Hi Reader -

Last week I wrote about the AI creative strategy agent we open-sourced and why it doesn't replace humans.

One of the questions we got on the post was:

How do you feed back performance?

Short answer: same shape as the agent.

The longer answer is more interesting, because performance analysis isn't actually one job. It's two completely different jobs at two completely different altitudes, and most teams treat it as one and end up doing neither well.

Both altitudes can run autonomously today. We've tested both. Neither output is worth acting on without a human reading it on the way out.

Altitude 1: The funnel shape changing underneath you.

This is the chart-level view of the account. Spend freshness, format mix, message concentration, placement distribution, the way your audience cohorts are quietly drifting from one week to the next. I talked about it here:

An agent can pull all of it every Monday morning. The chart updates itself, the data part is genuinely done.

Here's what isn't done.

The chart shows your format mix drifted toward statics this week. Is that bad?

Could be. Could also be you're testing into a TikTok-style set of ads on purpose. Could be Q4 dynamics doing what Q4 dynamics do. Could be the algorithm quietly finding worse users and the slow-bleed pattern starting up again (painful story, I wrote about it here).

The chart can't tell you which of those it is. The chart just tells you the drift exists. You still have to read it.

Altitude 2: The per-ad dimensional autopsy.

This is the BruteForce side. Every ad in the account, tagged on every dimension that might matter: angle, format, audience, theme, hook, visual hook, narrative flow, problem, benefit. Plus the actual performance numbers attached to each one.

An agent can run all of that. Tag every ad in the account, surface the patterns, hand you the report. The data work at this altitude is also genuinely done.

Here's what still isn't done.

The data tells you your winners share an "anxiety, not aspiration" frame. Cool. Now what.

Rewrite all your open briefs around it? A/B it against your existing pillars? Build a Tier-2 testing batch? Three different bets, three different stakes, three different ways to be expensively wrong.

The data tells you what the winners shared. It doesn't tell you which of those three moves is the right one for your business this quarter.

That's the part that doesn't compress.

Same shape as the agent, in other words.

AI surfaces, human decides.

AI surfaces, human reviews.

AI surfaces, human picks.

We've tested running both altitudes autonomously. You can. The output looks plausible if you squint. It also turns into a folder full of expensive bad bets the second you act on it without a human passing judgment on the way out.

So yes, you can have AI run your performance analysis. What you can't do is have AI run your performance analysis instead of you. Which is the same misconception we kept running into last week, just dressed up as a different question.

If you're spending $50k+/month and want a team that actually does this work for you, both the macro signal scan and the per-ad autopsy, reply "SIGNAL" and let's talk.

Or don't. I'm sure "the algorithm will figure it out" will start being a real strategy any quarter now.
​
Later,

Shamanth

P.S. Co-hosting a live session with FunnelFox later today about the full-funnel growth. Free, live, real practitioners. June 17, 12 pm EDT / 6 pm CEST.

Register here: https://funnelfox.com/webinar-from-traffic-to-revenue​