A (6:59)

I would have loved to see those, of course.

C (7:01)

Yeah, yeah, yeah. And are tightly governed under these protocols, which we can get into. And then another seller agent that sat over top of our freewheel ad server, which could pull digital impressions and bring all those together in the digital product catalog. Right, and bring all those together. And then on top of that, I think what was really exciting was we have APIs into a forecasting service that we have. So we have built in our own planning tool, a model that says if I buy this many digital impressions and this many linear impressions, ping my identity graph. And NBC has been investing in their identity graph now for many years. And give me a household reach forecast. If I divide the media plan by X amount on linear and Y amount on digital, and If I do 50, 50, how many households do I reach? If I do 70, 30, if I do 80, 20. And all of that is answered immediately by the agent. So you get not only the media plan, but the forecast and you can play with it. And we proved all of that could be is doable inside this poc.

A (8:12)

Ryan, just to make sure I'm following correctly, so these seller agents that you all had, you had one on the traditional TV side that was able to see, okay, do I have a spot in the third quarter of Sunday Night Football? What kind of ad format is supported for that? How much would I want to charge for that? How many people do I expect to be watching at that point? I'm going to package that up. Similarly, Free will had the one for like the peacock stream able to package up a lot of that same data. And then that got combine to also be able to reconcile, okay, between the linear broadcast and the peacock stream, how many people would that reach? Which is stuff that normally a human would be able to do, but this is a very particular kind of human who loves living in spreadsheets. Also humans, we can be a bit slow. It'll take us time to process all that data. Whereas AI agent, you give it the same data and it's able to package all of that up into. Okay, now I understand what I'm selling, what I have to offer to an advertiser here. Is that correct?

C (9:25)

Yeah, it's very well articulated. I would say on top of that, you had said that like a human does this in Excel we have a planning dashboard where you could put that data in and in that application it would give you a cross platform forecast. Okay. That's a traditional kind of web application for planning and the buy side has web applications and the sell side has web applications. What's interesting is that we're using the service that sits underneath that application via an API into an MCP server, which we can talk about MCP to provide the same answer, but just to an agent. So one of the taglines we've been kind of kicking around is AI is the new ui. And what I mean by that is that we're actually not creating something net new under the hood. The identity graph, the forecasting model, the product catalog, all of those things are built into traditional applications, databases and servers that have APIs on top. What we're essentially doing is making those APIs available to these MCP servers so that agents can just ask the questions of those applications. The answers about how many people am I reaching? Or what is the product name is surfaced through an AI chat agent versus through a traditional web dashboard that you would look at, right?

A (11:07)

Yeah. It's kind of like the idea that an MCP server, which stands for Model Context Protocol, is like the agentic AI version of a website that you would access through a web browser. So if you think about a website as we see the user interface of digitaday.com and you see all the articles and the very pretty art and everything arranged in a way that my human eyes can make sense of underlying that is just a bunch of code and a bunch of data that's pulled from various different sources and it gets just presented using HTML, CSS, JavaScript into the user interface that we all humans are able to see. MCP effectively allows all of that to get repackaged, but in a way that an AI agent would be able to understand. They don't need to see it, they don't have eyes.

C (12:04)

It's such a great analogy for I think the listeners, because you're. Everyone's experiencing this when they use ChatGPT. So let's stick on your. Your example. So Digiday has a website which is packaged in HTML and it's designed and you can go to Digiday and you can read, you know, Tim's latest article. Depending on that web publishers are making these decisions, right? Depending on the access that they're giving to AI agents to kind of crawl those websites. Another way I might understand about Digiday is just go into my ChatGPT or one of these chat AI models and just ask a question about advertising or digiday, and the answer is presented to me by inside the chatbot. But it's using the data from your website. So it's, you know, the way that I'm consuming it is through this agent, but the underlying data is being pulled from the source website. The same thing is true in kind of B2B and not just for web content, but for services like forecasting models or product catalogs or all these other things. Right. You know, if you go way back like, or not even way back, like even today, you know, one way to do that, the most kind of primitive way, is you call sales rep and say, what products do you have available? Like, that's an interface. Right. And then I think as we've tried to automate the business, there's now integrations between buy and sell side systems where some of that data is available via API into buyer planning platforms. Right. And so that was a new interface and now this is still another new interface and I would argue kind of the emerging UI experience where the buy side may just be in an agent and asking these questions and the answers are coming to them through the agent versus through a phone call, an email, a fax, or even a dashboard, which is the more modernized version of it. Does that make sense?

A (14:12)

Yeah, basically, ultimately what matters here is it's how the data is getting repackaged. Historically, data has gotten repackaged and visualized for humans. Now there's all this work that's going on to repackage that data and make it accessible for AI agents. And effectively that's the foundation of this.

C (14:32)

Yeah. And actually you call up something that I think is so critical because the place where the most amount of work needs to be done is in the data foundations. Like you need clean data. Like your product catalog needs to be well organized. Like you need to have like your data in order. You need to have these services available. Otherwise it's sort of garbage in, garbage out. I was actually just having this conversation earlier with a friend of mine. It's like it's so easy to build an agent now using natural language like vibe coding and stuff like that. But if you go ask that agent to fetch something, if the underlying data is not good, the answers you get back are not real. Right. And so I actually think that the biggest amount of work that we need to do is still in the sort of the old school realm of data cleanliness and data standards. So that these agents, especially as they ask questions across different types of publishers and platforms like can all kind of make, make sense of it. So I think as much as we're interested in building the agents, we're, we have to continue to be very focused on making sure that the underlying data sets are super accurate so that the.

B (15:51)

Agents don't hallucinate building on, on top of that. Right. Because I would imagine based on what you just said that, you know, a media buyer would still say, well, I kind of want to have my hands in the pot if that is still the case and there's still a ways to still want some human oversight here. So how much control does a buyer really have once an agent is optimizing in real time with this product?

C (16:14)

Yeah, I mean, listen, I think the humans are very involved and I think will be very involved for a long time. These are not like, these are productivity tools that speed up access to information and allow you to ask questions of information and get things in a much faster way. But you are still very much going to have humans in the loop understanding what they're buying and approving what they're buying. And along every kind of step of this process for the foreseeable future. As we see it as part of.

A (16:55)

That, the buy side would need to have its own agent. In this case with rpa, we mentioned how nbcu, you had your seller agent. Free will had a seller agent, but RPA with Noon Research had a buyer agent in this case.

C (17:10)

Yeah. And so the buyer agent is essentially a chat bot or chat agent. So someone, a human at rpa, you know, instead of them writing an email for example, and saying, you know, what, what, you know, what Sunday Night Football or live sport thing do you have available or making a phone call or you know, all the older ways of doing it, they just go into or if they have a self serve dashboard with data interoperability within our system. Right. With via like traditional sort of API connections or things like that, they just go into this, this buyer agent, this chatbot which is in this UI that, that Newton has built and they just ask for the information and the information comes to them. They then decide I like that media plan or I don't or I want to see a different mix. All that stuff is still done by the human. Right. I ultimately want to go to order or I have another question. We don't foresee that going away anytime soon. But the speed at which they can get the information to do the planning process is highly accelerated. Which gives people on both sides, on our sides and on their side more time to talk strategically about what they want to do with the campaign. I've sort of said that it's like automating the operational logistics, right? So the human has more time to focus on the strategic importance and things like that.

A (18:51)

And it's funny because, like, I mean, so every time I think about like programmatic advertising, I think of the parallels to other commercial lines. And so in this case I think about like travel booking. You know, historically there were travel agents that you would call or like go to see in person to say, hey, I'm looking to go on vacation to Mexico. Can you help me find a flight and find a hotel? And that human travel agent would then call airlines, call hotels, figure out, okay, what's the best deal for me? And then they would get back to me and like, okay, I got your travel booked for you. And then things like kayak came along and automated a lot of that process where I could just, in a self serve way, go on, hey, I'm looking to go to Mexico. These dates. This is my budget. This is, you know, I want a hotel with these and amenities. This is kind of like my list of demands for what I want. Kayak would be like, cool. I'm gonna plug into a bunch of different APIs and I'll come back to you with a list of hotels that fit your parameters. And then I'm able to go through that list and click, okay, this one looks good. I'm gonna click buy, but I'm handling that. Now we're kind of going back to the travel agent. But now the travel agent is an AI agent.

C (20:18)

Well, you're talking as if you're talking to a travel agent, but yeah, yeah, but to the point that I think you're getting at, like would you ever. Now just because you're, you're chatting with an agent versus using maybe Kayak.com not to pick on them, would you allow that agent to just buy the ticket without you knowing the price of the ticket? Like, probably not, right? Just because you put this into an agentic workflow doesn't mean you don't want to be in the loop and doesn't mean you don't want to be checking those things. And so I think that's like a good kind of analogy for media planning. There is the ability to get answers more quickly using human language. I think your analogy is good with travel, but there are still things as a, as a buyer, whether you're buying an airline ticket to, you know, somewhere sunny and warm, which I'm thinking about these days, living in New York in the Snowstorm or a media plan that you still very much want humans processing those ultimate decisions.

B (21:26)

I do want to ask about. I always ask about success metrics. I feel like that's my sticking point for everything. But in the travel agent example, if I plan a vacation, I go through my travel agent to do so, and things fall apart. I go back to that travel agent agent and I say, hey, what happened here? Right? If a campaign underperforms, who is accountable in this situation? Is it the buyer? Is it nbcu? Is it the agent?

C (21:53)

I mean, I still think it's the. The human buyers and sellers, the same as it's always been. I mean, you can even just look at the programmatic, you know, so you're letting DSPs and SSPs do, like, bid by bid decisioning. But, like, if the campaign doesn't perform, they don't blame OpenRTV. It's like the buyers and sellers have a discussion. I think the theme that we're all hitting on here is that the technology is evolving. The way that we automate the operations and the speed at which we can operate, but it's not. It's not fundamentally changing a lot of things that have already been occurring with technology. In the evolution that you pointed out with the travel agent to the E commerce website, it's very much an evolution of that to make the process still more efficient.

A (22:53)

The thing that still kind of boggles my mind, sticking with the travel agent kayak analogy, I remember. And even still now, honestly, when I'm kind of booking bigger trips, I get a little wary of, like, trusting the kayak booking sometimes, you know, I want to then go to the airline website directly or the hotel directly. Sometimes I'm just like, God, I kind of want to just like, call the hotel to make sure that, like, this is actually real. Because this is a lot of money at stake. Mm. NBC Sunday football for a live NFL playoff game. There's a lot of money at stake there. And yet you all use that as a proof of concept. Can you run us through what you all did here? Because this has already happened, this proof of concept. Right? NBC, unless you're doing it for the super bowl, you all still have the Super Bowl.

C (23:50)

Yeah, I think that speaks to a couple of different things. One, the data foundations again, and the governance that needs. You need a lot of certainty that those things are accurate, especially when it comes to things like rates. And you do still need humans in the loop, I think, for, you know, as people get used to this technology, the other thing. And maybe we could, you know, we'll talk a little bit about like standards, like emerging standards, you know, I think we're probably going to need eventually and we're part of the IB Tech lab. We had some conversations with them about that. Things like trusted registries, like how do you know that your agent is talking to the real NBCU agent? Right. We had this version in digital with ads Txt, Right. To make sure that who you're buying from is the actual person you're buying from. So I think that there's also things that the industry needs to do to create assurances around, you know, that you're, that the agent you're talking to is the agent, the real, real agent that they're representing themselves to be. And I think all that stuff is going to, you know, is going to take time and work, you know, to create, to create trust. To your point, what did you all.

A (25:10)

Learn from this experience? Because again, this is run already. So I imagine you have some takeaways from it.

C (25:20)

I mean, God, we learned a lot. I mean, I think some big takeaways is that there is the ability for these TV platforms that are both digital and linear to offer an all encompassing form of automation that is next generation and really exciting, which is going to I think help our big agency and advertising partners interact with us in a way that's more turnkey and kind of easy button. But also I think bring in smaller companies that don't have the bandwidth to go through all of the more manual logistics of buying complex campaigns. And actually I didn't get this in earlier, so maybe I'll speak to it for a second here. AgentIC AI and the ability to orchestrate these AI agents offers another tagline which I don't know if it's a good one or not, but what I was kind of calling premium automation, right? So OpenRTB, you know, can streamline and automate the buying of certain things, but like understanding, you know, bespoke ad formats, different sponsorship integrations, like, you know, OpenRTV doesn't really do that, but you can ask those sorts of questions of an agent and get those sorts of answers and start to automate some of that higher end part of the market. And frankly, TV is sort of uniquely in that space unlike some other forms of digital advertising because it does offer a mix of media and sponsorship and integrations and creative formats and things like that that I think will be well served by agentic AI. And so I think the first big kind of learning is that that stuff's on the table. I think there's a way to really automate that and give companies time and productivity back.

A (27:50)

Got it. So this agentic ad buy for an NFL playoff game, it worked. It wasn't like, oh, we set it up to do it, and actually we're going to have to postpone a bit.

C (28:02)

No, I mean, yes, it works is a functioning technical proof of concept that accurately represents what the buyer wants to buy and what the seller has to sell.

A (28:15)

Was it the Bears Rams game? Because I know that went into overtime, and overtime is one of those things where all of a sudden new inventory opens up and so things like automation can be really su towards being able to fill that inventory. Was that the game in which you tested this?

C (28:31)

I actually don't. I don't know that. I don't know which game it was off the top of my head. But yeah, but yeah.

B (28:43)

For all of that we've talked about so far. Right. Do you think that we'll get to a point where buyers trust full automation, right? From top to bottom, agent to agent? And is that an ambition of you guys?

C (28:59)

I mean, I think that as the technology evolves over time, with the right governance and standards, I think people will grow to trust it more. But I'm not sure that that's a radical statement, because I feel like that's a general statement about technology. Right. I mean, I'm old enough, I don't know if you guys are to remember a time where it's just like, am I really going to buy something over the Internet and put my credit card information on the Internet?

A (29:33)

You know, again, I still feel that way sometimes.

C (29:36)

Yeah, yeah, yeah. And maybe, maybe you should. Right. But I think most people don't at this point. Right. Like, most people feel very comfortable putting a ton of information about themselves and their credit card information and doing all sorts of things online that they never imagined they would ever do or were even very highly skeptical of doing. And you have your early adopters and you have your laggards and you have that whole process that works out. But a lot of people want to paint the AI revolution as something that's never happened before. This time is different, you know, And I think it certainly is a huge opportunity, and it does feel like a watershed moment, like the invention of the Internet, like, where, you know, a lot of new things are going to come out of it. But at the end of the day, I do still believe that some of the, some of these things hold true, you know, that have held true for all these new technologies, technology revolutions. Like, I think there's a Lot of, like, commonality between past revolutions as well. I don't, I don't think that this is sort of like something never, ever been seen before. I think there's a lot of evidence that there's analogies to the past, and we've talked about some of them that will hold true for, you know, for this technology as well.

A (31:04)

And that's kind of leads me to something that I think we can end on as I'm talking with more people about agentic AI in advertising and kind of what it means. So much of the conversation has been around automating processes, freeing up human employees to be able to spend more time on strategy or other things that I've been wondering, okay, but what does any of this mean for the audience, for the regular people on the other end of these ads? Because when things transitioned from direct IO, the traditional ad buying and selling process, to programmatic, what that meant was more relevant ads was kind of one of the bigger things. Data was able to be applied to how ads were served so that Kimiko gets a certain ad because she's in Atlanta. I get a different kind of ad because I'm in San Diego, and you get a different ad because you're in New York with agentic AI applying to advertising. Ryan, what does any of this mean for the end user experience, the person on the other end of these ads?

C (32:17)

Well, I mean, it's a great question. And I think, to be clear, we've really kind of dug into one corner of the agentic AI possibilities right now, which is really kind of like buy, sell workflow. So there's whole other categories of AI that I guess we haven't fully explored in this interview. But I'll give you one example of where we're using large language models that does affect the user experience. And that was also kind of an announcement that we made at ces, which is that the contextual relevance of the ads. So for a long time you could do some contextual targeting, genre level targeting, show level targeting, but what you can do with AI now is far more, I think, compelling and advanced. So we have two products that are going to be live in 2026, one for our video on demand library and one for our live content. But just to go with the video on demand stuff, for example, if you're an advertiser and you want to be around something super specific, like I want to be around family scenes of people around campfires because I'm selling camping equipment or something like that, we can now load. And not just I want to be around campfires. I want to be around campfires that feature families in having positive interactions. Right. So maybe I don't want to be around like college dorm parties around campfires or something. Right. We can load all of that into a search function and it will scan 30,000 assets, television shows, movies that we have in our peacock library, and it will return in seconds the scenes that are the most relevant and also suggest the ad pods where you would place that ad to be closest to those moments within those assets. And that means that for that advertiser and for that user coming off of that scene, you're seeing ads that like, are hyper relevant contextually to the content that you're in, so much so that the ads can actually reference the content. And so that's sort of what we're starting to also do with live. So, you know, moments in sports where there is a fumble. This has been one example that we've shown. We can use the AI to scan the live feed in real time. And if there is a fumble, we can then communicate that to the ad server. And advertisers that have prepared for that moment might have custom creative that says, like, you know, don't, don't fumble this opportunity. You know, buy this product, you know, xyz and that is loaded and triggered in the ad server because of the ability to scan and see these things in real time and communicate to the ad server, then make the ad decision. So you're going to start to see more, more things like that.

A (35:28)

So basically, like, if we thought automation, programmatic advertising led to relevant ads, and it did, you're basically saying, well, with agentic AI, you ain't seen nothing yet.

C (35:43)

Yeah, I mean, I've been calling it the targeting trilogy. So the right ad to the right person at the right time. We've made a lot of advancement in the right ad format. Peacock's done a ton of ad innovation. The right person is all the work on the identity graph. And I talked a little bit about that and like television moving from like a retail model to a D2C model and having more information about who's watching and you'd be able to personalize the ad to the right person and now adding in the right, the exact right time. And I'm not saying like I'm generally in the right category, I'm in entertainment or I'm in sports, but like I'm at the exact right campfire moment that I want to be at. That's good for my brand. And we've done testing, initial testing, with this and we've compared it to baseline and the recall of the ad, the website visits, all of those metrics coming off of these hyper contextualized ads, they all go up. So we know that this sort of trilogy works really well the more precise that you can get it.

A (36:44)

And that's because the AgentIC AI workflow can handle things so much quicker than humans, but even than existing ad technologies.

C (36:56)

Yeah, yeah. I mean being able to comb through those assets prior and find all those moments and then know the exact ad pod to put them in and like that's. Yeah, there was no real simple way to do that prior. So that's a. It's a good example.

B (37:11)

All I can think about is how Oreos dunk in the dark moment years ago would have absolutely blown with this.

C (37:17)

Technology, you know, but it's actually a good point. But imagine that it's scale and automated. Right. You know, and as like the creative generation gets better and I'm not, this is not like a roadmap item, but this is just sort of an observation, you know, the ability to kind of do real time dynamic creative optimization and actually build the creatives kind of on the fly. I think you'll start to see that. I mean I think that might be a little bit further out, but if you squint, you can kind of see all that on the horizon. Right?

A (37:53)

Absolutely right. Well, this is fascinating conversation and really even fascinating more to think about post conversation in terms of where things are going to go from here, which I imagine Kimiko and I are going to be talking plenty about on the Digiday podcast throughout the rest of this year and in the years ahead. But thanks so much for coming on.

C (38:12)

Thanks for having me. Foreign.

A (38:18)

Thanks for listening to this episode of the Digit Day podcast. If you enjoyed it, please leave us a rating and a review on Apple Podcasts, Spotify or wherever you're listening. Get more from Digiday with our daily newsletter sent out each weekday morning. Visit digiday.comnewsletters to sign up.