Google AI Max vs Agent Planners: who decides the budget the automation spends?
AI Max is Google's most complete automation yet: keyword targeting gone, Gemini matching queries to landing pages, controls layered on top rather than a product you opt into. Assume it beats manual optimisation, because it usually does. The question it never asks is the one that decides your quarter — should this campaign, this channel, this product have the money at all.
Take AI Max seriously
The reflex among ad-tech vendors is to imply Google's automation is a black box you need saving from. It's more useful to be honest: AI Max has signal you don't — full query streams, cross-property behaviour, real-time auction dynamics — and it acts on them per impression, continuously. On the job of spending a given budget well inside Google, you will not beat it with rules.
The question it structurally cannot ask
AI Max optimises toward the goal you set, using the budget you allocate, on Google inventory. Every one of those inputs is given to it. It has no mechanism to conclude that the budget should be smaller, that the goal is measuring the wrong thing, or that Meta deserves the next increment instead — and no incentive to, since all three conclusions move money off the platform.
That isn't a criticism. It's the boundary of what an in-platform optimiser is for.
Side by side
| Decision | AI Max | An agent above it |
|---|---|---|
| Which query matches this landing page? | Yes — Gemini, per impression | No |
| What bid clears this auction? | Yes | No |
| Which creative serves here? | Yes | No |
| Is this campaign's target CPA the right target? | Optimises to it; doesn't question it | Checks it against your real margin |
| Should Google or Meta get the next $10k? | Structurally can't — sees only Google | Reads both and compares |
| Is this campaign scaling a low-margin product? | No view of your margin | Yes, with COGS connected |
| Did conversions actually stick after refunds? | Sees platform-attributed conversions | Joins your own settled revenue |
The failure mode automation makes worse, not better
Better optimisation toward a wrong objective gets you to the wrong place faster. If your target CPA was set against revenue rather than margin, AI Max will efficiently scale whatever hits that CPA — including the product you lose money on. The automation is working perfectly; the objective was wrong, and nothing inside Google will tell you so.
That's the specific gap worth closing: not the optimisation, the objective. See optimising on profit instead of ROAS for what it takes.
The honest recommendation
- Use AI Max. Fighting Google's automation with hand-tuned rules is a losing position.
- Own the layer above it — budget allocation across channels, whether the target reflects real margin, and whether the automation is scaling something you'd rather it didn't.
- Keep a governed write path for the changes you do make, so an account with several people touching it stays explainable.
Frequently asked questions
- Should I use Google AI Max or a third-party AI tool?
- Both, at different layers. AI Max optimises delivery inside Google better than any external tool can. An agent works above it — allocating budget across channels, checking whether the target reflects real margin, and joining results to revenue Google can't see.
- What can't Google AI Max do?
- It can't question the budget or goal you gave it, can't see other platforms, and can't see your product margin. All three of those conclusions would move money off Google, so no in-platform optimiser is positioned to reach them.
- Does AI Max still use keywords?
- AI Max moves away from keyword targeting, using Gemini to match user intent against your landing pages instead. Advertiser controls are layered on top of the automation rather than replacing it.
- Can automation make things worse?
- Efficient optimisation toward a wrong objective reaches the wrong outcome faster. If your target CPA was set against revenue rather than margin, the automation will scale the low-margin product that hits it — working exactly as designed.