The one thing AI can't automate

Why you're getting this: You've interacted with Haven or AdTribe before, and this is my weekly-ish update on my journey to building an AI-native agency. Zero pressure to stick around, just click unsubscribe if you don't want to get it.

Alright, let’s get into it.

I ran a company townhall this week. One question I kept coming back to: as AI gets better at execution, what actually becomes more valuable?

Judgment.

Clients used to pay us to know things they didn't. Now they open Claude and ask "why is my CPA going up" or "what should I test next."

If knowledge is commoditizing and execution is automating, the only thing left is the ability to read a situation and make the right call.

Which ad to kill and when. Which test result is signal vs noise. Whether a 20% CPA drop is a win or a trap. You still need someone with experience to figure that out.

The people who are fast at execution but mediocre at judgment are going to get squeezed. The people with genuinely good instincts are going to be worth more than ever.

Here's what we're actually changing: we're asking everyone on the team to log their hypotheses before running a test. Before you launch anything, write down what you're testing and why you think it'll work. A specific offer, a new creative angle, a different audience. What you expect to happen.

AI can already see every test we've ever run and every result. What it can't see is why we thought it was worth testing in the first place. Right now that lives in people's heads and disappears when they close their laptop.

Every logged hypothesis becomes part of our AI brain. A new media buyer joining two years from now will have access to every bet this team ever made and the thinking behind it.

We've structured media buyer comp the same way for a while now. 30% base, 70% rev share. The idea being: you eat what you kill, everyone's aligned with the business.

I'm starting to think it's actually backwards.

Here's the thing I kept glossing over. Media buyers don't control how many accounts they have. I do. My sales pipeline does. So if I lose a client, or go through a slow quarter, their income takes the hit through no fault of their own. Their rent is tied to my ability to close deals, not their actual performance. That's not a great deal for them and I'm not sure it's motivating them the way I thought it was.

What really made me rethink it was watching the team over time. The best people on our team don't work harder because of the bonus. They just work hard. It's who they are. You could cut their rev share and they'd still be the first ones to flag an underperforming account on a weekend. The bonus doesn't create that. It just rewards it after the fact.

Which got me thinking: if the bonus isn't what drives the people I actually want, what's it really doing? And for the people who do need financial pressure to stay engaged are those even the right people to build around?

Where I'm landing: flip it. Much higher base, smaller rev share on top for the people who want upside. Give everyone stability first. Let the bonus be a reward, not a lifeline.

Haven't pulled the trigger yet. Still running the numbers. But the 30/70 split probably isn't surviving the year.

We onboard 2-3 new ad accounts every week. For years that meant the same 2-hour checklist every time. Create the Slack channels, set up the GitHub repo, build the Asana project, spin up the Drive folder, fill in all the Airtable rows. All manual. Every single time.

We built a tool that kills most of that. When a new client gets added to Airtable, a message pops up automatically in our team's Slack with a link to the onboarding console. The team lead opens it, types the brand name, picks the platform and media buyer. About 20 seconds of actual work.

Then everything else spins up on its own. Two Slack channels get created, one internal and one for the client. A GitHub repo with all our templates. An Asana project with every task and the right team members already assigned. A Drive folder with all the docs. All the Airtable rows. The only thing left to do manually is invite the client to their Slack channel.

What used to take 2 hours now takes 20 seconds of human input.

The bigger win isn't even the time. It's that the process is now the same every time, regardless of who's running it that day.

Talk soon,

Dmitry