AI in the marketing stack

The honest summary: production got much cheaper, judgement did not, and the discovery layer is moving. Most of the disruption is in the first; most of the value is in the second; most of the risk is in the third.

Production is cheap now

Drafting, variants, resizing, translation, summarising research — all collapsed in cost. This is genuinely useful and mostly boring. The teams getting value are using it to remove the work nobody wanted to do, not to replace the work that required a decision.

The failure mode is obvious in hindsight: publishing more because you can. Volume without judgement produces commodity content, and every distribution system — search, feeds, inboxes — is being tuned to discount exactly that.

Judgement did not get cheaper

What to say, who to say it to, what to promise, whether the result was incremental. None of that is a generation problem. A model can write fifty variants of a headline; it cannot tell you that the position is wrong, and it will happily write fifty variants of a wrong position.

This is why the fundamentals in this guide got more valuable rather than less: they are the part that does not automate.

Discovery is shifting

Answer engines and assistants increasingly sit between a person and a source. That changes how demand is created and captured — a summarised answer with three citations is a different competitive surface from ten links.

It is early, the numbers are contested, and anyone selling certainty here is selling. What is safely true: being genuinely useful and clearly structured helps in both worlds, and being a commodity hurts in both.

Agents that act

The newer development is agents with write access — to ad accounts, to publishing, to CRM. That moves the risk from "wrote something bad" to "changed something live". The control that matters is not the prompt; it is scope, change history, and knowing who or what made a change.

What to actually do

• Use it to remove production drag, and keep an editor.
• Do not increase output just because you can.
• Give agents narrow scopes and check the change history.
• Keep measuring incrementality. Faster production of unmeasured work is not progress.

What trips people up

• Treating AI as a strategy rather than a cost change.
• Publishing at scale into systems built to detect scale.
• Assuming your category's discovery has shifted as much as the commentary suggests.

You have got this when

You can name which parts of your own marketing got cheaper, which got no easier, and what you changed as a result.