Beyond the AI Hype: What Makes an Agentic AI Platform Actually Effective?

By Melinda Han Williams, Chief Data Scientist, Dstillery
There’s no shortage of agentic AI platforms promising to change how advertisers work. But as more enter the market, how can marketers tell which ones will actually make a difference?
At Advertising Week New York, I shared some of what we’ve learned over the past year developing Dstillery’s agentic AI platform, DS-1.
When evaluating an agentic platform, I think there are three important questions to ask:
- What does the AI know?
- What actions can it take for you?
- And does it make your results better?
Answering these three questions will help you understand the effectiveness of any agentic platform. Here, I’ll focus on how they apply to the media buying workflow.
Start with the right data
What data does your AI have access to? For media buying, this is where you’ll see the difference between LLM-driven campaigns and data-driven campaigns.
So, what is the right data to solve your problem? If you’re trying to run the best performing digital advertising campaigns, you’re ultimately trying to predict behavior: Who is most likely to be influenced by the brand’s message?
The best data for solving a behavioral prediction problem is behavioral data.
At Dstillery, we’ve spent 18 years applying AI and machine learning to predictive targeting. Our models learn from digital behavioral journeys, identifying connections between actions and understanding which behaviors are predictive of others.
Here’s an example I demoed at Advertising Week.
Imagine you’re a media buyer looking to reach new customers for a pet supply brand, and you’re looking for signals that reflect interest in pet supplies. Of course, you could ask a general-purpose LLM to suggest relevant audiences or search terms, and it would give you plenty of reasonable ideas.
When I asked DS-1 to find behaviors predictive of pet shopping, one of the search terms it surfaced was “can dogs eat sesame seeds.”
LLMs “know” a ton of information, pretty much all the information available on the internet. But LLMs don’t have access to real behavioral journeys taken by actual people. They can’t calculate the probability that someone who searches “can dogs eat sesame seeds” will then go on to buy a dog bed, because this is a behavioral prediction problem, and LLMs don’t have access to behavioral data.
“What does the AI know?” might be the most overlooked question when evaluating an AI tool. If the problem you’re tackling relies on behavioral prediction, then you need AI with access to behavioral data.
Campaigns run on more than good ideas
Data-driven insights on a brand are interesting, but an effective agentic AI platform does more than research, it also takes action.
With DS-1, marketers can move from audience planning to building, activation, and in-flight optimization. In the pet supply example, a media buyer can identify those search terms, then ask DS-1 to build Custom Search Lookalike audiences targeting those terms and send them directly to a DSP of choice for activation.
Behind the scenes, DS-1 finds the user journeys from an anonymous panel data source where someone typed in those search terms, builds a multimodal AI model that identifies behavioral patterns predictive of those search terms, and activates the resulting audience across whatever channels and activation surfaces needed for the campaign.
Behavioral data and predictive AI are necessary ingredients. To realize that value, it’s essential to have a platform that takes action in the real world, to put those results to work.
Are your results better or just faster?
This is probably the most important question, and one we’ve spent a lot of time thinking about as we’ve developed DS-1.
When we first launched the platform, I was excited about how easy it would be for advertisers to build so many more highly customized audiences. They could test more strategies, compare performance, and continually refine their targeting.
Once our clients started using DS-1 to create audiences I realized something: We were giving them homework.
We were providing the raw materials for super optimized campaigns but leaving them with the work of doing those optimizations: pulling reports, comparing results, figuring out what’s working and what to change, launching those changes, only to do it all over again.
That realization changed our approach.
Today, DS-1 monitors live campaigns, identifies opportunities to improve performance, recommends changes, and implements those changes with human approval.
We’re already seeing the impact. In a recent campaign with Canvas Worldwide, the second-largest U.S. independent media agency, DS-1-driven optimization resulted in a 38% improvement in cost per acquisition (CPA).
That’s what marketers should be looking for. Not just whether AI can make a task faster, but whether it can help them achieve better results.
Don’t overlook the workflow
There’s one more thing I’ve learned: If a tool doesn’t fit your workflow, then it doesn’t work.
I know the workflow in the demo is never anyone’s real workflow.
That’s why we’ve designed DS-1 to be customizable. Our ideal engagement with an advertiser is to build out a tailored instance of DS-1 that’s designed around their specific workflow, and our AI-native backend lets us deliver this in days, not months. Some clients use DS-1 through the UI you might see in our demo, others integrate specific capabilities into their own platforms via MCP.
MCP connections open up an opportunity that you might miss when thinking through those three questions to evaluate an agentic platform. If your current agentic platform doesn’t give you what you need for one of those criteria, you can always enhance it by plugging in an MCP that does. We’ve worked with agencies who prefer Claude as their main agentic platform, and when souped up with a DS-1 MCP, it becomes a full-featured agentic media buying platform.
Whatever the access point, remember the point of your AI platform is to remove complexity and extra steps to your workflow, so if you can’t customize it to how you work, it’s not likely to work for you.
What comes next
As agentic AI becomes a bigger part of advertising, I expect we’ll spend less time talking about what these platforms can do and more time asking what they’re actually accomplishing.
At Dstillery, we’re building toward that with DS-1. The platform brings together more than 15 specialized AI agents, can analyze over 25,000 audiences, and lets users easily define and activate custom audiences. DS-1 is proven to reduce media-buying workflows from days to minutes. But the real value of that behavioral data, predictive AI, and tools that take action are the successful campaign outcomes it drives.
For marketers evaluating agentic AI, I’d come back to the same questions: Does it have the right data? Can it actually take action? And does it fit your workflow well enough to improve your results?
Those are the questions that will separate platforms that look impressive from those that actually deliver.
You May Also Like
A Sneak Peek at the Pillsbury Bake-Off Classic, Starring Host Drew Barrymore and the Iconic Doughboy
The Pillsbury Bake-Off is making its grand return to television just in time for Thanksgiving.
Brand Control Is Holding the Best Social Ideas Back
The best social ideas leave room for audiences to remix, reinterpret and personalise them. The brands winning on social are the ones confident enough to let that happen.
The Geolocation Ban Isn’t a Compliance Problem. It’s a Technology One.
For years, precise geolocation sat near the top of every performance marketer’s wish list. It offered something most targeting signals could not: a direct link between where a consumer was and what they were likely to do next. Store visit attribution, competitor proximity targeting, location-based audience enrichment – these became core to a meaningful subset of performance and location-based advertising strategies.