The pitch from every AI media buying vendor is identical: set the objective, upload the creatives, let the algorithm handle everything else. After running Advantage+ and Performance Max across a dozen accounts over the past 18 months, here is the honest picture. AI has fundamentally changed what I spend time on in media buying. It has not replaced the decisions that determine whether a campaign works in the first place.
What the algorithm actually optimises
Meta Advantage+ and Google Performance Max are both AI-driven campaign types that take creative inputs and a conversion objective, then decide targeting, placement, and bids. They genuinely outperform manually-targeted campaigns in most cases when two conditions are met: the conversion signal is clean, and the creative volume is sufficient. Advantage+ Shopping consistently beats catalog-based retargeting when the product catalog is well-structured and the pixel has meaningful conversion history. PMax outperforms single-channel campaigns when asset groups are properly organised and audience signals are provided. The performance improvement is real. The misunderstanding is in thinking the algorithm runs without inputs, and that those inputs do not require human judgment. They require more human judgment than manual campaigns did, just different judgment.
The three things AI still cannot do
First: AI cannot tell you if your offer is wrong. Advantage+ will optimise delivery of a creative to the audience most likely to click it. If no one wants what you are selling at the price you are charging, the algorithm will find this out efficiently and you will have a CPL that is technically optimised and commercially useless. The algorithm works within the constraints you set. It cannot change the constraints. Second: AI cannot interpret context. A CPM spike in October during festive season requires different budget logic than a June CPM spike after a competitor runs a major launch. The numbers look identical from inside the platform. The correct response is different. Third: AI cannot set the right conversion event. If you tell PMax to optimise for form fills and 20% of your form fills are qualified, PMax will find people who fill in forms. That is not the same as finding customers.
Where AI actually earns its place in the workflow
The operational loop that works in practice: AI reads the Meta account and flags what needs attention — frequency above 4.0, CPL more than 25% above the 7-day rolling average, ad sets with spend above the significance threshold and zero conversions. It then drafts the actions: which creatives to pause, where budget should shift, which audience overlaps are cannibalising delivery. I review and approve or modify. The AI handles detection and the structured draft. I provide business context the data cannot show: the product launch next week, the seasonal behaviour I know from three years on this account, the reason last month's CPL spike was a creative fatigue problem and not a signal quality problem. What has changed is not whether I make decisions, but how quickly I am briefed.
What this means for how you set up campaigns
AI media buying works in proportion to the quality of its inputs. Clean conversion signal — server-side CAPI, not just a browser pixel — is the foundation. If the algorithm is learning from 60% of your actual conversions, it is finding the wrong people with increasing confidence. Creative volume matters more than it ever did: Advantage+ exhausts a two-creative set faster than a manual campaign would, and it needs variety to test against. Audience signals for PMax are not optional, they tell the algorithm where to start, and starting without them is expensive. The setup work that used to go into manual campaign management now goes into signal quality and creative infrastructure. AI does more of the campaign-level work than it did two years ago. But only if you do the right things before it starts.