I continue to be amazed in everyday conversations with senior business and nonprofit executives at how few still have significant hands-on experience with the most common simple chatbots that have now been around for nearly three years.
In many/most cases, they are not even aware that ChatGPT is not the only major player in the game. If they have experimented with AI, it is usually ChatGPT, and they tend to use it like they would Google, which is not a great sign for Google, but also only scratches the surface of what these tools are capable of.
The Rise of AI Agents
Recent AI hype has focused heavily on "AI agents" – systems that can take instructions, use tools, and work toward goals with minimal human oversight. The eventual goal is to be able to assign AI a task and let it figure out how to complete it autonomously.
Large language models face a fundamental hurdle in becoming true agents. These models were built to predict the next most probable word in a sequence, using patterns learned from all the human generated content that the large AI companies have been able to get their hands on. This word-by-word prediction approach differs significantly from the kind of goal-oriented reasoning that autonomous agents require.
Despite these limitations, organizations are successfully deploying AI agents for specific tasks. Early implementations focus on simple, repetitive work where the steps are well-defined. Popular AI assistants like ChatGPT, Claude, and Gemini now include basic agent-like features that can break down requests into actionable steps.
What we're seeing isn't true autonomous reasoning, but rather sophisticated engineering. AI companies are building systems that use business logic and predefined workflows to transform user requests into step-by-step plans. The language model then executes these plans, creating the appearance of independent decision-making. What is currently lacking is creativity and the ability to plan at a higher level.
Using the progression of crawl, stumble, walk, run – we're still in the early stages of stumble when it comes to Agentic AI.
Current Capabilities and Future Vision
Chatbots and CoPilots work pretty well to help us write, fix our grammar, and code at a certain base level. Intelligent assistants can perform autonomous research for us with reasonable reliability. LLM-centered process automation (Task Agents) is being deployed with considerable success for point solutions, albeit with a lot of work to define the processes and establish guardrails.
If you want to think about the continuum in its entirety, the best description I have heard is from a consultant from G2 and Harvard. Think about the continuum from Waze to Waymo. With Waze, you get in the car, pull up Waze, and it provides turn-by-turn directions to help you drive to your destination on time. With Waymo, you get in the back seat, say a prayer, and Waymo gets you from point A to point B without any additional input from you.
On the path this technology is currently on, the agentic AI equivalent of Waymo will be when you can define a problem, the same way you would to a human, and a system of AI agents with different capabilities will coordinate a swarm effort to complete the task. Conceptually, this sounds viable, but there is still a lot of ground to cover between what is possible today at scale with the kind of reliability required for commercial use and that capability.
