Agentic AI: Good or bad?

In the past few years, as AI started getting used more and more in the workspaces, it started to develop a reputation as the master helper. The assistant you wish you had. Explain it as a task, and it will do it for you. But hey, your instructions better be clear. That’s what’s known as a prompt, and people understood the importance of learning how to prompt better. Things were good after that for a while. Prompt well, and AI will get your favourable results. But for technology, even the sky is not the limit. And this year, we have amongst ourselves Agentic AI — a technology capable of understanding the task by itself, realising what needs to be done and executing it as well. 

Say you have ordered a product from an e-commerce website and you want a refund. A regular AI will write the refund model for you if you prompt it right. But an Agentic AI will draft the email by itself, check the order status and submit the refund request in the system. While normal AI helps you complete a task, Agentic AI can take all actions necessary to finish a certain goal. It can also make decisions. If going through a task, it can decide if it is capable enough to solve the task at hand or if the task needs human intervention. Another key difference is that Agentic AI does not follow a rigid script; it can switch depending on what is happening. Its actions are based on context. While regular AI acts like a tool, Agentic AI can be looked at like an artificial teammate that is capable of overseeing an entire project to its completion. It also learns from feedback and improves its decision-making skills automatically. 

How green is this valley, right? But with this level of automation also comes a long list of concerns. Firstly, since it can execute steps without human intervention, it can also make mistakes without it. And since it can execute a project end-to-end, the mistakes won’t just stop at harmful suggestions but will have a negative real-world impact. Because the AI literally is autonomous with agency, if it misunderstands your instruction, it would be very hard to control its agency after it has begun executing, because an Agentic AI closes the human suggestion loop. Such a mistake at scale can be preposterous. Finally, once such a mistake has happened, it becomes very hard to pinpoint who to blame. Is it the AI’s fault? The user’s? The developer’s? Accountability becomes a grey area.