8 years ago, I learned to drive in an old stick shift hatchback with a stuck clutch. The gearbox jammed randomly, so I never had a long uninterrupted sequence of independent driving. I got a license without much confidence in my ability to control a clutch.

Fortunately, I haven’t seen a manual car since. So it’s never been a problem.

The shift in how we work feels similar. I learned to build and write by hand. Those skills are going the way of the clutch pedal. Today’s narrative is all about learning how to steer and control speed, similar to driving an automatic.

But driving stick gave me a feel for what the engine is doing. When something sounds wrong, when the car is working harder than it should be. That awareness has helped me more than once. Work is going to be similar.

Accountability Starts With You

There are real fears around AI taking people’s jobs. I won’t dismiss that. But there’s a second concern I hear often from leaders: when AI does the work, who owns the result?

Good organizations run on ownership. Every decision can be explained by someone. When something goes wrong, that same person explains what happened. Reviews, planning, trust between teams, all of it assumes the people who produced a piece of work understood it.

AI breaks that assumption. Models aren’t a real part of your workforce, so they can’t own a result. You can designate a human owner, commonly this is the person who initiated that work but increasingly they’re out of the loop too. Agentic work happens independent of any human. Now, you have a person responsible for output they didn’t produce, generated by a process they don’t understand, and a sequence they can’t retrace. Of course they’re concerned.

It also gets worse with scale. Supervising one model on one task is manageable. But the whole point of AI is one person doing the work of ten. Accountability doesn’t scale at the same rate. Ten tasks means ten sets of results to own, explain, and defend.

To AI’s credit, the models of today are genuinely excellent. That said, correctness is the baseline, not the accomplishment. Nobody remembers the thousand things that went right. Nobody forgets the one that didn’t. When the mistake is yours, you can explain it, learn from it, own it. When the mistake came from a process you can’t retrace, you’re left holding something you can’t account for. That gets more uncomfortable as AI moves into higher-stakes work, where the tolerance for unexplainable errors approaches zero.

Beyond Observability

There are many LLM observability tools on the market, all focused on making model decisions legible. The hope is that if we understood how models arrived at the decisions they do, the accountability problem goes away. I don’t think it would. It will certainly help, but this issue isn’t technical.

Understanding a single decision isn’t the same as understanding a pattern of decisions. Knowing a worker’s thoughts at any given instant has never been enough to understand their judgment across a project. You’d need to understand how hundreds of individual decisions relate to each other, and that is a fundamentally different kind of understanding.

What matters is the habit of paying attention. The instinct to check before things go wrong. To develop, over time, a sense for what normal looks like and what doesn’t, built from watching the same kind of process run enough times.

Most organizations deploying AI right now check whether the final output looks right. Without procedural checks, you don’t know if your AI got there the right way. Consequently, you don’t know if the correct result was by chance or through reason. The distance between where organizations are and where they need to be isn’t about better monitoring tools. It’s about the people developing attention habits that make any tool useful.

To Remember, to Learn, and to Forget

The accountability gap won’t be closed by better models or better tooling. It’ll be closed by people with the ability to work alongside autonomous tools and know what to pay attention to. This requires being honest about which of your skills still matter, which ones you need to grow, and which ones you can let go of.

Remember what correct looks like. Not at the level of how to produce it by hand, but at the level of structure, tradeoffs, and consequences.

When AI produces a model, know whether the assumptions hold. When it drafts strategy, review the soundness of its reasoning. When it proposes a plan, gauge if it will survive contact with reality. These are judgment calls. They require having done the work yourself, at least once, and well enough that you have a sense for what good looks like and how it’s different from passable.

That’s the awareness from driving stick. You’re not shifting gears anymore, but you can tell when the engine sounds wrong. Letting that judgment atrophy makes you a worse driver. At that point you’re a passenger.

Forget the manual execution. The formatting, the templates, the repetitive steps that are identical every time. Those were never the valuable part of anyone’s skillset. They were the tax you paid to do the valuable part.

It’s easy to confuse deep and routine knowledge. Knowing the why behind a choice and when not to make it, versus remembering the exact steps to execute it. The first makes you good at something. The second made you faster at it in a world where execution took time. That world is shrinking.

Learn how to observe. When you did the work yourself, observation was free. Now you oversee a process that works on your behalf, and you need the instincts for it.

Start by testing the AI’s output. What happens under unusual conditions? At volume? When the assumptions change? Good professionals already ask these questions from their own work. The adjustment is asking them consistently about work you didn’t produce, at the pace AI generates it, which naturally discourages careful review.

Over time, you start reading patterns. An LLM that consistently overcomplicates things is telling you something about how it approaches problems. One that defaults to the same approach regardless of context hasn’t understood the problem. One that produces clean output with subtly wrong reasoning is the most dangerous kind. These patterns are invisible in a single interaction. They only show up after dozens. Recognizing them is a skill that develops with repetition, like anything else.

The hardest adjustment is a new category of uncertainty. When you do work yourself, you generally know what you know and what you don’t. When AI does it, there’s a third possibility: things you believe you know because the output looked right. Consistently interrogating that belief every time is the hardest part of this experience.

AI also requires more explicitness than working with people. People adapt. They read context, absorb preferences, fill in gaps with reasonable judgment, and over time need fewer instructions. AI doesn’t do this reliably (not yet). You have to state your standards and expectations clearly, every time. Getting efficient at that is its own learned skill.

Irony in automation

There’s an uncomfortable circularity here. The most important skill for working with AI is judgment: knowing what correct looks like, recognizing when something is off, understanding the tradeoffs behind a decision. Ironically, judgment comes from the experience of doing the work yourself.

This is the paradox of AI in professional work. The thing that makes you good at using it is time spent not using it.

That doesn’t mean you should avoid AI. It means you should be deliberate about when you use it and when you don’t. Try doing the work by hand. Build the judgment. Develop the feel. Once you have it, let AI handle the execution and focus your attention on evaluation and direction. The ordering here matters.*

For people who already have years of experience, the situation is simpler. You have the judgment. What you need to build now is the observation habit, and that only comes from actually working alongside AI, consistently, and paying attention to where it succeeds and where it quietly fails.

The clutch pedal is going away. Keep your ear for the engine.

*There is a tension here: Where do you spend this time without AI? Can’t be at work, tokens are cheap and speed matters. Hopefully in school, but homework rarely mimics real world experiences which become muscle memory. It seems like not everyone will have the chance to build judgment the old way. I don’t have a clean answer for how to resolve this, and I’m skeptical of anyone who claims to.