What is Agentic AI? Definition, features, and governance considerations

Key Takeaways:

• Agentic AIs are systems that can make decisions with some level of autonomy and adapt to changing inputs from multiple sources without much or any human oversight.
• An Agentic AI system differs from non-agentic AI systems in its ability to make independent decisions and proactively adjust its approach to meet its goals.
• Agentic systems live on a spectrum of autonomy. It’s important to understand the different levels of agency to decide how to govern usage, from non-agentic AI systems through to fully autonomous and self-governing systems.
• Implementing Agentic AI requires regulatory compliance, ethical safeguards, and transparency.

Agentic AI is the latest buzzword being thrown about by technology heavyweights. Forbes referred to it as “the next big breakthrough that’s transforming business and technology”. However, when you ask people to define it clearly, things quickly get vague. 

So what is Agentic AI exactly, how does it differ from other AI systems, and is it really going to change the world?

Agentic AIs are artificial intelligence systems that are able to act with some level of autonomy, meaning they are able to make decisions to achieve specific objectives, based on input data. They are called Agentic AIs because they have “agency” or freedom to make decisions as they see fit.

Agentic systems can adapt to changing circumstances or inputs to take the best course of action with limited or even no human oversight.

Imagine you work for a large shoe retailer and you’re responsible for supply chain work. A highly autonomous Agentic AI could independently manage your company inventory by predicting product demand based on real-time sales data and external factors (like economic trends, fashion trends, or even the season).

It would then be able to adjust stock levels, place orders with suppliers, and optimize shipping routes to ensure timely deliveries, without human intervention.

Agentic AI systems exist on a spectrum, so while this example is highly autonomous, other systems with autonomy over certain things or with stricter limitations are still “Agentic AI” systems.

Agentic systems dynamically and autonomously adjust their approaches to meet their goals,  whereas regular AI systems do not because they lack autonomy. Non-agentic AI systems include non-autonomous generative AI and analytical/predictive models that also lack autonomy. Non-agentic AI/ML systems typically learn parameter rules from the data itself but do not have any level of autonomy. However, Agentic AI systems can assess their environments and make informed decisions about the best next course of action towards their goals with some level of autonomy. 

Taking into account our example of managing inventory for our shoe retailer above, here’s how a regular AI and an Agentic AI would manage tasks differently:

• Regular AI: A traditional AI system might generate demand forecasts or suggest optimal inventory levels based on predefined models or historical data, but it would require human intervention to make decisions, place orders, or adjust strategies. For example, a regular AI could be set up to ping the supply chain manager when demand for a particular shoe style is rising, but the manager would have to take initiative and act on that information to order more of those shoes to meet demand.
  
• Agentic AI: An Agentic AI would theoretically be able to independently and autonomously assess the current stock situation, decide how much inventory is needed to meet demand, place orders with suppliers, reroute deliveries, and even adjust pricing if necessary. 

Agentic AI systems have autonomy, can adapt or spontaneously change their approach to meet a defined goal, and are context aware. 

• Autonomy: The ability to operate independently once given an objective.
• Goal focus vs task focus: A focus on achieving specific outcomes or goals vs performing defined tasks without an understanding of the overarching goal.
• Adaptability: The ability to adjust strategies or next actions if the situation changes spontaneously to ensure it reaches its goals.

Many of the most relatable Agentic AI use cases are found in the fields of robotics and autonomous systems. For example, a self-driving car includes an Agentic AI system because it is able to take in environmental data and deploy safety precautions if needed. Even though research into autonomous vehicles has been underway since before the term “Agentic AI”, the system would still be classified as having some level of agency.

However, even these systems are not fully agentic, because they make decisions based on rules or constraints set by humans.

Not all Agentic AI systems are created equal. Agency in AI exists on a spectrum vs in black and white terms. So the level of autonomy a system has can vary significantly. 

The more autonomous and adaptable an AI system is, the greater the potential risks and governance challenges are.

Here's a breakdown of the various levels of agency in AI systems:

Agentic AI naturally gets hyped up. The possibilities are pretty cool. But in reality there are huge hurdles to overcome to implement and govern these systems safely and responsibly. Given their autonomy, ensuring compliance with legal and ethical standards is critical – especially in high-risk sectors like finance, healthcare, or infrastructure.

The EU AI Act and the US Executive Order on AI both set out guidelines for governing AI systems. Agentic AI could fall within any of the risk classifications of the EU’s AI Act based on its implementation. So an Agentic AI system within a video game would not fall into the high risk category. But an Agentic AI system running a nuclear power plant or flying a plane would. In these high risk environments companies need to follow strict protocols, including: 

• Transparency: Users must understand how the AI system works and how it makes decisions.
• Data Governance: The data used by Agentic AI must be carefully managed to prevent biases or discriminatory outcomes.
• Documentation and Traceability: Every decision made by the AI must be traceable to ensure accountability, particularly if things go wrong. 

In addition to these legal concerns, organizations must establish internal governance frameworks that allow for human oversight while still enabling the AI to function autonomously.

This means you also need to build "fail-safes" to override AI decisions in critical situations and ensure that all actions are auditable, which can be a real challenge with AI systems.

Agentic AI has a lot of potential in various industries, which is one of the reasons it is hyped. But real-world use cases are few and far between due to the strict governance requirements and high levels of risk posed by autonomous systems:

• Supply chain optimization: Companies can use Agentic AI to autonomously manage inventory, predict demand, and optimize logistics, reducing costs and improving delivery times.
• Autonomous vehicles: Self-driving cars involve some level of Agentic AI to process environmental data in real-time, which helps them avoid obstacles and make safety decisions without human intervention.
• Healthcare diagnostics: Agentic AI can be used to autonomously analyze medical images, identifying potential diseases or conditions faster than human doctors could, and suggesting treatment options.

Agentic AI offers huge potential across industries, from supply chain optimization to autonomous vehicles. However, despite the hype we are still quite a long way from having many truly ambitious agentic systems in operation. 

The more autonomous AI systems become, the more responsibility, governance requirements, and risk there is. 

Organizations must implement strong governance frameworks, adhere to evolving regulations like the EU AI Act, and be able to explain how Agentic AIs have made decisions and reproduce results – which can be very difficult to do since AI models are often non-deterministic.

While Agentic AI is currently a much hyped advancement in AI, safe and secure implementations and use cases are few and far between. We can expect to see AI systems becoming gradually more autonomous in line with the 7 levels outlined above, but fully autonomous systems are still entirely in the realm of science fiction. For now at least.

Agentic AI – unlike RPA systems – can make autonomous decisions and adapt to changing environments. RPA on the other hand follows predefined, rule-based workflows for repetitive tasks. While RPA automates specific processes, Agentic AI is capable of reasoning and responding to dynamic situations independently.

Generative AI creates new content based on patterns learned from data, like generating text or images. Agentic AI makes autonomous decisions and takes actions to achieve specific goals, adapting to new information and environments, usually without human intervention. However, the current iteration of Agentic AI does rely on generative AI models like GPT4o as a central component, so they are not always perfectly distinguishable.