What Surprised Me Most About Working With AI Every Day

The most common questions I get from people who know I work with AI systems every day aren’t about what I’ve built or what it’s like. They’re skeptical: Will it actually scale? Can you rely on it when it matters? Doesn’t it just end up replacing the human judgment that makes work good? I’ve been in daily practice long enough now to have some honest answers — and they’re not the ones I expected.

I’ve been building with AI tools daily for long enough now that I have a before and after. There’s a version of me that thought this work would feel like ongoing discovery — like having a superpower that kept revealing new things. That version was partially right and mostly wrong. What daily AI work actually taught me had almost nothing to do with the AI.

It Slows You Down Before It Speeds You Up

Nobody puts this in the productivity case studies. In the first weeks of daily AI use, I was slower, not faster. Not because the tools were hard to learn — they’re not — but because they exposed something uncomfortable: I didn’t have clear standards for most of what I was generating.

Here’s the mechanical truth of it. You can generate 20 versions of a piece of copy in three minutes. But if you don’t know what “good” looks like before you start, you haven’t saved time — you’ve produced 20 things to hesitate over. Fast output amplifies the clarity of your criteria. Or the absence of it. And hesitation at scale is worse than slowness.

The speed only becomes an asset once you can evaluate quickly. Evaluation requires standards. I had to build mine before the speed became useful to me — and that took weeks I genuinely didn’t expect to spend. I thought I was going in to learn a tool. I came out having done a significant audit of my own creative criteria.

Most of the Work Is Editorial

Here’s what a working session with an AI system actually looks like for me on a regular day: I start with a brief. The system generates something. I read it, form a view, redirect. It generates again. I pull the one paragraph that landed, the one structural move that worked, reframe the rest, and give it back. Repeat.

That’s not prompting. That’s editing — with a very fast, very prolific first-draft writer in the room.

The AI didn’t move the creative work somewhere else. It made it more frequent and more explicit. My editorial eye got sharper because the first draft always arrived in 30 seconds and I always had to decide immediately.

What I wasn’t prepared for is how consequential that compression is. When you’re making editorial decisions constantly — dozens of times a day instead of a handful — the decisions themselves become more deliberate. You start noticing patterns in what you keep and what you cut. You start seeing your own aesthetic preferences as a system rather than as instinct. The AI gave my judgment more material to work on. The judgment got better.

This reframing matters because “prompting” is how most people describe working with AI, and it’s the wrong mental model. Prompting suggests that the input is where the skill lives. It’s not. The skill lives in the evaluation. Anyone can generate. The ability to know — quickly, precisely — what’s wrong with an output and how to fix it: that’s where the work is.

It Reveals Your Standards to You

I’ve used AI systems on projects with crisp, complete briefs and on projects where the goals were still being figured out. The difference in experience is so extreme it’s almost embarrassing to describe.

On a project with a clear brief — defined audience, defined voice, defined constraints — working with AI is fast and genuinely satisfying. You know immediately when the output is close. You know what “close” even means. You can redirect with one sentence. The system follows.

On a project with a vague brief, working with AI is a hall of mirrors. The system produces something plausible. You don’t love it, but you can’t say why. You redirect, it gives you something different. Still not quite right. You redirect again. The problem isn’t the AI. The problem is that you haven’t defined what “right” is. The AI is content to generate indefinitely. You’re the one who has to call it.

This forced me to front-load clarity on every project — not for the AI’s sake, technically, but because the AI makes the absence of clarity so immediately painful. I spend more time in definition now than I ever did before. More time asking “who is this for and what do they need to feel?” before I generate anything. More time building the brief into something that can actually hold.

The Boring Parts Are Where It Gets Interesting

I work on education products. One of my ongoing AI projects involves generating large volumes of read-aloud content at specific reading levels for early literacy — content that has to be accurate, engaging, appropriately complex, and consistent across thousands of individual pieces produced over months.

This is not glamorous work. It’s painstaking and repetitive by design. The Dragonfly literacy system I built has six characters, each anchored to a distinct developmental stage and a specific emotional vocabulary. When generating content, every piece has to fit inside that character container. Every sentence has to hit a Lexile target. Every concept has to be decodable at the right stage. The voice has to feel consistent whether it was written on a Tuesday in January or a Friday in October.

No human writer could sustain that without the system degrading. There would be drift. Characters would start bleeding into each other. Lexile levels would creep. Voice would shift across months of production. With a well-structured AI system anchored to explicit, documented constraints, we get consistency at scale — not because AI is better than a human writer, but because the system is better than human memory.

What surprised me wasn’t the AI. It was the structure. Building something rigorous enough that the AI could actually work inside it turned out to be 80% of the job — and the 80% that mattered most.

I’ve found this pattern everywhere in daily AI work: the interesting problem is never the generation. It’s the architecture. What guardrails make the output trustworthy? What criteria make it evaluable? What structure allows one person to produce the volume of a small team without losing coherence? The AI is fast at the part that used to be slow. The new slow part — the important slow part — is the design of the system it runs inside.

What Actually Changed

I went into daily AI work expecting to come out faster. Instead, I came out more structured. More deliberate about briefs. More rigorous about constraints upfront. More willing to slow down at the beginning so the middle goes fast. More honest about when a project isn’t ready to generate yet — because I’ve learned the expensive lesson of generating before the brief is solid.

I also came out with a much cleaner understanding of what creative work actually consists of. The AI does the generation — and it does it well, fast, and without complaint. My job is everything around the generation: the brief, the criteria, the constraints, the evaluation, the decision about what ships and what doesn’t. Those things can’t be automated. They’re not even close to being automatable. They are where the real work lives, and they always were.

Working with AI every day didn’t change what I value in design. It clarified it. It made visible a set of practices — definition, constraint, editorial judgment, structural thinking — that I had always relied on but never quite articulated. They were baked into how I worked without being legible to me. The AI didn’t teach me those practices. It just made them non-negotiable.

And that, honestly, is the thing I least expected: that working with a machine would make me more aware of the parts of the work that are irreducibly human. Not because AI can’t simulate them — it can simulate a lot. But because when the simulation falls short, it falls short in the same place every time: judgment. Someone has to decide what’s right. Someone has to know. That part hasn’t moved.

The AI got faster. The work got clearer. I’m not sure I would have predicted that trade.