Agentic AI: Are We Building Castles on Quicksand?

Ari Ramkilowan (Head of Machine Learning) & Stef Adonis (Head of Marketing)

Artiﬁcial Intelligence is in a strange spot. With the explosion of AI tools and applications, we ﬁnd ourselves teetering between two inseparable yet intertwined paths – the promise of extraordinary capability, and the peril of unmitigated risk. This precarious balance gives rise to the question:

‍

Are we building something truly enduring, or are we rushing ahead on unstable foundations, building castles on quicksand?

‍

Agentic AI covers quite a diverse range, spanning from simple chatbots to the vision of fully autonomous systems that can act, reason, and take initiative. While the current hype often overshadows practical discussions, there is undeniable potential for rapid advancements in this ﬁeld. Agentic AI systems go well beyond mere button-based conversational interfaces, offering tools that integrate into complex enterprise operations.

‍

While the appeal is undeniable, a leap of this magnitude toward fully autonomous systems in enterprise-level applications could lead to unforeseen risks. While the threat of these risks remains a reality, we should instead be focusing on human-led Agentic AI – a level where intelligent tools enhance operations while ensuring human oversight.

‍

The key distinction lies in initiative, and the ability to plan. For example:

‍

An LLM is like an incredibly well-read librarian who can instantly recall and synthesise vast amounts of information from books. Ask this librarian a question, and they'll provide a comprehensive, eloquent response drawing from their extensive knowledge, or the wealth of information at their disposal. If prompted, and asked really nicely, they might even respond as a pirate. They're exceptional at retrieving and combining information, but they always wait for your speciﬁc query.

‍

An agentic application, on the other hand, is like that same librarian, but instead of simply answering your question, they take it a step further by showing some initiative. They might say, "Based on what you're asking, I think you might also want to explore these related topics. I'll go ahead and pull some additional resources, draft a preliminary research summary, and even reach out to some subject matter experts who might provide deeper insights.”

– Ari Ramkilowan, Head of Machine Learning

‍

The agentic application introduces a layer of goal-oriented behaviour, breaking down complex tasks into sub-tasks, making decisions, and taking actions beyond mere information retrieval. It has the capacity to perceive an environment and take purposeful actions toward a speciﬁc goal rather than following a speciﬁc query or a predetermined sequence of events.

‍

This holistic approach underlines its superiority to rigid, workﬂow-based tools that falter in handling edge cases.

‍

While the journey toward fully autonomous agentic systems may still be on the horizon, enterprises are beginning to invest in the technology. The interest lies in faster iteration and broader scope, where agentic systems introduce ﬂexibility without replacing existing workﬂows.

‍

However, the promise of agentic AI comes with a great deal of risk, especially for businesses – misalignment of goals, unpredictable behaviour, loss of human oversight, ampliﬁcation of bias, and security risks – all of which demand careful navigation.

‍

So, we must ask ourselves not whether we can build this but should we build this.

‍

There is a path forward that is more of a hybrid model – one that lies between structured processes and autonomous agents. This will give us the efficiency of agentic AI and the security of human involvement.

‍

The allure of agentic AI is immense, but so are the responsibilities that come with it. Oversight, accountability, and ethical alignment must serve as the foundation of our innovation. These systems should enable autonomy within controlled parameters, minimising risks while maximising potential.

‍

As we look ahead, human-led Agentic AI may just emerge as the "sweet spot" - a balanced middle ground where technology supports rather than replaces human expertise.

‍

The evolution of agentic AI is not just about technology; it’s about deliberate and thoughtful integration. While the idea of fully autonomous systems tempts us with the promise of efficiency and innovation, it also demands vigilance. Building robust AI systems isn’t about surrendering control but exercising it wisely.

‍

So we don’t need to build those castles on quicksand after all. We have the power to create a much ﬁrmer middle ground that combines the strengths of agentic AI and human expertise.

Agentic AI: Are We Building Castles on Quicksand?

"

"The allure of agentic AI is immense, but so are the responsibilities that come with it. Oversight, accountability, and ethical alignment must serve as the foundation of our innovation."