AI Scales Expertise. It Doesn't Replace It.

When I first got into Clay, I asked myself the question everyone asks. What can I do now that I couldn't do before?

It sounds like the right question. It's the whole promise of these tools. But it took me down some genuinely weird rabbit trails. I kept building automated loops in domains I didn't fully understand. The automation worked, technically. The output was mid, and I couldn't always tell you why, because I'd never done the underlying work by hand.

That was the inefficient version of my learning curve. Here's the lesson I'd give anyone earlier on theirs.

The wrong question and the right one

"What can I do now that I couldn't before?" is a trap. It points you at work you've never done, in domains you don't understand, and hands you a machine that produces that work at scale. You end up scaling output you're not qualified to judge.

The better question is the boring one. What am I already doing, or what does someone already know cold, that AI makes dramatically faster?

A friend put it to me as two categories:

1. You've never done the role and you think AI gives you the role. You're going to fail.
2. You've done the work before, somewhere, sometime, pre-AI, and now AI makes you faster. That works.

For the speed to be real, you needed to have been doing the thing anyway before AI showed up. The speed only matters if the underlying work already exists. "Create ad variants at scale" is a real unlock at a company that was already producing variants. At a startup that never needed variants, it's a machine that makes nothing faster.

There's a corollary I missed at first. You also need to know what good looks like. Not in theory, but well enough to spot a bad output on sight. Otherwise the confident, well-structured, entirely wrong version sails straight through.

Why the people winning at programmatic actually win

The clearest proof I've seen is Chris Shuptrine, VP Marketing at Torii, who demoed his programmatic SEO system at a Demand Collective roundtable. Around 300 listicles published over four to five months in an IT-automation niche, roughly 4x blog traffic, from about 500 to about 2,000 visits per week.

The factory is impressive. A ticket holds the program, Python handles the deterministic parts, and the LLM writes and revises against guideline files that get smarter every time something ships bad. But the factory isn't why it works.

It works because Chris has done each phase of the pipeline himself. He knows what a good listicle looks like because he's written them. He knows Google only indexes about 10 to 15 articles a week from a typical domain because he's run into the ceiling. He knows LLMs fail at randomness because he caught the failure, since asking an LLM for a random number gets you 6 or 7 nearly every time. Every constraint encoded in his system is a scar from doing the work.

The automation is downstream of the expertise. That's the order. You can't run it in reverse.

So I stopped trying to replace our SEO agency

Here's where this got practical for me. I started building a programmatic SEO play at Vector and ran straight into my own version of the problem. I'm strong on content, context, and workflow. I'm not as strong on technical SEO, meaning schema, metadata, and how pages should be structured inside the CMS. I could have AI'd my way through it and shipped something that looked right but was wrong in ways I'd never catch.

The old version of me would have ground through it solo, because the implicit goal was to prove I could replace the agency. That instinct is everywhere right now, and it's mostly wrong.

Instead I went to an agency and said I'm building this thing, I'm not firing you, I'm asking you to build it with me.

So now we share a repo. I commit PRs. I handle the content, the context, the brand voice, the topic clusters, and the workflow that generates briefs. They handle the technical schema, the page structure, the metadata. We launch it together.

What unlocked it was a reframe. I don't actually care about replacing anyone. I just want a programmatic play we're not running. We weren't running programmatic at all. So why wouldn't I do this with the people who already have the expertise I'm missing?

The expertise doesn't care whether it lives in your head or your partner's. The question is whether it's in the system.

Document the judgment, then scale it

There's a sharper version of this idea. The lasting value of an expert, whether that's an agency, a designer, or a senior operator, is front-loaded in their judgment. Which fonts to use, what the brand is, how to position. Historically you paid for that judgment monthly, forever, because it lived in their heads and got re-applied by hand every time.

Now you can document the judgment once and turn it into a system that scales output from it. Chris's revision-guideline files are exactly this. Every time output comes back bad, he asks the model why, then updates the guideline file so that failure never recurs. The guideline files are the product. The articles are just what the product emits.

That's what the agency partnership really is for me. We're not buying their hands. We're encoding their judgment into a repo we both commit to. After that, the system carries it.

How to run this on your own stack

1. Inventory what you actually know well enough to judge. Not what you're interested in. What you've done enough times to evaluate output on sight. That list is where your AI leverage lives.
2. For everything else, find the person who has the expertise and build with them instead of around them. Shared repo, shared system, clear division of labor based on who knows what.
3. Encode judgment as you go. Guideline files, correction logs, documented standards. If a quality call only lives in someone's head, the system doesn't have it.
4. Treat "AI lets me skip learning this" as a red flag. It usually means you're about to scale output you can't evaluate.

I took the inefficient learning curve so you don't have to. Ask the boring question. It pays better.