Every growth team I talk to right now is building an "AI growth agent."
Most of them will ship nothing that moves revenue.
I say that as someone genuinely excited about AI in growth and genuinely skeptical of the current wave of it. Those aren't in tension. The technology is real. The way most teams are pointing it is not.
The pattern
It goes the same way almost every time:
- Someone gets excited about Claude, ChatGPT, or Cursor
- Someone suggests automating growth experiments with AI
- Six weeks later there's a janky script generating forty ad variants nobody ships
- Everyone quietly goes back to doing growth the way they always did
What makes this hard to catch is that every step is locally rational. The excitement is warranted. Automating experiments is a reasonable goal. The script usually works — it really does produce forty variants. And going back to the old way isn't laziness; it's a team correctly noticing that the new thing didn't help.
The failure isn't in any single step. It's in what got chosen for automation before step one.
The visible work is not the constraint
Growth work splits roughly into two kinds.
There's visible work: copy variants, ad creative, subject lines, landing page permutations. It's legible, countable, and easy to point a model at. You can watch the output pile up.
Then there's expensive work: figuring out what to test, in what order, and why. Reading a messy result and deciding whether it's signal. Holding fifteen competing hypotheses and choosing which one earns engineering time next quarter.
Teams automate the first kind because it's the kind you can see.
I can write forty ad variants in an afternoon. I cannot, in an afternoon, figure out which of fifteen competing growth hypotheses actually deserves engineering time next quarter.
That sentence is the whole argument. The variants were never what slowed anyone down. Prioritization was — and it stays slow no matter how many variants you generate. If anything, generating more makes it worse, because now someone has to review forty pieces of copy that nobody had capacity to ship in the first place.
Where the expensive work actually lives
If you want to find it, look for the parts of your process where the work is invisible and the calendar is the bottleneck. It's usually one of these:
- Analysis — the gap between a result landing and anyone knowing what it means
- Prioritization — the standing queue of ideas nobody has ranked, because ranking needs context that lives in four people's heads
- Hypothesis generation — not "what could we test," which is infinite, but "what would actually change our mind"
These share a property that makes them unattractive to automate: none of them produce a visible artifact. There's no pile at the end. A week of good prioritization looks, from the outside, like a week where nothing shipped.
That is exactly why they're the constraint. Work that produces no artifact doesn't get resourced, doesn't get measured, and doesn't get fixed. It just quietly sets the ceiling on everything above it.
It's also where a model would genuinely help. Not by deciding for you — by holding more context than one person can, surfacing the contradiction between two results nobody reconciled, and making the ranking argument explicit enough to disagree with.
Almost nobody is building for that.
Ask a different question
If you lead a growth team, "how do we use AI?" is the wrong opening question. It starts from the tool and goes looking for a job.
Ask instead: where does our current process break down under load?
Not where it's annoying. Where it breaks when volume doubles. The answer is rarely creative production, because creative production scales with headcount and always has. It's usually a decision one person makes slowly, using context they never wrote down.
Answer that first and the AI use case falls out on its own. You'll know what to build because you'll know what's actually expensive.
Skip it, and you'll ship something that works exactly as specified and changes nothing.
Otherwise you're just automating a workflow that wasn't working anyway.