The Agentic Journey

All AI news this week has focused on two key things:

1. Sam Altman commented that we are in an AI bubble, comparing the current market to the dot-com crash.

2. MIT’s report “The GenAI Divide: State of AI Business 2025” found that 95% of GenAI pilot programs failed.

While it may seem that a bubble is about to pop, much like the dot-com bubble, it doesn’t mean the technology (or at least the intent of its promise) won’t change things forever. The Internet’s initial intent was to share information. I’d say we’re meeting that. It’s an efficiency bump that revolutionized how we live.

For AI, that promise is to carry out tasks and decisions that humans would typically have to do. We’re not quite there yet with AI, but we’re on our way on that journey. And I don’t see us turning back.

But what does that journey look like?

In a few conversations this past week, I heard references to the levels of autonomous driving. I think it’s a great way to think about it, especially given how I see the parallels of lagging security controls to automotive fatality rates.

Let’s break it down and compare it to the AI journey.

Level 0: No Driving Automation: The human controls all aspects of driving, with no automation support. It does include “warnings and momentary assistance.” These are features like blind spot or lane departure warnings and automatic emergency braking.

The AI equivalent is essentially what life looked like before ChatGPT took off, and you had to Google things yourself or the low-code/no-code platforms of the past (e.g. Zapier).

Level 1: Driver assistance: This is the lowest level of assistance available. While the human is still in full control, it includes features that provide steering OR brake/acceleration support. That OR is important because it allows for only one feature to be performed at a time. Examples include lane centering OR adaptive cruise control.

As we get into AI-enabled task augmentation, I see this as being small, trivial tasks that require basic reasoning that can be automated. Nothing too crazy here. I would lump research support or even Google’s AI search results here. Essentially, these are copilots.

This is where the vast majority of companies are operating today.

Level 2: Partial driving automation: The next level up introduces the AND. Its features that provide steering AND brake/acceleration support. So it’s the car peforming lane centering AND adaptive cruise control at the same time.

To me, this is a more complex AI workflows that combine multiple tasks. More advanced enterprises that have been experimenting with AI for a few years, along with a subset of B2B SaaS companies, are here.

Quick brake (pun intended): Levels 0 to 2 are all driver support features, meaning that the driver is still operating the vehicle. As we shift to Level 3 and beyond, we enter into automated driving features where you are not driving. I see this as the threshold between agentic workflows and agentic agents.

Level 3: Conditional driving automation: Allows for driving the vehicle under limited conditions and will refuse to operate unless all conditions are met. When the vehicle requests it, you must take over driving.

In the AI world, I see this as agents that require human-in-the-loop approvals. Agents are given specific goals and can handle known paths. If there is an unknown path or a sensitive operation that needs to be taken, it defers to the human for intervention. This is very much so still a human-in-the-loop (HITL) process.

There’s a lot of experimentation happening here, and I’m sure we will see companies deploying AI agents in the near future, but we’re not here yet.

Level 4: High driving automation: This is where we get into actual self-driving cars. This is the level that Waymo operates. To put it into perspective, pedals and steering wheels are optional at this point. But, like in a Waymo, things can still go wrong, and you can call a human to help. Like this guy who got stuck in a Waymo that was just driving in circles.

The AI equivalent here is an AI agent that has a goal given to it and is operating with full agency to accomplish it. Only in rare circumstances will a human have to intervene. This is the point where we reach AI playing the role of a full-time employee under the guidance of a human manager.

As much as the AI-hype marketing machine will say we are here, we aren’t.

Level 5: Full driving automation: The ultimate goal of self-driving cars. This is where the car can drive in all conditions.

For AI, this is agents achieving full autonomy, being able to define its own goals (hopefully under some constraints), and having the ability to improve itself over time.

If you like visuals, I ported this concept to the pathway to agentic AI.

If you have questions about securing AI, let’s chat.