Programmatic, but make it agentic.
For years, the mechanics of programmatic have been surprisingly manual, requiring long email chains and hands-on data analysis to refine audience segments before launching a campaign.
Removing that inefficiency is the vision behind DS-1, a new agentic AI interface developed by audience platform Dstillery. DS-1 allows Dstillery’s clients to develop custom audiences in minutes and push those segments to The Trade Desk for activation.
Under the hood, DS-1 was built using Model Context Protocol and can integrate into whatever platform a client chooses, including Slack, Microsoft Teams, web browsers or a brand’s own custom collaboration platform.
Keynes, a tech platform that helps brands target audiences for CTV advertising, is an early adopter of DS-1 and has partnered with Dstillery since 2019.
Slow and steady loses the race?
Keynes is “90% tech and 10% humans,” said CEO Dan Larkman, with humans “steering the ship” in the right direction if the tech is slightly off in its targeting suggestions. But there’s a lot of room for error in that 10%, including all of the manual back-and-forth communication between Keynes and Dstillery.
In the past, Keynes would compile audience data, including pixels, first-party brand data and performance data from past campaigns, which it would use to predict audience traits based on the product for sale. Keynes would then communicate those attributes to someone on Dstillery’s team.
From there, Dstillery would “seed” a model by gathering signals or behaviors indicative of the optimal customer, said CEO Michael Beebe, such as whether they viewed a specific product page.
In the past, a lot of that data came via cookies, but once Google started talking about deprecation (ha!), Dstillery shifted to anonymized data sets like opt-in consumer panels. (Later, when Google reversed course on plans to deprecate third-party cookie, Dstillery brought both kinds of signals into a multimodal AI model that was able to process and connect the data across formats, Beebe said.)
After pulling the data together, Dstillery would then send the curated audience back to Keynes, which would inform Dstillery of any tweaks that needed to be made.
Inevitably, some of those changes would get lost in “a game of telephone,” said Larkman, which could be very disruptive. Something as small as getting the targeting age slightly wrong could set a campaign back several steps, from when the mistake was caught until it was later communicated and revised.
Finally, Keynes would be able to activate the audiences within its DSP partner and launch the campaign.
From brief to launch, this process took a minimum of 48 hours, said Beebe, and often closer to a week.
If it looks like a chatbot …
With DS-1, however, the process is condensed to a matter of minutes.
The LLM-powered interface has a direct integration with Slack and functions not unlike “a Slack channel that is a chatbot, essentially,” said Beebe. (Though he prefers you not call DS-1 a chatbot. It’s just a conversational AI interface, he said. Whatever floats your boat.)
Keynes can query the model just like it would a standard LLM chatbot by inputting its desired audience characteristics, said Larkman, like age, income and interests. DS-1 will return prospective audiences that Keynes can revise and re-prompt if needed.
For instance, Larkman said, if Keynes is working on a campaign for a brand selling sports gear, DS-1 might suggest audiences that index highly against Nike shoppers.
But it can also get more specific. Say the campaign is focused on tennis gear. Keynes could tweak the audience to only include people who are like Nike tennis shoppers, rather than looking at the brand’s entire customer base.
Keynes can push the resulting audience or audiences directly to The Trade Desk via preconfigured DS-1 connections.
That same setup also allows Keynes to test audience variants – like slight shifts in age or income – over much shorter timeframe, speeding up the trial-and-error process.
With a human at the helm, Larkman said, it wasn’t feasible to test as many versions. “It’s so much extra work” to track those small refinements, he said, that leaders often don’t want to ask that of their team.
But by removing the “emotional layer,” Larkman said, more work gets done without eating up hours, or even days, of human time. No one feels guilty adding to a machine’s to-do list.
