AI

AI earns its place in a growth stack when it removes work a person was doing slowly or badly. Not when it appears in a deck. I build the layer myself, in Python and Laravel, wired into the ad accounts and the warehouse rather than sitting politely beside them.

Where it usually starts

Almost every engagement begins with a job nobody enjoys: a Monday report assembled by hand, campaign hygiene checks that get skipped when the week is busy, creative briefs written from memory. Those are the first things worth automating, because the cost of doing them badly is already visible to everyone.

How I build it

The script does the deterministic work — pulling spend, joining cohorts, checking that events still fire. The model is handed only the part that genuinely needs interpretation: summarising why a cohort turned, drafting the next angle from what performed. Every run logs what it read and what it changed, so an automation stays auditable instead of magic.

Anything that touches budget gets a human approval step. Anything that only writes a draft does not need one. That line is where most AI projects quietly go wrong.

What you end up with

• A small set of jobs that run on a schedule without supervision
• Alerts that fire before budget is wasted, not after the invoice arrives
• Drafts and summaries that start the work rather than replace the judgement
• Code you own, documented, in your own repository

The test I apply before building anything: if this stopped running tomorrow, would anyone notice? If the honest answer is no, it should not exist.