How Google and Meta decide
There was a version of this job where you chose the keyword, the bid, the placement and the audience. You typed them in. If the campaign spent money somewhere you did not want, you could find the setting that put it there.
That version is gone, and it did not go gradually.
What you no longer choose
The bid. You set a target and a model sets every individual bid, in every auction, thousands of times a day. You cannot see the bids and you cannot overrule one.
The search. With broad match, you no longer pick which queries you appear on. You offer a starting point and the system decides what is close enough.
The placement and the creative. Performance Max assembles the ad and picks where it runs across search, YouTube, Display, Gmail, Maps and Discover. Meta does the same shape of thing. You supply the parts.
The audience. What you upload is an audience signal — the platforms use that word deliberately. It is a hint about where to start looking, not an instruction about who to reach.
Each of these was sold as a convenience, and each of them is real: the models are better at the mechanical part than a person with a spreadsheet. But together they moved the steering wheel. You are not driving any more. You are describing where you want to end up, to something that drives.
So what are you actually controlling?
What the model optimises towards, what it is able to see, and what it is forbidden from doing. That is the whole set.
Those three things are not settings you configured once and can now recite. They accumulated — from a tracking change three years ago, from a conversion somebody added for a campaign that ended, from a sales process that feeds something back without anyone deciding it should. They are instructions, and nobody wrote most of them on purpose.
What you told it counts as success
This is the objective. Everything the model does, it does in service of this, and it will be relentless about it.
If your conversion action fires on a form submit, the model buys form submits. If half of those are students doing research and the model does not know that, it will find you more students, efficiently, for as long as you let it.
And it reaches past the ad accounts. When your sales team marks a lead qualified on a hunch, and that judgement flows back as an offline conversion, Google takes it as fact and goes looking for more people like that lead. A judgement call in your CRM becomes a buying instruction with your budget behind it.
What it is able to see
Consent, tagging, what fires and what does not. Every model is working from a partial view of what actually happened, and the size and shape of the missing part is different in every company.
This is not a data quality footnote. It decides what the model believes worked. A channel that converts well but is poorly measured looks like a bad channel, and the model will spend less there, correctly, according to what it was shown.
What you told it not to do
Exclusions, negative keywords, brand safety settings, placement exclusions. These are among the last hard constraints left — the model genuinely cannot cross them.
Which is exactly why they are dangerous when they are old. A negative list written before a rebrand still blocks the terms it blocked then. Nothing in the interface will ever tell you that the list has an opinion about your current business.
You have already teleported to the top
People ask when they should start using AI in their marketing. The answer is that they started the day somebody opened their first Google Ads or Meta account.
An AI model is already deciding your bids, your queries, your placements and your creative combinations. You are running at the most automated level there is. It is just that the AI model is not yours, the rules are not yours, and none of them will tell you what it decided on. Not Google, not Meta, not TikTok.
That is the honest position, and it is not an argument against automation. It is an argument about what is left for you to do, which is to be deliberate about the three things above instead of letting them accumulate.
Who decides, on what basis, and who is responsible
That is the question this whole field now turns on, and it applies whether the decision was made by a person in a meeting or by a model in an auction.
When the client asks why the budget moved, or the owner asks whether the quarter was worth it, the answer cannot be that the system decided. The accountability did not move when the decision-making did.
What to do about it
Read what you are currently telling them. That is the audit: we take read access to your accounts and your analytics and write up what is actually in there — every finding with the thing it rests on.
Most of what we find is not a mistake anybody made. It is an instruction that was right when it was written and stopped being right quietly, on a day nobody was looking.