Agentic Ad Buying: Why I Won’t Let AI Spend My Clients’ Money (Yet).

Every ad platform is currently trying to convince you to hand over the keys. Let the AI set the bids. Let the AI pick the audiences. Let the AI decide the budget splits across campaigns in real time, no humans required. It’s pitched as the natural next step after Smart Bidding, and on paper it sounds efficient. In practice, I’ve watched what happens when nobody’s checking the AI’s homework, and it’s not efficient. It’s expensive.

Only about 32% of marketers say they actually trust AI to autonomously buy inventory right now. I think that number is roughly right, maybe even a little generous. Here’s why I’m still in the skeptical camp, and why I think that’s the correct place to be for now.

The Core Problem

Black-box bidding without oversight is how budgets quietly die

Agentic ad buying systems optimize for whatever signal you give them, and they’ll optimize for it with total conviction even when the signal itself is broken. Feed a bidding algorithm bad conversion data, or let it start optimizing toward a vanity metric because that’s what was easiest to track, and it won’t second-guess itself. It’ll just spend faster and more confidently in the wrong direction. A human media buyer at least has a gut feeling that goes “wait, this doesn’t look right.” An autonomous bidding agent doesn’t have a gut. It has a loss function, and it will chase that function off a cliff if the cliff happens to be where the function points.

I’ve seen this play out with plain old Smart Bidding, which is a much tamer version of agentic buying. Budgets quietly drift toward the easiest, cheapest conversions available, which are often the lowest-quality ones, because the algorithm was never told to care about quality, only about hitting a target CPA. Now imagine that same blind confidence extended to picking inventory, creative rotation, and audience targeting with even less human checkpoint in between.

Where I Actually Land

Use AI to accelerate decisions. Don’t let it make them unsupervised.

This isn’t a “ban all AI from your ad account” take. I use AI-assisted bidding and automation constantly, it’s genuinely useful for pattern recognition at a scale no human can match. The line I draw is between AI that recommends and AI that executes unsupervised with real budget on the line. One of those needs a human checkpoint before money moves. The other one doesn’t ask permission.

Full agentic ad buying, where the system independently shifts significant budget across channels or inventory sources without a human reviewing the logic, is the one I’m not comfortable with yet. Not because the technology is bad, but because nobody’s proven the guardrails are good enough, and “we’ll figure out the guardrails after a few expensive mistakes” is not a strategy I’m willing to run with a client’s money.

A Practical Middle Ground

What I actually recommend right now

• Let AI handle bid optimization within tightly defined guardrails you set and review weekly, not autonomous budget reallocation across campaigns.
• Audit conversion data quality obsessively. Garbage in, confidently wrong output out, every single time.
• Keep a human reviewing creative and audience decisions, even if AI is generating the first draft of both.
• Treat any “fully autonomous” ad buying pitch from a platform with the same skepticism you’d give a stranger asking to hold your wallet “just for efficiency.”

Quick Answers

A few direct questions, answered directly

What is agentic ad buying?
Advertising systems where AI autonomously makes and executes media buying decisions such as bidding, budget allocation, and inventory selection, with minimal or no human review before spend occurs.

Is agentic AI safe for ad budgets?
It carries real risk without proper oversight, since AI bidding systems will confidently optimize toward flawed signals if conversion data or goals aren’t carefully managed and reviewed by a human.

Should marketers use AI in paid media at all?
Yes, AI is highly effective for pattern recognition, bid optimization within guardrails, and analysis at scale. The concern is specifically with fully autonomous, unsupervised budget decisions.