Where AI belongs in a campaign, and where it ruins one
The honest split between the work machines do better than people and the work they cannot do at all, written by people who use both every day.
Every agency now says it is AI-native. Most mean they have a subscription. Here is the actual split we work to, written down so you can hold anyone — including us — to it.
Give the machine the grinding
There is a category of work in advertising that is genuinely mechanical. It has always been mechanical. It was simply expensive because it was done by people on salaries.
- Resizing and versioning a campaign across every placement you buy
- Cutdowns, subtitles and language localisation
- First-pass desk research and competitor scraping
- Assembling weekly reporting from four platforms
- Generating creative variants for a test matrix
- Transcription, tagging and asset management
Every hour of that used to appear on an invoice at a thinking rate. That is the saving, and it is the whole economic argument for working this way.
Keep the deciding
What a model cannot do is know which of a hundred good options is the one that will work on a Tuesday in a market it has never sat in. It has no stake, no taste and no memory of the last four things that failed.
The campaign on this site that drove three million app downloads turned on a single judgement: that the country did not need educating about crypto, it needed provoking into wanting it. That judgement is worth more than every asset produced afterwards, and no amount of generation substitutes for it.
How slop actually happens
Not from using AI. From using it at the wrong end. Slop is what you get when generation happens before the decision instead of after it — forty variants of nothing in particular, produced fast, judged by nobody, shipped because they exist.
The tell is volume without a point of view. If you cannot say in one sentence what the work is arguing, more of it will not help.
Three questions for anyone selling you AI marketing
- Which specific tasks does the machine do, and which do people do? A vague answer means nobody has thought about it.
- Has the saving reached my invoice, or your margin? Compare the fee to what you paid two years ago for the same output.
- Do I own the system afterwards? If the workflow lives in their account, you have rented a dependency, not bought a capability.
The third one is the one nobody wants asked, which is exactly why we build every system to be handed over.
Get the next one
Occasional memos on how this industry actually prices itself. No sequence, no drip, no webinar. Unsubscribe in one click.
Want this applied to your actual budget?
Build a scope and we will tell you which lines to cut before we quote them.