It’s Just Work: The Part of AI No One Likes to Talk About

A mentor of mine likes to use a phrase that has stuck with me: it’s just work.

Lately, I’ve been thinking about that phrase a lot, especially as I scroll through the constant stream of posts about AI effectiveness. Every day, there seems to be a new statistic about who’s using AI, how small businesses are using it, and how quickly it’s transforming the way we work. I’ve commented on quite a few LinkedIn posts about this, and I keep coming back to the same idea:

One of the biggest misunderstandings about using AI effectively is that people forget it’s still work.

It’s not a magic pill in which work goes away and machines do everything.

It still requires work.

But then again, it’s just work.

People love AI because it feels fast. It gives quick answers. It creates the impression that hard things have suddenly become easy. And to be fair, that appeal is real—I like that part too.

But when you move beyond casual use and start talking about agentic workflows, about agents actually doing meaningful business tasks that humans used to do, the gap between “easy output” and “real implementation” gets very wide, very quickly.

In fact, the word workflow can be misleading. It sounds neat, linear, and simple. In reality, building useful AI systems is none of those things.

It starts with a better question.

Before you build anything, you have to ask the question: What problem am I actually trying to solve? What value am I trying to create? How am I improving my business, my sales process, or my marketing motion? What specifically is broken, inefficient, or underperforming?

That kind of thinking matters because there’s a major difference between automating an existing human workflow and rethinking a process around what AI can uniquely do.

A lot of people define AI workflow as simply documenting what a human already does and assigning that process to an agent. That approach feels tangible, which is why it’s so popular. It’s easy to understand. Easy to explain. Easy to sell.

But it’s also limited.

The reality is that agents can do some things humans cannot. Humans can do some things agents cannot. The goal is not always to copy the human process exactly. Sometimes the better question is whether the process itself should change.

Take personalized outreach as an example. Mass personalized email is not something humans do especially well at scale. We rely on templates, shortcuts, and repetitive messaging. AI can often do that part better—at least in theory.

Of course, that can also lead to a flood of terrible emails, which we’ve all seen. I set up a new email account recently, and the amount of spam hitting it was awful. Most outreach is still lazy, generic, and painfully easy to ignore.

So the better question isn’t just, “Can AI send more email?” It’s, “What problem are we solving with outreach in the first place?”

Maybe the real issue is that the messaging isn’t personalized enough. Maybe it isn’t relevant enough to the audience. Maybe it doesn’t offer enough value to earn attention. AI can help with that, but only if you’ve thought deeply enough about the problem.

And that’s where a lot of teams stop too early.

In sales and marketing, people often say they want to solve “outreach.” But that’s too broad. “How do we reach people?” is a generic problem. The more useful question is: How do I add value to this specific niche, this specific buyer, this specific moment?

That takes thought.

That takes strategy.

And yes, that takes work.

But then again, it’s only work.

Once you’ve identified a smart problem to solve, the next step is building the solution. This is where the romantic version of AI starts to fall apart a little.

Because now you have to choose tools. And those tools cost money.

This is one of the uncomfortable truths in the AI conversation: meaningful implementation usually isn’t free. I’m not even talking only about LLMs. I’m talking about the integrated tools, platforms, connectors, and systems that make a real workflow function. You are going to pay for them, so you need to be deliberate.

You have to ask yourself: Does this tool actually solve the problem? Is this the smartest way to solve it? Is the stack I’m building efficient, or am I just layering technology on top of confusion?

There are plenty of ways to build a stack, but it is surprisingly easy to end up spending more on tooling than you would on human effort if you aren’t careful.

So now you need more than a good question. You need a solid plan for the steps, the stack, the context, and the logic behind how the system will operate. You need to understand what inputs matter, what success looks like, and how the system will improve over time.

Then comes the hardest part: actually finishing it.

But then again, it’s only work.

This is where people, especially those supervising the work, underestimate the difficulty by the widest margin. Once you get into the tools, you get pulled into credentialing, integrations, permissions, backend communication, broken handoffs, and all the messy realities of getting systems to talk to one another.

That last mile is rarely clean.

If you’re a CEO or business owner looking at AI from 30,000 feet, it can be tempting to think, “How hard can it be? AI tells you what to do.”

But that’s not how it plays out in practice.

AI might point you in the right direction, but then it fails. You try again. It fails differently. You refine the prompt. You send the error message. You explain the problem again. You test another approach. Maybe you add screenshots because visual context helps, but now you’re burning usage credits while troubleshooting.

This is what implementation actually looks like: iteration, correction, re-explanation, and persistence.

You have to get it across the finish line. You have to make the workflows real. You have to make the agents operational.

And that work is substantial.

But then again, it’s only work.

A lot of owners and executives don’t fully appreciate how much effort this takes because they are not the ones in the weeds doing it. Their own use of AI is often relatively simple and high-level, which makes the deeper complexity easy to miss. But once you’re connecting multiple systems and expecting agents to execute across them, the work becomes very real.

And even then, you’re still not done.

Because after “the work,” there’s another truth:

It’s just more work.

You don’t build the system, turn it on, and declare victory. That would be like posting once on social media and concluding that your content strategy failed because one post didn’t perform.

It doesn’t work that way.

Once the system is live, now comes the part I actually like best: evaluation.

You let it run. You watch what happens. You study where it failed. You identify what worked. You look for bright spots. You resist the urge to blow everything up the first time it underperforms.

That restraint matters.

Too many people try something once, decide “it didn’t work,” and move on. But effective AI implementation—like effective marketing—is usually the result of thoughtful iteration. You make one small change. Then you run it long enough to see whether that change helped. Then you make another adjustment. Then another.

This is not glamorous work. But it is the work.

And it’s the only way to make agentic workflows actually effective.

So when people say, “AI can just do it,” I think what they’re often missing is that the technology does not replace strategic thinking. It doesn’t replace judgment. It doesn’t replace understanding your business well enough to define the right problem. And it definitely doesn’t replace the patience required to test, observe, evaluate, and improve.

That’s still on us.

In the end, I think many people are not failing with AI because it’s too hard. They’re failing because they’re not patient enough—or resilient enough—to do the work.

But this isn’t hard labor. It is focused, time-intensive, occasionally emotionally draining work. The kind of work that requires you to keep going when the first version breaks, when the system doesn’t behave the way you expected, and when the promise of “easy automation” turns out to require real operational discipline. Then the system works but the results aren’t there yet.

That is what effective AI use actually demands.

It’s just work.

But it’s still work.