Every agency deck now has an AI slide. Very few of them say what actually changes in the account on a Tuesday morning, which is the only thing that matters when you are paying for the work.
Here is the version without the slide. This is what we automate, what we deliberately do not, and how to test whether an agency's AI claim is real.
What models are genuinely good at here
Advertising accounts generate far more readable material than any team can read. A mid-sized account produces thousands of search terms, hundreds of ad comments, dozens of feed errors and a constant stream of performance deltas every week. Most of it is noise. A small part of it is the difference between a good month and a bad one.
Reading at that volume is exactly what language models are for. Four jobs carry almost all the value:
Overnight anomaly checks
Every account gets scanned before the working day starts: spend spikes and collapses, conversion tracking that stopped firing, ad disapprovals, feed items that lost a price or an image, landing pages returning errors. The output is a short list of things that changed and are worth a human look.
This is unglamorous and it is where the money is. A feed error that goes unnoticed until Thursday costs three days of Shopping impressions. Caught at 06:40 it costs nothing.
Search term and comment mining
Thousands of queries a week, read and grouped into themes rather than skimmed for the obvious waste. The same goes for ad comments on Meta, TikTok and Reddit, which are the cheapest customer research available and are almost always ignored.
What comes out is a negative keyword list, a set of new ad groups worth building, and a list of objections that keep appearing. That last one usually belongs on the landing page, not in the account.
Creative pattern analysis
After enough tests, patterns emerge in what works: opening on price versus opening on a problem, a face in the first frame or not, captions burned in or not. Summarising those patterns per platform, from your own data rather than from a general best-practice article, is what turns the next creative batch into something better than a guess.
First-draft reporting
Numbers assemble themselves. Pulling spend, conversions and revenue into a monthly picture is a solved problem, and there is no reason a person should spend a day on it.
What we never automate
The line is simple: a model can read and it can draft, but it does not get to spend.
- Budget decisions. Every change that moves money is made and signed off by a media buyer.
- Creative sign-off. Brand risk is not a probabilistic question.
- What counts as a conversion. This is a business decision about which customers are worth having, and it is the single most consequential setting in any account.
- Saying no. Telling a client that paid media is not their constraint takes a person who is prepared to lose the retainer.
The reason is not caution for its own sake. It is that models are confident in exactly the situations where they are wrong: thin data, unusual weeks, anything seasonal, anything where the right answer depends on context that never made it into the account.
Automate the reading. Keep the deciding.
How to test the claim
If an agency tells you they are AI driven, three questions settle it quickly.
- What did it catch last month? A real system produces a list of specific incidents with dates. A slide produces adjectives.
- What does a human still approve? If the answer is "everything is automated", ask who is accountable when the automation is wrong. If the answer is "nothing really", the AI is decorative.
- Show me the output. Anomaly logs, mined search term themes, creative briefs. It either exists or it does not.
The short version
AI in a media account is not a strategy. It is a way of removing the excuse that nobody had time to look. The advantage it produces is speed of noticing, and speed of noticing only becomes money when a person acts on it the same day.
Want this applied to your own accounts? Book an audit; the first read costs nothing.