AI media buying is a desk with better instruments, not a robot that spends your money while nobody is watching. Girard Media uses models inside a buying practice we staff. The brief, the brand risk, and the kill switch stay human.
If you want channel-level tactics on Performance Max, we already published a Performance Max optimization guide. This page is about who is actually accountable when the algorithm is in the loop.
Models do not replace the brief
Every broken "AI buying" story we inherit has the same hole: nobody wrote a brief the model could fail against. The account is a pile of campaigns named after moods. Conversion actions are duplicated. Brand terms and prospecting share a budget. Then someone turns on a smart bid and is surprised.
Before we let any system move money, we write:
- The job of each dollar (defend, harvest, or discover).
- The audiences we will never buy, in plain language.
- The creative that is allowed to spend, and the creative that is in review.
- The ceiling for a bad day, not just a bad month.
That brief is ours to maintain with you. It changes when the offer changes. It does not change because a dashboard went green overnight.
Girard Media will not "plug in AI" on top of a dirty account. We will say the account is dirty and we will price the cleanup. Cleanup is not optional if you want the models to do anything but amplify waste.
Tracking is part of that cleanup. If the pixel fires twice, if the CRM never sees the lead, if "purchase" means a thank-you page view, the model is optimizing a lie. We fix the lie first. We do not add another smart campaign on top of it and hope the lie averages out.
Where automation helps and where it burns budget
Automation is good at reallocating inside a well-defined pool: shifting toward queries, placements, and hours that already converted for this offer. It is bad at deciding whether the offer is right, whether the landing page is honest, or whether a new category is worth teaching the market.
We use automation inside the pool. We do not use it to invent the pool.
Practical split we run:
- In the model: bid, budget pacing inside a cap, creative rotation among approved assets, suppression of known waste.
- On the desk: new markets, political or medical adjacency, partner conflicts, landing-page changes, and any campaign whose failure would be a press story.
Connected TV, search, and social do not get the same leash. A misfired search query is a wasted click. A misfired CTV buy is a brand sitting next to content you cannot explain to your board. For that inventory we keep a tighter human pass. Our CTV advertising strategy is the companion for that channel.
We will not quote public cost-per-lead or return figures. Those belong in your report, against your offer, in your window. Anyone publishing a universal ratio is selling a screenshot.
The buying desk we actually staff
Girard Media's media buying is a managed service. Traders (human) set the brief and the caps. Models work inside those caps. A person looks at the account on a schedule you can see, not "when the AI alerts us."
You get:
- A weekly note in English: what spent, what it produced, what we changed, what we refused to change.
- A named person who can get on a call without a ticketing portal.
- Creative testing that is queued, not dumped. One variable when we can. A stop rule when we cannot.
You do not get a login and a prayer. If we run the buying in our seat, the seat is ours. If you leave, the service stops. Exportable reporting is yours. The live campaigns are not a souvenir.
This is the same ownership line we use for AI automation elsewhere in the stack. Instruments stay with the operator.
What you see every week
No mystery. The week has a shape:
- Monday: cap and creative check against the brief.
- Midweek: anomaly pass (broken pixels, rejected ads, inventory that does not match the brand).
- Friday: the note, and one decision we want from you only if the brief itself must change.
If nothing material happened, the note says so. Silence from an agency is usually hidden spend. We would rather send a short "held" than invent activity.
New platforms get a sandbox cap. We will not pour the brand budget into a beta because a rep said "AI native." A sandbox has a dollar ceiling, a time ceiling, and a written question it is supposed to answer. If it cannot answer in that window, it ends. That is how we keep experiments from becoming the whole account.
Creative still decides more than the bid type. We will not let a model rotate ten weak ads in search of a miracle. We cut to a small set, we write the next set from what the first set taught, and we keep brand-risky categories on a human pass. Speed without judgment is how you buy a headline you have to explain to legal.
If your current buying is a vendor dashboard plus a monthly PDF, talk to us. Bring last month's spend and the landing page it hit. We will tell you, in that meeting, whether a model would help or whether the brief is the whole problem.