AI for marketing: agents on my own marketing data
Why I am not learning AI in marketing as a text machine but as a colleague who reads real numbers, and why your offer may sit exactly there.
01
The moment
October 2026. The first campaigns of my Meta ads experiment are live, and the log of what Claude changed, and why, grows every day. No ad-management subscription, no optimisation platform. Claude Code creates the ads, sets them live, reads the results and optimises for leads.
It started with a question you probably know from your working day. Why do so many businesses pay every month for tools meant to manage their advertising, when the real problem is almost never the tool? The campaigns run, the numbers are there, and still on Monday morning nobody knows exactly what worked last week.
I tell you this because you probably sit on the other side of that question. You are the person who collects the numbers from several tools on Monday, copies them into a deck and is then asked why the cost per enquiry went up. That work is changing right now. Whoever understands how early has something to sell.
02
What went wrong
What goes wrong at the start goes wrong for almost everyone. People think AI in marketing means text. That is the most visible part, and the part that loses its value fastest. When everyone can produce decent ad copy in seconds, decent ad copy is only the baseline. Whoever builds their independence on writing copy with AI faster than others builds on ground that is sinking right now.
The second mistake is subtler. You give an agent access to everything and expect it to handle it wisely. An agent looking at scattered, contradicting numbers guesses. It guesses politely and convincingly, but it guesses. When the ad account says one number and the website another, each tool is right on its own, and together they make no sense. Only someone who decides what each number means can resolve that.
And the third mistake: you start with collecting instead of with the questions. Dumping everything into a database first and then seeing what to do with it almost always ends in a data graveyard nobody opens again. The right order is the reverse. Questions first, then the numbers that answer them, then the agents.
03
What I tried
The Meta ads experiment is running, it is early, and I deliberately do not write in advance that it works. How much budget it has moved so far and how many leads came out of it is not solid enough yet to write down here. As soon as there are numbers I can stand behind, they go on the experiment page and as an entry in the changelog here.
Next to it I am building the foundation agents need to work sensibly at all: a marketing data warehouse in BigQuery. For a small business it is surprisingly cheap, because you only pay for what you store and query. Under Do you will find how I turned it into an offer.
04
What I know now
What I know now fits in one sentence: the AI is replaceable, the data is not. Next year there is a better model. Whoever ties their work to one tool starts from scratch every time. Whoever has put a business’s data in order puts the new model on top and is further along at once. For you that is good news, because putting things in order is work you already half do today, only without anyone paying for it.
I also know how narrowly an agent’s job has to be framed. An agent that writes every Monday what changed last week and why. An agent that checks campaigns and says where money flows without results. The narrower the job, the more reliable the result. An agent that is supposed to do everything ends up doing nothing reliably. You know that from every good brief.
I know where the human has to stay. An agent can pause a campaign. Whether it may do so on its own is a question of trust, and with software you build trust in small steps. With a small budget it may do more, with a big one less. Drawing those lines and answering for them is a job businesses will need someone for, someone who understands marketing and can lead agents.
That shifts what people get paid for in marketing. For a long time you were paid for knowing how campaigns work. That knowledge now sits in the model. In future you get paid for building the system in which that knowledge works with one particular business’s numbers, and for making sure it runs. If you are wondering what you could sell after leaving: for this offer you need no agency and no team, only judgement and a few clean decisions.
Small businesses have a real advantage here for the first time, and so do you if you work for them. A corporation needs a project with a budget and a timeline for a warehouse. A small business needs a few clear decisions and a few weeks. The tools are the same.
05
Open question
How much freedom do you give an agent that spends money? I do not have a rule yet that I would recommend to everyone. Right now I decide case by case, and that does not scale. When I have an answer, it goes here as a new changelog entry.
06
Changelog
First version. Nothing here is rewritten silently: every later change gets its own entry in this changelog.