Why Your AI Project Might Be Stuck in 'Cruise Control' Mode

I have a confession to make: around ten years ago, I almost convinced my wife to postpone buying a new car because "self-driving cars are just around the corner." Well, it's been 10 years, and I still have to drive it myself. This expensive lesson in technology hype versus reality feels particularly relevant as I watch history repeat itself with GenAI.

Vendors like Devin AI promise 'autonomous knowledge work'. Some key stakeholders buy the promise so it is crucial time we get real about AI autonomy levels - before more of us make the same overoptimistic predictions I made about self-driving cars.

The automotive industry offers a sobering parallel. It took 15+ years to move from basic cruise control (Level 1) to limited autopilot (Level 2). And despite billions invested, we're only seeing highly restricted Level 3 pilots in 2024. Each level isn't just an incremental step - it's a massive technological leap.

Let's map this reality to GenAI capabilities:

Level 1 (Basic Assistance): Current GenAI can complete clearly defined tasks with human guidance. Like cruise control, it handles one aspect while humans manage the rest. It'll write that first email draft, but you're still driving the conversation.

Level 2 (Partial Automation): Today's best implementations can handle complex workflows - think Tesla Autopilot for knowledge work. It can analyze lengthy documents, generate comprehensive reports, identify patterns in data, and maintain context across multiple tasks. However, like Autopilot, it requires constant human oversight. You need to verify its outputs, correct its course when it hallucinates, and maintain responsibility for the final decisions. The AI is a powerful co-pilot, not an autonomous operator.

Level 3 (Conditional Automation): The massive leap. Just as cars needed sophisticated sensor fusion and millions of road hours to achieve limited Level 3 (think modern vehicles that can handle highway driving but require human takeover), GenAI needs similar breakthroughs in reasoning and reliability. We're seeing early experiments in narrow domains, but widespread Level 3 GenAI is likely quite far away.

Level 4 (High Automation): While Waymo has achieved this in specific geo-fenced areas for driving, it's largely theoretical for GenAI right now. We'd need systems that can handle entire knowledge domains without supervision - like a lawyer bot that could independently handle standard contracts from start to finish. The complexity jump from Level 3 is enormous.

Level 5 (Full Automation): Those vendor pitches promising full knowledge work automation? About as realistic as promising fully autonomous cars handling every edge case globally - from unmarked dirt roads in rural India to complex construction zones in Tokyo. While companies like Waymo show impressive progress in structured environments, true "go anywhere, handle anything" autonomy remains a massive challenge. The same applies to GenAI - we're very far from systems that can truly replace human knowledge work across all domains and contexts.

Most successful enterprise GenAI implementations operate between Levels 1-2. And that's perfectly fine - augmentation often delivers more value than automation. Problems appear when vendors promise and stakeholders expect AI solution to operate on Level 3, or even Level 4. We need to be pragmatic - it is simply not possible at the moment. Yes, technology accelerate at frightening speed, but the reality is complex and messy and with any pace of technology development it can't simply jump from Level 2 to Level 5 (yeah, there is a slight possibility of self-improving autonomous system reaching technological singularity over few weeks, but if this happens I will not worry about enterprise AI adoption that much anymore)

Today, in February 2025, the key to success isn't chasing full autonomy but understanding how to effectively collaborate with AI at its current level and be pragmatic about current capabilities and legal/ethical limitations. Just as Tesla owners need to keep their hands on the wheel, knowledge workers need to stay actively engaged with GenAI tools.

Let's not make the same mistake I made with self-driving cars. Instead of waiting for a future that's further away than vendors claim, let's focus on capturing value from the assistance and augmentation capabilities we have today. And maybe, just maybe, we should listen to our spouses when they suggest taking a more practical approach to innovation. They tend to be right about these things.

P.S. After finishing this post, I remembered that Google DeepMind offered this framework for assessing autonomous systems capabilities almost 2 years ago. Take a look at their model comparing narrow vs general capabilities across different performance levels - from "Non-AI" to "Superhuman." What's fascinating is how many capabilities have remained at the same level despite the apparent rapid progress.