Behind the Buzzword: What is agentic AI?

In this edition of Behind the Buzzword, we cover a hot topic in the tech industry right now: agentic AI. Here's what you need to know about it, explained in a four minute read.

What is Agentic AI?

Agentic AI is an AI system that can complete complex, multi-step problems with limited to no human supervision. With traditional AI, you ask it to complete a single task at a time. However, agentic AI can proactively work to achieve a broader goal until it is complete.

An example: Comparing traditional AI vs agentic AI

To illustrate the difference between the two types, let's take a simple task like ordering a pizza. For a traditional AI, you might type in a query like:

Can you suggest a good list of pizza places in San Francisco?

And the AI might respond with:

Here are some of San Francisco’s top pizza spots for a variety of styles—and the latest additions to the scene: Tony’s Pizza Napoletana (North Beach), Del Popolo (Nob Hill), Pizzeria Delfina (Mission & Pacific Heights)...

One query, one result! However, with an agentic AI, you could ask:

Can you order me a great pepperoni pizza from somewhere nearby?

And the agentic AI could:

• Consider your current location as a starting point
• Make a list of all your favorite pizza places and highly ranked locations nearby
• Select the best of these locations that have a pepperoni pizza on the menu
• Make an order, including entering your address and payment details
• Track the status of your order and let you know the pizza is on the way

In this case, the AI has acted as an agent on your behalf. They’ve taken the initiative, made decisions, interacted with external systems, and executed a full workflow autonomously. All of this would have taken a lot more work---and human intervention---with a traditional AI system.

An example of an agentic AI solution would be ChatGPT agent, which I've written a review about. It does indeed pass the pizza test (though not without some hiccups, as the technology still has a lot of improvement to be truly agentic.)

Agent washing vs Agentic AI

Unfortunately, agentic AI is currently being used as a buzzword, with many products claiming to have agentic AI, when behind the scenes, it’s really just plain old chatbots or task automation with no intelligence, adaptability, or autonomy.

This is known as agent washing, and is a very common practice. So, whenever someone says "our solution is agentic!" treat the claim with a large dash of critical thinking. Put the tool or service through its paces, and see if it can really replicate near-human autonomy, or if it's just good at doing one narrow task (often with a lot of hand holding.)

Where is agentic AI useful?

If we're talking true agentic AI---the type that is fairly autonomous and able to complete complex tasks without issue---then you're looking at something that's benefical across practically every field: software development, customer support, cybersecurity, manufacturing, finance, and so on.

Many agentic solutions currently available are in their infancy, and require a significant amount of human oversight, which sometimes defeats the point. Make sure you're properly timing your actual efficiencies when using an agentic solution: count the time it takes you to create the prompt, gather the contextual information, wait for the prompt to load and execute, and answer the AI's prompts when it asks you to jump in. Also, make sure to keep track of task error rates.

The downsides of agentic AI solutions

To get an AI agent to act on your behalf, you have to often hand over your authority and permissions. For example, to purchase you something online, you'd need to give it your account and payment details (and I can attest to this if you're using ChatGPT's Agent feature.) Naturally, there's security and privacy concerns with that, not to mention if the AI decides to hallucinate and order you a bunch of things you don't want---or decides to delete all your account data unprompted.

Here are a list of some of the potential risks involved with using agentic solutions:

• Indirect prompt injection/data poisoning: The agent may visit a site with malicious instructions hidden on it which causes it to leak your sensitive information or perform destructive actions on the platforms you’ve connected it to. E.g. Destroying your Google data.

• Excessive autonomy: With great power comes great responsibility. Since you’ve given an AI the keys to the kingdom, it can perform any actions you can. E.g. Deleting all your account data without any confirmation from you.

• AI hallucination: AI are notorious for incorrectly interpreting instructions, making up information, or inventing unintended tasks. Because it’s got your login permissions (and potentially payment details) this means it can do things you didn’t ask it to do. E.g. Buying something unrelated to your request, compromising your data, etc.

• Lack of auditability: You’ve given the agent your login details. So, if it does something sketchy, how does someone tell the difference between you and it in the system logs? “An AI did it” isn’t going to necessarily fly in this scenario.

Most of these revolve about what could be called the "God Problem": If you're giving a solution enough power to be useful, you're giving it enough autonomy to act against your best interests.

Aside from this, there are also concerns about human replacability with agentic solutions. Obviously, if you get to the point you can completely replace a human with an agent, this has wider societal implications, many of which are being discussed across various sectors right now.

Conclusion: A slice of knowledge about agentic solutions

Hopefully from this ultra-quick article, you're now familiar with what agentic AI is, and how it works if you were ever use it to order a pizza. That's a wrap on the latest Behind the Buzzword!

Want to learn more about Generative AI?

Pluralsight offers a range of beginner, intermediate, and expert AI and ML courses, including dedicated courses on generative AI. Since you can sign up for a 10-day free trial with no commitments, it’s a great way to take some professionally authored courses with a set course structure. Why not check them out?