AI agents are exciting but not a silver bullet

Two years have gone by since ChatGPT captivated the world. Upon its initial release, everyone was amazed by its seemingly magical ability to answer any question in any manner users requested. People recognized not only its potential to disrupt the world of knowledge but also predicted that natural language and voice interfaces might become standard in technology applications, changing how software is designed.

What is agentic AI?

While 2023 was the year of ChatGPT and the introduction to generative AI capabilities, as 2024 has progressed, global interest has increasingly turned toward what’s known as “agentic AI”.

Agentic AI is an extension of existing generative AI functionalities. The difference from existing large language model capabilities — which produce a single output in response to a single input — is that agentic AI takes a single input and performs a series of “reasoning” steps to produce a more considered response to the initial prompt.

It does this by feeding the AI-produced output back into the large language model to determine what the next step should be. This means that a simple prompt given by a human can be broken down into a series of tasks by a large language model to produce a more thoughtful response compared to classic generative AI capabilities.

For many, this kind of technology could be game-changing. It democratizes bureaucratic processes and allows humans to get from point A to point B without necessarily knowing the route they need to navigate to get there. When combined with integrations into other generative AI capabilities and system integrations, this technology seems closer to the autonomous functioning many envision when thinking about artificial intelligence.

Risk aversity and control

Nonetheless, adopting agentic AI is not without risks. While it may be suitable for quick or one-off tasks, many professionals are likely concerned about a computer deciding the exact steps to take in a given process. This is especially true in risk-averse professions, such as lawyers. (When I used mail merge to carry out a process, a partner once asked me to make sure the computer had done its job properly).

Product designers would be wise to keep this in mind when leveraging generative AI capabilities. If control is important in a given process, they should ensure there is an opportunity for users to understand all the steps taken and to “course-correct” the AI if it seems to be heading in an incomprehensible or inadvisable direction.

Agents v. process design

At an organizational level, there may be concerns about leveraging agentic AI given how large language models operate. It is fundamental to large language models that they function in a stochastic manner rather than a deterministic manner.

In practice, this means that the output of large language models is not always predictable and that different actions can be taken in response to the same prompt. This could have several negative repercussions if different people across the business are trying to do the same thing but using generative AI results in different routes being taken to achieve it, most notably related to risk, efficiency and data management.

Indeed, for many processes, organizations may prefer to give users as little latitude as possible in how they conduct their work. This means that some degree of design must occur when considering agentic AI. For example, is the process in question best accomplished through an agentic AI-powered route, or would it be better to use a deterministic method such as an expert logic application? Or perhaps a combination of both?

All of this reminds me a little of the craze about “robotic process automation” a few years ago, where computers could carry out rote operations on your computer, e.g. downloading things from websites, placing them in a folder, OCR-ing them, capturing information from a specific paragraph, copying it into a form etc. While some of this technology seems impressive, the reason it exists is because a gap is not being bridged between different systems in a holistic process that works for users. I see robotic process automation largely as a sticking plaster for this, and I’d prefer to fix the underlying problem than deal with individual symptoms. I have largely the same concern for agentic AI systems as well.

Human reasoning

A common thread runs through all applications of generative AI. Many creators of generative AI models are keen to emphasize that these models mimic human reasoning. In reality, the fundamental mechanism through which large language models work is “next token prediction,” meaning that the models produce their output based on the semantic relationship between words and phrases.

While we do not know exactly how the human brain carries out its reasoning function, it seems an oversimplification to equate how machines and how humans think. From what I read, it seems that those who are “all-in” on AI are eager to make this comparison, and those who are “anti-AI” are eager to dispel it. I care less about the debate as to whether machines can “reason” and more about whether we are using the right tool for the job.

We should bear this in mind when considering whether a process requires a human to be in the driving seat or whether it is sufficient for AI to be in control with human oversight. We should also consider the impact on a human’s capability to think and whether we lose something if AI repeatedly carries out a cognitive task that humans should be performing. These concerns have been present since the introduction of large language models, but they are perhaps brought to the forefront with the introduction of agentic AI capabilities.

Agentic AI is an advancement that is exciting in multiple ways. First, it is a new tool in the arsenal for creating game-changing applications that can transform the lives of consumers and businesses. Second, it calls into question the nature of what we do in our day-to-day lives as humans, whether we act for specific reasons, and if it is time to change. The actual repercussions of embedding agentic AI in applications will continue to be felt over the coming years, but this is an interesting space to watch in the meantime.