Every company in our space is wrestling with what their agentic AI strategy is. The answer that seems obvious is to wire up an LLM, give it access to Meta or Google through MCP and put it to work.
The reality is more complicated, and the complications matter a lot when it’s your budget on the line.
Ask an LLM the same question five times, with the same context and instructions, and you will get five different answers. Sometimes the differences are trivial. Sometimes they’re not. These models are probabilistic by design, and that is exactly what makes them useful for open-ended tasks like drafting, summarizing, reasoning and exploring.
It’s also exactly what you don’t want when real money is moving.
I use LLMs to write, to code and to think out loud. But I wouldn’t let one publish a strategy document on my behalf without reading it carefully first, just like I wouldn’t let one file my taxes without reviewing everything it did.
The pattern is the same across these cases: You don’t hand LLMs the keys to anything with irreversible consequences without a human in the loop and without guardrails. And yet some of the current conversation about agentic advertising is proposing exactly that.
But, if you wouldn’t let an LLM publish a blog post without checking it first, why would you be comfortable letting one spend an advertising budget without the same level of scrutiny?
This isn’t a reason to avoid LLM-powered agents in advertising. They’re going to be part of how buyers interact with advertising systems, and that shift is already underway. But the industry needs to have a discussion about what the agent should be responsible for and what it shouldn’t touch.
Three lessons from running automated advertising at scale
Anyone who’s spent serious time managing automated campaigns across the major ad platforms – walled gardens and programmatic – runs into the same set of realities:
- Knowing what a platform’s API can do is not the same as knowing how to run a campaign on that platform.
Every major ad platform has its own operational rhythms, including its own preferences for how audiences are structured, how budgets are paced, how changes are sequenced and how automated bidding wants to be fed. None of this is in the API documentation.
The documentation tells you what calls are legal. It doesn’t tell you which ones are wise.
The difference between legal and wise is where campaigns either perform or waste money. Learning the wise part takes years and billions of dollars of real spend. There are no shortcuts.
- Automation only works at scale if you build governance into it from the start.
Running large volume campaigns means you can’t have a human reviewing every change. You also can’t have automation that makes decisions nobody has approved.
The only way this works is to encode the boundaries up front – what’s allowed, what isn’t, what requires human approval, what can happen automatically – and then let the system operate inside those boundaries deterministically.
The human stays in the loop on the things that matter. The system handles everything else reliably. Both parts have to be designed together or neither one works.
- The most valuable automation is the kind that takes a customer’s strategic thinking and lets it operate at scale.
Buyers know their businesses better than any platform does. They know their categories, their creative strategies, their audience priorities and important trade-offs.
The job of the execution layer isn’t to replace that knowledge; it’s to take that knowledge and execute on it across every campaign and every platform without the customer having to personally touch every campaign to make sure it’s being done right.
Human judgment in. Deterministic execution out. At a scale no individual team could hit by hand.
The future of agentic advertising requires putting the LLM where it belongs: at the interface, helping the buyer express intent, negotiating between the buyer’s goals and what the platform can actually do.
And we must make sure that when that intent is translated into actual campaign operations, the translation is deterministic, rule-governed and placed inside a governance framework that an LLM can’t silently circumvent.
This is the architecture the agentic era requires. For agencies and brands moving toward agentic advertising, the answer isn’t to build everything from scratch or to hand an LLM the keys to your campaigns and hope for the best. It’s to be deliberate about which parts of the stack should be probabilistic. And we can’t afford for the execution layer to be probabilistic.
The parts that touch real money should always sit on deterministic infrastructure that’s been doing this long enough to know how to do it right.
“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Follow Fluency and AdExchanger on LinkedIn.
For more articles featuring Eric Picard, click here.
