Introduction
There is an increasing trend recently to call any action completed with an AI tool an AI Agent. While in many cases this might be true, it is equally possible that this has been mis-labelled.
This may stem from the fact that in this AI era there is confusion about the definitions of the terms we use daily fueled by people who may not fully understand what they are discussing disagreeing with established definitions. In some cases there may even be an outright refusal to define some terms.
The truth is that these terms can, and should be defined. Not fully or correctly defining the terms we use will hamper our ability to use these tools, may result in mis-use, or impact future iteration.
So what is the definition of agentic AI?
The simple definition is:- an AI tool that receives an input, uses that data to independently set goals, planning actions, making contextual decisions, and executing complex tasks with minimal human intervention by utilising tools and interfacing with external systems.
The Definition in Depth
First we should understand what agency means (as in AI agent).
Agency refers to the capacity of an entity to act, to make choices, to have intentions.
In AI, ‘agency’ draws on some of the earlier research into AI on intelligent agents and multi-agent systems, which for decades have examined systems that perceive, decide, act, sometimes in environments with partial information or changing conditions.
We can do this more easily on modern hardware, however the possibility of error means that human intervention means that we exercise caution.
In the definition we emphasised input and output. This is because the autonomous system requires a trigger or a reason to perform the action. An input can be anything such as a text prompt, an image, a sound etc. The output can be just as varied, but depends upon the predetermined action that is being taken.
The part in between is what makes a truly agentic AI. We expect the agentic AI to combine that input with previous known data and instructions, combine that with some thinking process (such as through an LLM), make decisions and execute based on those decisions. That last part is what sets agentic apart from a regular query and programmatic solution. It has to make a decision and execute on that decision.
What is and What is Not An Agent
We often confuse a simple chatbot with an AI Agent. I think the confusion comes about because of the term ‘Agent’. Simple chatbots in the past have often been misnamed as agents, and so that incorrect naming comes across into the field of AI.
Of course a chatbot can be truly agentic, and really should be.
For example, using the chatbots again, imagine a company that sells baby clothes. They have a chatbot on their ecommerce website.
-
A user asks if they have a particular style and size in stock. The AI, using Retrieval Augmented Generation (RAG), accesses the database and finds that they are in stock, letting the user know. This is not agentic.
-
A user asks where a brick and mortar shop is located. The bot replies with an address and map. This is useful, but is not agentic.
-
In another case a user reports that they have not received their order. The bot finds the order form the user’s email address, cross references with the Australia Post API and finds that after 4 weeks the item has not been delivered. Based on past interactions the AI model knows that an item not delivered in that time period is likely lost. It makes an autonomous decision to offer the user a refund or a voucher to purchase again. This is agentic AI.
-
A user is frustrated as the item they wanted is out of stock. They ask the chat bot when it will be available. The bot finding that it is not due to be in stock any time soon finds another similar product from the range. It is more expensive, but the bot has detected the frustration in the users, and discounts it heavily for the user. It provides a ‘buy now’ button in the chat window with the reduced price, completing a sale that otherwise would not have occurred and probably left a user annoyed. This is Agentic.
Of course there are many many more examples.
Some company’s products might give you the idea that AI agents are something else, or include a much broader scope, but that is, at the end of the day, marketing. And it is not really a problem - they can use whatever term they like, and the only thing you should assess on it is “what can it do for me”, and don’t get stuck thinking it is something else.
The Pitfalls
I hear you say “Agentic AI sounds amazing. There should be more of it!”.
And you may be right, agentic AI can certainly be used more. There is a catch though, our modern AI systems can still get things wrong. Sure they are better than a couple of years ago, and even better than they were at the start of the year, but they can still make errors. And when we let them make decisions about pricing, discounts and other potentially loose creating things, we certainly need a human in the loop.
As it happens, image recognition tools are very good at spotting things like cancerous cells inside an otherwise healthy lung, sometimes even before a doctor. Imagine if we allowed that to be agentic without a human doctor intervening. This could mean that an AI could ingest an X-Ray, diagnose cancer, schedule and execute a course of treatment for a patient.
A doctor might take a more considered approach. They may appreciate the early diagnosis, but the treatment might be more appropriate. Or in an extreme case, they might realise that it was not cancer at all, but rather just cheese on the lens (Futurama joke).
And so we have that little part of the definition “minimal human intervention”. Perhaps one day we could abandon that, but certainly not yet.
Another interesting thing about agentic AI, is that it seems to also include a partial programmatic solution. Interfaces with external systems, decision trees and execution on interfaces all require old-school programming. We in fact have a hybrid of AI and non-AI tooling being used. Again, this is not a bad thing, just useful to know if you are planning on developing your own agentic tools.
Conclusion
The definition of Agentic AI is actually not nebulous, it is “an AI tool that receives an input, uses that data to independently set goals, planning actions, making contextual decisions, and executing complex tasks with minimal human intervention by utilising tools and interfacing with external systems.”
It is important to understand this definition if you are planning on building, iterating or using these types of tools. Understand what it is, and what it is not.
Use the tools to benefit you or your business, and don’t let them waste your time or money.
Ps. if you have any questions
Ask here
###You May Also Like:
Imagine a new way to create AI customer apps with Salesforce