The bottleneck isn't AI

Happy Sunday!

I read a paper on Saturday morning that described my job back to me in language I hadn't been using yet.

It was 113 pages out of MIT and a couple of other research groups, on the economics of AI as it scales. The argument I couldn't shake: the scarce resource of the AI economy isn't who can do the work, it's who can verify the work is correct.

They call the gap between what AI can produce and what a human can actually check the measurability gap, and they argue it's getting wider, fast.

I sat with that for most of the afternoon, because if you take it seriously, it changes the trajectory of what I think 11 Agency is going to sell.

Where this is heading

For five years, I would have told you we sell Klaviyo expertise. Flows, campaigns, segmentation, deliverability, the mechanical parts of retention.

A lot of that, today, is still work the AI cannot actually do. The judgment in real time, the messy parts of getting a flow live in someone else's account, the things that don't fit a clean prompt. That is still most of the day job.

But the trajectory is hard to miss. Every quarter, more of the mechanical work gets faster, cheaper, and more automated. The piece that holds its value as everything around it speeds up is verification.

We are not fully there yet. But "we caught the thing the AI brief missed" is becoming a bigger and bigger share of what an experienced operator gets paid for.

A founder running an agentic Klaviyo brief can already ship a flow in 20 minutes that looks polished, sounds right, and hits every checkbox. Some of those flows will be subtly wrong in the way only someone who has seen hundreds of accounts catches.

The $20k attribution drop I dealt with earlier this week is the small version of that. The big version is 24 months of marketing decisions made on output nobody verified, on top of a number that was partly fictional the whole time.

The bottleneck isn't AI. It's verification.

Which means the question I keep seeing founders argue about, "will AI take my job," is the wrong question. The right one is harder.

What part of my job is verification, and is that part scaling with the AI, or is it staying flat while everything around it speeds up?

If you can't answer that out loud in one sentence, you should be a little worried. Not because AI is coming for you, but because you didn't have a clear answer before AI either, and it just made the gap visible.

What I haven't figured out yet

Two things from the paper are sitting with me that I do not have answers to.

The first is what they call the codifier's curse. Every senior operator who teaches AI to do their job is converting their tacit knowledge into training data, so every Loom, every SOP, every "here's how I think about welcome flows" is fuel for the thing that competes with you.

I have written and recorded a lot of those over five years. I am not sure how to feel about that yet.

The second is what they call the missing junior loop. If AI handles entry-level work, where do the next senior verifiers come from? The pipeline of people who can check AI output is built from people who did the junior work and learned the pattern by getting it wrong a hundred times.

Cut that loop and the verification capacity at the top eventually goes with it. Maybe not in 18 months, but in five years, definitely.

I do not have a clean answer to either of those, and anyone who tells you they do is selling something.

The bet

What I am pretty sure of is this. The agencies that compound over the next few years are not the ones using Claude faster than everyone else.

They are the ones who notice that verification is becoming the load-bearing piece, name what they verify, and start charging for it before the rest of the market catches up. Most operators have not noticed the product is changing.

So that's the question I'm leaving you with on a Sunday morning. Not as advice, just as the thing I was actually thinking about yesterday.

What part of your job is verification, and are you building more of it, or less of it? Most of us, if we are honest, are accidentally building less.

If you have a clear answer, hit reply and tell me. I'm collecting how other operators are thinking about this, and I'd genuinely like to hear yours.

Have a good Sunday. More Wednesday.

- Raymond