No matter how bullish you are about artificial intelligence, there are some things that current AI models just can’t do. Like successfully run a company with an all-agent workforce or figure out when it’s fed made-up idioms as a joke, to highlight a few recent news stories.
But as many ad tech veterans have known for years, machine learning tools can yield pretty consistent results when it comes to analyzing and optimizing large amounts of data.
CTV outcomes measurement company EDO is no stranger to large data sets. Since its founding 10 years ago, the company’s own data set has grown to include information related to 20 million smart TVs, 300 million TV airings, and a whopping 102 trillion TV impressions. (Some context: 102 trillion seconds is about 32 million years.)
And those trillions of impressions bring along other data points, such as geography, time of day and even the type of platform – an SVOD platform versus a cluster of FAST channels, for example.
“Every impression has got some expected outcome on it in our measurement, and so we’re able to feed that into these algorithms and basically train these AI models for any given advertiser in a given category,” EDO CEO and President Kevin Krim told me.
So it’s no wonder that EDO is all in on what Kevin calls “vertical AI,” meaning automated tools that are designed for ultra-specific use cases within the ad tech industry.
I sat down with Kevin in March to get some more of his hot takes about how AI will continue to transform programmatic CTV advertising.
On the future of AI in CTV: “We just think that the scale of streaming TV is such that automation is inevitable. Because AI models can take into account all these variables that there’s no way we can keep track of in our brain or our spreadsheets.”
On what AI is good for: “Sometimes it’s just prohibitively expensive to do certain things with human eyes and hands, but when you apply models to it with the right data, the economics become pretty compelling.
“For example, we’re scanning all of linear TV and all of streaming ad-supported TV to see what new creatives are being run, so we can add them to our database. If we try to put people on that – first of all, who would really want that job? Second of all, it’d be just too expensive. We’d have millions of people watching every hour of everything, all the time. It just couldn’t be done.
“But you give that to computers with decent AI models, and it’s instantaneous. It’s really not expensive at all, and the accuracy levels are way higher than human beings could do.”
Why “vertical AI” is superior to off-the-shelf AI models: “What we’re hearing and learning from all these large language models is they’re only as good as what they’re trained on. And short of stealing people’s copyrighted material, when you have a proprietary data set that you can train on, it’s a big advantage.”
On who (or what) makes the final decisions: “In our world, there’s still going to be a final human check on a lot of the data to make sure that we’re comfortable with the quality. Because if some new ad format pops up that it’s never seen before, by definition, it’s not going to handle it well, right? It only knows what it’s been trained on.”
Answers have been lightly edited and condensed.
Questions? Thoughts? Let me know what you think of the newsletter at victoria@adexchanger.com.
For more articles featuring Kevin Krim, click here.
