionizing the way we do business. The impact could be huge: McKinsey estimates that current gen AI technologies could eventually automate about 60-70 percent of employees’ time, facilitating productivity and revenue gains of up to $4.4 trillion. These figures are astonishing given how young gen AI is. (ChatGPT debuted just under two years ago -- and just look at how ubiquitous it is already.)

Nonetheless, we are already approaching the next evolution in intelligent AI: agentic AI. This advanced version of AI builds upon the progress of LLMs and gen AI and will soon enable AI agents to solve even more complex, multi-step problems.

The promise of agentic AI

While AI has rapidly advanced in natural language processing (NLP) and computer vision, most gen AI and LLM applications today remain relatively “narrow.” That is, they can only complete simple or one-step tasks by retrieving information from a publicly and quickly available source, such as the internet. More contextual or complex tasks are impossible to achieve because multi-step prompts require the system to access various interfaces or tools (e.g., your personal calendar app, weather app, and notes app).

Agentic AI differs from these traditionally narrow systems in a few crucial ways. First, agentic AI systems possess a degree of autonomy. They can adapt their behavior based on the user’s inputs and goals. This is possible because agentic AI systems are built on foundation models that are trained on an extensive and diverse set of training data. As a result, agentic AI systems are more well-rounded in their reasoning capabilities and less likely to be affected or misled by unexpected outcomes or inputs.

Another critical differentiating factor is agentic AI’s interoperability. Unlike robotic process automation (RPA), agentic AI systems can converse with several tools, from business software to human input and additional foundation models. So, these systems can solve highly complex problems that require contextual business information, seeking insights from one source and then another -- just like a human would.

In a sense, agentic AI signals a departure from automation and a step toward orchestration. Just imagine: Virtual assistants may soon become highly skilled workers capable of not only suggesting a travel itinerary but also booking flights, hotels, and other accommodations.

Now, apply this higher degree of reasoning and logic to critical business initiatives and decisions. The possible implications of this evolution are immense.

Nascent risks to consider

Potential ethical issues with LLMs are well-documented. These engines are prone to bias and “hallucinations,” occasionally generating inaccurate or fabricated information. Most AI practitioners know these risks and have become diligent in cross-checking. For instance, many financial institutions (FIs) have enacted robust AI fact-checking processes, never allowing AI outputs to reach clients or the public without a thorough review.

Because agentic AI relies on multi-step LLM processing, it has the potential to amplify these unmitigated biases, errors, and unintended behaviors. The consequences of unchecked agentic AI can be severe, from discriminatory decision-making to data breaches and physical damage. To get ready for agentic AI and to minimize the risks, I urge all leaders to enact the following programs:

• Organize your data for use by AI: Understand where data resides, how to access it, and ensure compliance with data security, governance, and privacy laws.

• Build internal AI guidelines and controls: Establish clear policies, processes, and accountability measures for agentic AI projects, baking in ethical considerations from the ground up. For example, it’s wise to implement a human-in-the-loop system that ensures all AI outputs are verified by a human.

• Conduct comprehensive risk assessments: Thoroughly evaluate potential harms and unintended consequences of agentic AI applications, considering everything from algorithmic bias to cybersecurity vulnerabilities.

• Implement robust testing, monitoring, and course-correction mechanisms: Rigorously test agentic AI systems, continuously monitor for emerging issues, and be prepared to intervene quickly and make adjustments if problems arise. Remember, if your AI systems regularly operate differently from a human, that’s a problem.

• Understand your vendors’ approach to agentic AI: As more advanced forms of AI emerge, it’ll become even more critical to understand your vendors’ policies and regulations surrounding its use. Most industry-leading vendors have adopted a pragmatic approach to all AI by prioritizing consumer data privacy and model transparency -- these stipulations become even more essential as the power of AI increases.

The future of AI and business

It’s difficult to pinpoint when agentic AI will become mainstream. However, it’s already starting to emerge as we speak; leaders must take action to acknowledge and prepare for its rollout.

Specifically, organizations should approach agentic AI’s development and deployment cautiously and responsibly. The ethical risks of autonomous, self-learning systems must be carefully managed through rigorous testing, monitoring, and governance frameworks.

By successfully positioning themselves to navigate the questions surrounding agentic AI, leaders can help ensure they’re at the forefront of this technology when it inevitably becomes a critical competitive advantage.

Vall Herard is the CEO of Saifr.ai, a Fidelity labs company. He brings extensive experience and subject matter expertise to this topic and can shed light on where the industry is headed, as well as what industry participants should anticipate for the future of AI. Throughout his career, he’s seen the evolution in the use of AI within the financial services industry. Vall has previously worked at top banks such as BNY Mellon, BNP Paribas, UBS Investment Bank, and more. Vall holds an MS in Quantitative Finance from New York University (NYU) and a certificate in data & AI from the Massachusetts Institute of Technology (MIT) and a BS in Mathematical Economics from Syracuse and Pace Universities.