How much control should AI have over media buying? - The Digiday Podcast

Loading summary

Transcript45 lines

[00:00]
A
Foreign.
[00:10]
B
Hello, hello, and welcome to another episode of the Digiday Podcast, a show about reducing the risk of AI agents blowing ad budgets. I'm Kamika McCoy, senior marketing reporter here at Digiday.
[00:22]
A
And I'm Tim Peterson, executive editor of Video and Audio Digital Media. Kamiko, I want to talk about the word governance, which I feel like is a word that left my vocabulary when I was no longer taking social studies classes, but this year has become a word that's like, firmly in my vocabulary to the point where I've had to, like, relearn what it even means or what it means outside of the context of, like, you know, actual governments.
[00:55]
B
I think we've been talking a lot about it when it comes to AI agents and how much they are allowed to do, what the reins look like. We talked a little bit about this on stage at DPMS earlier this year with Bayer, about how there's a trust gap and governance is kind of in the middle of that.
[01:14]
A
Yeah, yeah, it's. It feels like now this whole agentic advertising conversation has finally moved more so from the like, hey, check out what we can do. Look what we can do. Look what we can have AI do to. Okay, how do we actually keep these AI agents on the rails? And maybe part of this is also colored by. We've had examples now, Multiple examples of OpenAI's large language model breaking out of sandboxes and hacking companies anthropic. Then like a week later came out and was like, oh, crap. Actually, Cloud did that too, earlier in the year. We're just letting you all know now. And so maybe there's just something in the air. But that's all to say, felt like a really timely conversation we had for this week's episode with Scott Ensign, who is the chief Strategy Officer of Butler Till, which is, I believe, Butler Till claims to be the first agency to have done an agentic ad buy that. You know, they started testing AI agents for ad buying back in December of last year with PubMatic. They've done multiple tests since then, as they shared with us in this interview.
[02:31]
B
Yeah, I think it's absolutely coming along. And it's crazy how from when we had this conversation back in May, which was still kind of like not necessarily shiny object, but a lot of the railroad tracks were being built, built. Now it's about like, okay, how do we maintain these railroad tracks? And some of the best practices and things like that are starting to come to fruition, at least for Butler Till. I know it's fairly early days and when we talked to Scott. One of the interesting things that he mentioned is about, like, what the aspirational room for error is in these things and what the actual room for error is for these things. So to me, juicy conversation.
[03:07]
A
Yeah, because I mean, even like you mentioned. So I interviewed him on stage at DPMS in May. At that point, they had done, I think, two or three case studies to talk about. And then we just spoke to him. I think two weeks ago was when we recorded this conversation. And they had more than twice as many new case studies to talk about. New channels to talk about, too. Previously they were doing online display and ctv. Now they've expanded and we get even into talking about using AI agents for buying audio ads, which is interesting. So, yeah, it felt like really timely conversation. This is definitely not the last time I expect to say the word governance on this show.
[03:54]
B
Absolutely not. And if it's not about AI, what else is it about? So with no further ado,
[04:08]
A
Scott, welcome to the show. Thanks for joining us.
[04:11]
C
Good morning. Thanks so much for having me.
[04:14]
A
Happy to have you. So you and I spoke at the Digital Programmatic Marketing Summit earlier this year, in May of this year. Kimiko was in the room as well. And in that session we talked about a couple major case studies you all at Butler Till had done in using AI agents as part of ad buying workflows. There was one with PubMatic buying CTV inventory, and then there was another with Yahoo and Scope 3 that was more of a. Like using AI agents as part of more of a direct deal. So different types of campaigns or different types of buys that you all tested. That was as of May. Have there been any new major tests Butler Till has conducted involving AI agents since then?
[05:01]
C
There have, actually. There's a bunch. So as we talked about then, we've been pretty early and bullish on this technology for a variety of reasons. And so it's interesting, programmatic, as we all know, and we say here all the time, it's not a channel, it's a method of buying. But even that is those lines are being blurred in so many ways, as I'm sure we're all feeling and know. So what is and is not programmatic? I think kind of the Yahoo test that you mentioned was, and the scope 3 test was a good example of that. Since then, we've had half a dozen other tests, and that is that has expanded into different partners, different formats. So the initial one that you were referencing, which was with our friends at PubMatic and Jaloso, who owns Club Tales, that was A CTV specific test, we've expanded that into OLV online video. Then we are testing across pretty much every other format. So we have tests going with display, we have test going with streaming audio. And you mentioned direct. So we've been really expanding that into multiple formats and use cases and really kind of testing rails and approaches across the board.
[06:21]
A
Any of the tests felt like they've been kind of testing the boundaries or testing new things as opposed to strictly testing new partners.
[06:31]
C
Well, I think the next real phase of this is going to be using this as a way more to drive performance. And so that's where I think it goes beyond testing, like, does this work or not? Which Tim, I remember is what we talked a lot about in May, as this stuff is still very early, despite all of the things that we've talked about. And kind of everyone wants to know,
[06:55]
A
is this the thing I should just put all my money towards and then I can take a vacation?
[07:00]
C
And the answer to that is definitely not. Don't do that is my official advice. I think it's an important technology that we know is going to be at the center of a lot of buying in the same way that we knew that about those of us who remember knew that about digital many, many years ago. So I think some of the ways that what we're doing now is different is we're using Agentic to do different parts of the workflow. The initial test was just like do these Rails work? That was with ADCP at the very beginning of that protocol and technology. But now we're building things like Claude skills to monitor pacing and performance, to make updates to the media buys. That's becoming, and we were doing that to a certain extent early on, but that's becoming more and more sophisticated as this matures. There's different, there's different technologies with these different partners too, which makes it interesting, makes it both interesting and I think a little bit unclear where all of this stuff is going and where we're going to put all of our chips. So now I would say very much it's a very, very small, small percentage of our overall media efforts and media spend, but it's growing fast. Would you put a number on it percentage wise? I mean, low single digits in terms of the amount of spend that's going through this right now. And that's intentional. And we have the opportunity to do more and we see even bigger opportunities in the near future, but we're being very methodical about it as a media agency. Our stewardship is really at the heart of what we do and really, really important. So you know, we're bringing our clients along on this journey at a, at a pace that works for. Because again we do a lot, as I think I mentioned on stage, we do a lot in the pharma space and so tends to be, for very good reasons, a conservative category. And then the rest of this is across agriculture. You know, we mentioned the alcohol brand that we've done it with. So lots of things going in, lots of different areas.
[09:10]
A
Yeah, because if you're a fund manager and you tell your clients like, oh, hey, instead of buying the index funds and you know, things safe like that, we're going to throw everything on Kalshi right now.
[09:22]
C
Probably get a little unnerved, something like that. Yeah, yeah.
[09:27]
B
As you guys are, it sounds like more is being automated and handing over, handed over, excuse me, to machines. Is there anything that you guys want to keep strictly human?
[09:40]
C
Yeah, I mean I think the evaluation of insights is largely human. And I mean when we talk about the future, I always struggle a little bit because people will ask questions in the direction of like, is this something that will always be done by a human and always is a long time. But certainly I think the campaign insights in terms of what we're going to do next and where we're going to put the money. Well, those, those might be started ideas might be generated by the agents in these cases, but those are all being evaluated by people all the time. And so when I mentioned performance monitoring and insights, that's as you said, to automate some of those processes that historically have been a little bit manual. We can do things like, hey, this is underpacing here or our CPL is a little bit higher here than expected and that's arming our traders in the case of the programmatic campaigns with a list of things that are kind of helping them start on the third rung of the ladder. But the actual evaluation of those and execution of those, I don't see a near, even frankly medium term where there's not a human involved in, in those decisions for the most part.
[10:51]
B
What helped you guys determine that? Right. Because theoretically I can handle a lot of things, if not most things when it comes to like the media buying process. Was there something that happened during the test that made you determine this is the stopping point, we've got to have a human here?
[11:06]
C
Yeah, I mean, I think it's some of the outputs that we see. And so I'm not sure I completely agree with the statement that AI is capable of doing it. I guess Maybe in a technical sense AI is capable of doing it, but AI is not there yet in terms of making those human like decisions at scale that we are looking to make within media campaigns. I think I mentioned before being a media agency and stewardship being our prime directive if you will. We basically tolerate a zero or very very near zero. My boss always likes to remind me we're never going to have a goal that's less than zero or that's greater than zero percent error rate. Of course that's what we do and we see stuff in the things that we get back from the models. It's like eh, that's just a little bit off. And it's a little bit off to a degree, whether it's a budget recommendation, an audience recommendation. That's why I say I don't see a really near term or probably even medium term future where we can let that stuff go.
[12:15]
A
How do you determine what those thresholds are or should be? Because it sounds like they can vary depending on what specifically we're talking about. So I don't know if you can have hard and fast rules of if it's 0.3% there's the cutoff.
[12:33]
C
Yeah, I mean that's really hard I would say. And again that's why we're not ready to let machines let the agents make a lot of these decisions and why there's human oversight really into everything that we're doing today. But as I said before, I think the threshold for an actual error with this stuff is zero like our clients really expect. And we endeavor to deliver completely error free work. I mean we're spending money, we take that very, very seriously on behal of our clients. You know we do have thresholds in terms of like what kind of percentages of budgets we want to apply to different things and so we can set more I think hard and fast guardrails within the agentic frameworks that we're using, which is helpful too. But for me the really exciting thing has been what this has done is given that human trader that's working on these campaigns the ability to make much better decisions and to make them much faster. And so people talk a lot, which I totally understand about. Well, is this going to replace the job of that human trader? And you know I'm, I've been in the business long enough to remember when, you know, I need to stop telling these stories because it just reminds people how old I really really truly am. But I mean I remember recording, I worked in radio for a little while early in my Career recording things onto reel to reel and digital audio tape, you know, and then moving into a, into a world where we can create radio spots and email them around. And I think that's the kind of revolution that we're seeing is that the, the work that that trader is going to be able to do is just going to be amplified so much. They can do things faster, we can have planning cycles happen faster, we can get to insights faster. It's not all going to be gated by what, you know, a single human being can do in a day. That human trader is going to have at their disposal an army of agents who can help them do more and do all of it much faster. And so expectations on the client side are going to change as a result,
[14:39]
A
of course, going back to. So now at this point, across all of the different sites, you started in ctv, then you did, I believe, Yahoo's online display, you've now done online video streaming audio. Are there any major important differences when it comes to the different channels when you're using agents? Or is the idea like the bones are the same, the infrastructure is largely the same and so from a purely. When it comes to the agentic considerations, it becomes channel is less of an important distinction.
[15:17]
C
Yeah, that's a great question because we think about, and I just had this conversation with some folks internally here yesterday, we think about like what is our AI vision and strategy? It's like AI is too big of a technology to have one vision associated with it and agentic in terms of its impact on media is too big to have one thing that it does or even one set of things that it does. So when you talk about those different tests that we did, which, you know, I'm glad you mentioned the Yahoo one again because yes, it was displayed, but it was also more an endemic single partner display campaign versus a programmatic campaign which is very different. And we actually today have different teams that are working on those across the organization. When you think about our more direct and endemic team sponsorship team versus the programmatic traders, I think those use cases are extremely different with those I mentioned before, optimization and performance insights that we're pulling out of some of the more programmatic things that we're doing. When I think about the direct and endemic work, we're talking also to our friends over at Google Ad Manager about buying directly into the publisher side ad server using agentic technology, which is a very different use case from the ones that we just mentioned. And so in that case we have media buyers who are RFPing a number of publishers who are getting those RFPs back, who are evaluating them, who are waiting for a rep to get back from vacation, all of those things that have gone into that process now we can have. And this is all very nascent and we're not up and running with it yet. But the idea is that on that ad server side, each publisher represented in Google Ad Manager on the ad server would have their own seller agent and our buyer agent would be negotiating. And what used to take a few days, maybe a week again, depending on how much back and forth or the size of the campaign, in theory, could happen nearly instantaneously. That's an agentic use case. That's more like a negotiation, like a real supercharged, super fast negotiation process. And one of the reasons that planning cycles are what they are today in the media world are because these things just take time. You can't do one of those plans every day the way that we've always done it in theory. And again, I don't want to get too ahead of where we are with the technology, but in theory that should shorten planning cycles from annual to quarterly to monthly, potentially to weekly. And if we can get to those insights faster, I think you can pretty quickly see a future where as soon as we have an idea on performance, even on the direct side, we're able to set our negotiating agents back out. Buyer agents talking to seller agents to refine plan. And then it's kind of where we start is what is and is not programmatic. I think those lines are blurred. Is it using rtb? Is it real time? Technically speaking, no. But it is leveraging technology to make this stuff happen much faster and to get to performance and insights much faster.
[18:41]
B
I want to ask the same question that Tim asked, but about the guardrails. Do they change when it comes to is it client by client, is it channel by channel? What determines the guardrails that you guys have around the agents?
[18:57]
C
Really, really important question too because we have always had and Butler till specifically, I mean, we deal almost exclusively with enterprise level clients. So each one of those we get to know very well. We have people who work on them and each one of them has their own unique set of security concerns, privacy concerns, regulatory environments. And as I mentioned before, we work in pharma, we work in med device, we work in health insurance, we work in alcohol, we work in financial services. So the governance on each of those is different and it's really, really important. And each one of those clients has a different risk tolerance and profile for very good reasons. So we Were just having a conversation yesterday. I think the future is going to be an enterprise layer of AI in general, but specifically with Agentic and those environments are going to have to be controlled on a client by client basis. Today we're doing that essentially with like there's a team working on it, that team understands that specific client and can set up the Agentic framework to align with that client's risk tolerance and specific concerns. But ultimately as we scale that and it gets bigger and that, you know, low single digit number I believe is going to become double digits pretty quickly. We're going to have to have infrastructure that allows us to manage that at an enterprise level and set up very, very client specific guardrails and instances inside of our data environments, inside of our AI environments that are going to give clients the confidence that we are in the same way that we talk a lot about human reasoning at scale in the same way a human would look at that and say, is this on compliance that we will have the right frameworks in place to do that in this environment as well?
[20:57]
B
Are there any clients that are, I guess hesitant? Right, because when it comes to AI and you talk about like a pharma client or something like that, there's concerns over data and whatnot. So what's the client appetite here?
[21:12]
C
I mean the short answer is yes, there are clients who are much more and I obviously won't mention any by name but like we've got some categories that just seem to be a lot more conservative and risk averse. And it's not even necessarily, I think people think of pharma as a category that is a little bit conservative, but it's really, we've seen more and I don't want to say hesitant, I would say maybe more cautious in the financial services area just because, you know, these are companies that do risk management as a core part of their offering. And I think a lot of them in my experience and conversation are taking a lot more of a cautious approach with AI overall. But certainly then that applies to the way that we're doing media buying on their behalf. And they want to have a lot more reviews along the way of like what's going to happen here? What's going to happen here? What data are we giving to this model? How's that data going to be used? Is it going to be shared? And if so, how, how is it anonymized? There are very different approaches depending on the client and the category for sure.
[22:24]
A
You mentioned earlier using Claude skills. Have you all consolidated your agentic work around The Claude model or are you using different AI models?
[22:35]
C
Different AI models. I would say Claude, I mean, CLAUDE was the first one that we used. I think in the agentic world right now in the. The use cases that we're exploring for us and I'm sure others are having a different experience, but for us, Claude has been the most useful. But no, we're also experimenting with ChatGPT, we're experimenting with Gemini, we're experimenting and we use a lot of Copilot here too because we're a Microsoft shop as an agency. So we're exploring a bunch of frontier models. But definitely Claude has been the one that's really at the forefront for the agentic work that we're doing.
[23:19]
A
How does that affect governance? Not even just managing different models, whether it's Claude, GPT, Gemini, et cetera, because those can be. You have to work with those models differently. I hate Gemini just because it doesn't care about the instructions as much as some of the other. It's not as obedient. But even with Claude, now there's Fable 5 and there's Opus 5. And one thing that changed with, I think it was Fable 5, maybe it was Opus 4.8, but you kind of need to be less prescriptive with the model for the model to be as effective. It's more about telling the model, here's my goal, kind of figure it out. Because it can do longer term reasoning and longer term thinking. And so whereas a year ago, even maybe six months ago, it was more. You had to write a 500,000 word prompt and tell it exactly what you wanted it to do. So how does governance change when you're using different models? Or is it the kind of thing where you start with, I think Opus 4.6 was, or maybe I think it was 4.6 was state of the art when you all were doing the first CTV test in December of last year. And so is it like a Microsoft Windows thing where if you're starting out on Windows 95, you're not trying to upgrade to XP because that's going to break every part of your workflow?
[24:50]
C
Yeah, I would say based on the work that we've done so far, we haven't found that kind of. The incremental model upgrades have made a huge difference in terms of what we've already put in place. It hasn't broken. Hasn't broken that stuff. I feel like I should be knocking on wood when I say that the next model is sure to break everything that I've said that out loud. And on tape here. But I think that goes back to what we were talking about before where I think having the human oversight is going to continue to be really important because as these models become more sophisticated, there is longer term implication for what we're doing. And to your point, just like with a new employee, you might need to be very, very prescriptive with like, here's how you do this, here's the next step, here's blah blah, blah, blah, blah. But then when you have your next meeting with someone on your team and it's someone who's been doing programmatic trading for five years, it's a very different conversation. And I think that that's kind of been our experience with the models as they've continued to be more sophisticated. Like the, the context is much more sophisticated and built in. But to me that's even more reason why it requires human oversight is because those inferences can still be as great as these things are, as amazing as they are, as much as when I use them, it still blows my mind every time. They are still unpredictable. And there are still things that happen within the inferences that are making that you're like, that is not right, not something that we would want. And it's really interesting things too. Stuff that you wouldn't expect. Like you have it create a PowerPoint for you which not related to agentic media buying obviously. And you're like, why did you make that logo look different? There's just stuff that happens that sometimes doesn't make a lot of sense. Which is really interesting and fun for us frankly because our team gets to tinker and experiment and see here's what it's coming back with and figure out why did it think this or why did it do this? But again, just more argument that I don't think anyone is anywhere near ready to set up an end to end agentic only media agency or media buying shop. It's a lot that's going to have to go into that.
[27:16]
A
And what about the seller agents? Because obviously you all can do a lot of the governance when it comes to your buyer agents. That's where you have the most control, that's where you have the most visibility. The seller agents though, there's a lot of trust that goes into those and obviously those are the responsibility of the sellers, of the publishers or the platforms. But how they interpret things that are said to them by your buyer agent can meaningfully change how a buy gets executed. Are you able to vet the seller agents? Are you able to see the system prompts that sellers have given as instructions for their seller agents.
[27:54]
C
Yeah, I would say overall, the work that we've been doing right now is so kind of early and collaborative that the answer to that is yes, they're very open and willing to share. Like this is how we're prompting our agent. This is what the agent is doing. I do see a future where that's going to change because at the end of the day it's a negotiation and they have their interests and we have our interests. And we've seen what kind of the proliferation of technology to wield greater negotiation power has, has warped and done weird things in our space over, over the years. Programmatic being, I think probably the most prominent example of that. So I think there's so much left to be written about the norms and how that's going to work. My fear is, and I think this is one of the main things that has caused issues in the programmatic space is what I just described, which is like the buyer has a certain amount of information, the seller has a certain amount of information. And there's a lot of trepidation about information sharing because you fear that it's going to be used against you in a negotiation. And just in a very. I'm not talking about people having ill intent or anything like that, just a very basic sense of like, I'm selling something, I want to get as much money as I can for it, I'm buying something, I want to get it for as little money as I can. And so it's not an environment that naturally lends itself to open conduit of vulnerable information. I think that is a really interesting construct as we think about agentic because the seller agent really needs good things like performance information. Obviously we're never going to share sensitive client information or things like that with, with a seller agent. But to me, I think we need to move towards standards where we can share as much information as possible to improve not just the buy, but the entire ecosystem which again, not to get too much on a soapbox, but part of the reason that we've seen such a proliferation of low quality content is the only signal that the buyer sends to the seller through dsp, SSP and all of those protocols is I want this for less money. That is not going to result in high quality content. It's not going to result in the monetization of things that I think are good for the ecosystem, frankly, not to get too heady about it, but good for society as we've seen. I think if we can come up with standards and ways of working that are going to allow us to. I would love to a future where through that kind of sophisticated agentic media buying model we can communicate to the sell side of the market. Like this is what our clients care about. So this is a financial services client. We don't want to be on some like low quality site with you know, eight ads loading and playing. We want to, we want to make a connection with a real person over here by by creating a value exchange that's really, really meaningful. And I want you seller and I'm obviously anthropomorphizing this to the extreme but like to make more of that stuff, I want the signal they get to be leading them in a direction that creates a better open Internet. And I do think agentic, not to be too maybe overly optimistic about it, but I do think agentic has the capability to do that because we can move from a hyper, hyper limited sophistication of more stuff for less money to here's what we're really looking for and here's the value that we're looking to create.
[31:37]
B
I think it's fair to say based on what you just said, the governing bodies of this industry that be have their work now cut out for them. They're listening to this episode as to setting some industry standards for this because
[31:51]
A
you're right and agreeing on those standards
[31:53]
B
and for them to be adopted.
[31:54]
A
Multiple governing bodies when it comes to hsd.
[31:58]
B
But nonetheless they've been taken to task based on what you said. And I guess we should be checking in with you at our next programmatic marketing summit and our next conversation given you guys are knee deep in the work.
[32:13]
C
Yeah, we've got, as I mentioned, more than half a dozen tests across formats, across categories, across platforms running right now. So we're excited, we're excited to see what this is going to, where this is going to point us next and what we can share with the industry because we think it's really important. And as I said, I tend to be an optimist. So I think that if we all kind of work together and share this stuff about what's working and what's making things better in our space, we could perhaps create a little wave of improvement instead of just racing to the bottom. So,
[32:53]
A
Scott, thanks so much for coming on the show. Really enjoyed the conversation.
[32:56]
C
Of course, it's a pleasure. Thanks so much for having me.
[33:06]
A
Thanks for listening to this episode of the Digiday 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 week day morning. Visit digidaire.comnewsletters to sign up.

Summary6 min read

Episode Overview

Title: How much control should AI have over media buying?
Podcast: The Digiday Podcast
Date: August 11, 2026
Host(s): Kamiko McCoy (Senior Marketing Reporter, Digiday) & Tim Peterson (Executive Editor, Video and Audio Digital Media)
Guest: Scott Ensign (Chief Strategy Officer, Butler/Till)

Theme:
The episode explores the rapidly advancing role of agentic AI in media buying, focusing on governance, trust, and the balance between automation and human oversight. Insights are drawn from Butler/Till’s real-world pilots with AI-driven ad buying.

Key Discussion Points & Insights

1. Governance and Trust in AI-driven Advertising

Tim opens by highlighting the resurgence and importance of the term "governance," especially regarding the risk and trust gap with AI agents in ad buying (00:21).
Kamiko notes the industry’s shift from initial amazement at AI capabilities to serious consideration of effective guardrails and risk management (00:54).

"It feels like now this whole agentic advertising conversation has finally moved... to 'Okay, how do we actually keep these AI agents on the rails?'" – Tim Peterson (01:14)

Recent Incidents: Mention of LLMs (e.g., OpenAI, Anthropic's Claude) breaking out of sandboxes, emphasizing the need for strict oversight.

2. Butler/Till's Early Experiments and Expansion with AI Agents

Scott details Butler/Till’s position as an early adopter of agentic ad buying, having started with PubMatic on CTV inventory and then expanded into online video (OLV), display, and now streaming audio (05:01).
The agency has conducted over half a dozen tests, each pushing into different formats and partners, evolving from CTV-specific to include more direct and endemic deals (06:20).

"Since then, we've had half a dozen other tests, and that has expanded into different partners, different formats... we're testing across pretty much every other format." – Scott Ensign (05:01)

3. The Role and Limits of Automation in Media Buying

Current Use: AI agents are now automating more parts of the workflow (e.g., pacing, performance monitoring), yet the share of media spend handled by AI remains in the low single digits (06:59).
Room for Error: Butler/Till’s approach is highly cautious—stewardship and client trust are prioritized, especially in conservative sectors like pharma and finance.
Critical Insight: Automation boosts efficiency but true human oversight remains non-negotiable, especially for error-checking and decision-making (09:26, 10:50).

"[AI] is not there yet in terms of making those human-like decisions at scale that we are looking to make... our clients really expect and we endeavor to deliver completely error-free work." – Scott Ensign (11:06, 12:32)

Aspirational vs. Actual Error Rates: Zero error tolerance is the official policy due to the direct management of client funds (12:32).

4. Channel-Specific Considerations for Agentic Media Buying

AI’s impact and required controls vary depending on the type of buy (programmatic vs. direct/endemic) and the specific media channel (CTV, display, audio, etc.) (15:17).
End-to-end agentic workflows involve both buyer and seller agents exchanging requests and bids almost instantly—potentially slashing planning cycles from weeks to days or hours (17:00).

"The idea is that on that ad server side, each publisher... would have their own seller agent and our buyer agent would be negotiating... in theory, could happen nearly instantaneously." – Scott Ensign (16:36)

These differences necessitate customized guardrails and governance structures for every client, campaign, and media format (18:56).

5. Governance Frameworks & Industry Risk Tolerance

Enterprises working with sensitive categories (e.g., pharma, financial services) require unique, client-specific AI governance environments, aligning with regulatory, security, and privacy needs (18:56, 21:11).
Financial services are especially conservative and risk-averse with AI, often demanding granular reviews, strict data controls, and full transparency at every step (21:11).

"Each [client] has their own unique set of security concerns, privacy concerns, regulatory environments... each one of those clients has a different risk tolerance and profile." – Scott Ensign (18:56)

6. Managing Multiple AI Models & Evolution of Prompting

Butler/Till is experimenting with a range of AI models: Claude, ChatGPT, Gemini, Copilot—each with different capabilities, controls, and quirks (22:34).
As models become more sophisticated (e.g., improved reasoning in Claude’s newer releases), governance often becomes trickier, requiring less prescriptive but more goal-oriented prompts (23:19).
Model upgrades haven’t broken workflows yet, but Scott cautions the need for ongoing human oversight due to unpredictability and emergent model behaviors (24:50).

"As these models become more sophisticated, there is longer term implication for what we're doing... that's even more reason why it requires human oversight." – Scott Ensign (24:50)

7. Buyer Agents, Seller Agents & Market Transparency

Today, close collaboration with partners allows for open sharing of system prompts and agent instructions, but competitive tensions and negotiation dynamics may limit this transparency as the system matures (27:15).
There’s a risk that buyers and sellers might withhold information—leading to market inefficiencies as previously seen in programmatic (29:00).
Scott advocates for developing new standards that allow safe, value-driven information-sharing to improve the overall ecosystem and not just chase lowest CPMs (30:00).

"We need to move towards standards... to improve not just the buy, but the entire ecosystem... [otherwise] the only signal the buyer sends... is I want this for less money." – Scott Ensign (30:59)

8. Call for Industry Standards and Shared Learning

The growing complexity and stakes necessitate the development of robust industry-wide standards for agentic media buying. Multiple governing bodies will need to collaborate for success (31:37).
Butler/Till continues to share learnings from its many ongoing cross-channel tests, striving for greater value and transparency industry-wide (32:12).

"We think it's really important... if we all work together and share this stuff about what's working and what's making things better... we could perhaps create a little wave of improvement instead of just racing to the bottom." – Scott Ensign (32:12)

Memorable Quotes (with Timestamps)

On trust and governance:
"There's a trust gap and governance is kind of in the middle of that." – Kamiko McCoy (00:54)
On the role of human oversight:
"Zero or very, very near zero... we see stuff in the things that we get back from the models... that's just a little bit off." – Scott Ensign (11:06)
On the promise and limits of automation:
"The work that trader is going to be able to do is just going to be amplified so much. They can do things faster... it's not all going to be gated by what a single human being can do in a day." – Scott Ensign (13:30)
On channel specificity:
"AI is too big of a technology to have one vision associated with it and agentic in terms of its impact on media is too big to have one thing that it does." – Scott Ensign (15:17)
On industry standards:
"Multiple governing bodies when it comes to HSD... they've been taken to task based on what you said." – Kamiko McCoy (31:54)

Timestamps for Important Segments

| Segment Topic | Speaker | Timestamp | |--------------------------------------------------|----------------|------------| | Introduction & Governance in AI | Tim | 00:21 | | History of Butler/Till's AI Ad Buying | Scott | 05:01 | | Automation vs. Human Oversight | Scott | 09:26–12:32| | Channel-Specific Agentic Approaches | Scott | 15:17 | | Governance & Guardrails Customization | Scott | 18:56 | | Client Appetites and Caution (Finance vs Pharma) | Scott | 21:11 | | Managing Multiple Models & Prompting Challenges | Tim & Scott | 22:34–24:50| | Transparency with Seller Agents | Scott | 27:15 | | Need for Industry Standards | Panel | 31:37 | | Looking Forward: Ongoing Tests & Optimism | Scott | 32:12 |

Conclusion

This episode provides an in-depth, real-world perspective on how far AI-driven, agentic ad buying has come—and how far it has to go. It stresses the indispensable role of human oversight, the importance of context- and client-specific governance, and the urgent need for industry-wide cooperation on standards. Above all, it presents a hopeful, pragmatic view that, with careful stewardship and collaboration, AI can elevate rather than erode trust and value in digital advertising.

Loading summary

Transcript45 lines

[00:00]
A
Foreign.
[00:10]
B
Hello, hello, and welcome to another episode of the Digiday Podcast, a show about reducing the risk of AI agents blowing ad budgets. I'm Kamika McCoy, senior marketing reporter here at Digiday.
[00:22]
A
And I'm Tim Peterson, executive editor of Video and Audio Digital Media. Kamiko, I want to talk about the word governance, which I feel like is a word that left my vocabulary when I was no longer taking social studies classes, but this year has become a word that's like, firmly in my vocabulary to the point where I've had to, like, relearn what it even means or what it means outside of the context of, like, you know, actual governments.
[00:55]
B
I think we've been talking a lot about it when it comes to AI agents and how much they are allowed to do, what the reins look like. We talked a little bit about this on stage at DPMS earlier this year with Bayer, about how there's a trust gap and governance is kind of in the middle of that.
[01:14]
A
Yeah, yeah, it's. It feels like now this whole agentic advertising conversation has finally moved more so from the like, hey, check out what we can do. Look what we can do. Look what we can have AI do to. Okay, how do we actually keep these AI agents on the rails? And maybe part of this is also colored by. We've had examples now, Multiple examples of OpenAI's large language model breaking out of sandboxes and hacking companies anthropic. Then like a week later came out and was like, oh, crap. Actually, Cloud did that too, earlier in the year. We're just letting you all know now. And so maybe there's just something in the air. But that's all to say, felt like a really timely conversation we had for this week's episode with Scott Ensign, who is the chief Strategy Officer of Butler Till, which is, I believe, Butler Till claims to be the first agency to have done an agentic ad buy that. You know, they started testing AI agents for ad buying back in December of last year with PubMatic. They've done multiple tests since then, as they shared with us in this interview.
[02:31]
B
Yeah, I think it's absolutely coming along. And it's crazy how from when we had this conversation back in May, which was still kind of like not necessarily shiny object, but a lot of the railroad tracks were being built, built. Now it's about like, okay, how do we maintain these railroad tracks? And some of the best practices and things like that are starting to come to fruition, at least for Butler Till. I know it's fairly early days and when we talked to Scott. One of the interesting things that he mentioned is about, like, what the aspirational room for error is in these things and what the actual room for error is for these things. So to me, juicy conversation.
[03:07]
A
Yeah, because I mean, even like you mentioned. So I interviewed him on stage at DPMS in May. At that point, they had done, I think, two or three case studies to talk about. And then we just spoke to him. I think two weeks ago was when we recorded this conversation. And they had more than twice as many new case studies to talk about. New channels to talk about, too. Previously they were doing online display and ctv. Now they've expanded and we get even into talking about using AI agents for buying audio ads, which is interesting. So, yeah, it felt like really timely conversation. This is definitely not the last time I expect to say the word governance on this show.
[03:54]
B
Absolutely not. And if it's not about AI, what else is it about? So with no further ado,
[04:08]
A
Scott, welcome to the show. Thanks for joining us.
[04:11]
C
Good morning. Thanks so much for having me.
[04:14]
A
Happy to have you. So you and I spoke at the Digital Programmatic Marketing Summit earlier this year, in May of this year. Kimiko was in the room as well. And in that session we talked about a couple major case studies you all at Butler Till had done in using AI agents as part of ad buying workflows. There was one with PubMatic buying CTV inventory, and then there was another with Yahoo and Scope 3 that was more of a. Like using AI agents as part of more of a direct deal. So different types of campaigns or different types of buys that you all tested. That was as of May. Have there been any new major tests Butler Till has conducted involving AI agents since then?
[05:01]
C
There have, actually. There's a bunch. So as we talked about then, we've been pretty early and bullish on this technology for a variety of reasons. And so it's interesting, programmatic, as we all know, and we say here all the time, it's not a channel, it's a method of buying. But even that is those lines are being blurred in so many ways, as I'm sure we're all feeling and know. So what is and is not programmatic? I think kind of the Yahoo test that you mentioned was, and the scope 3 test was a good example of that. Since then, we've had half a dozen other tests, and that is that has expanded into different partners, different formats. So the initial one that you were referencing, which was with our friends at PubMatic and Jaloso, who owns Club Tales, that was A CTV specific test, we've expanded that into OLV online video. Then we are testing across pretty much every other format. So we have tests going with display, we have test going with streaming audio. And you mentioned direct. So we've been really expanding that into multiple formats and use cases and really kind of testing rails and approaches across the board.
[06:21]
A
Any of the tests felt like they've been kind of testing the boundaries or testing new things as opposed to strictly testing new partners.
[06:31]
C
Well, I think the next real phase of this is going to be using this as a way more to drive performance. And so that's where I think it goes beyond testing, like, does this work or not? Which Tim, I remember is what we talked a lot about in May, as this stuff is still very early, despite all of the things that we've talked about. And kind of everyone wants to know,
[06:55]
A
is this the thing I should just put all my money towards and then I can take a vacation?
[07:00]
C
And the answer to that is definitely not. Don't do that is my official advice. I think it's an important technology that we know is going to be at the center of a lot of buying in the same way that we knew that about those of us who remember knew that about digital many, many years ago. So I think some of the ways that what we're doing now is different is we're using Agentic to do different parts of the workflow. The initial test was just like do these Rails work? That was with ADCP at the very beginning of that protocol and technology. But now we're building things like Claude skills to monitor pacing and performance, to make updates to the media buys. That's becoming, and we were doing that to a certain extent early on, but that's becoming more and more sophisticated as this matures. There's different, there's different technologies with these different partners too, which makes it interesting, makes it both interesting and I think a little bit unclear where all of this stuff is going and where we're going to put all of our chips. So now I would say very much it's a very, very small, small percentage of our overall media efforts and media spend, but it's growing fast. Would you put a number on it percentage wise? I mean, low single digits in terms of the amount of spend that's going through this right now. And that's intentional. And we have the opportunity to do more and we see even bigger opportunities in the near future, but we're being very methodical about it as a media agency. Our stewardship is really at the heart of what we do and really, really important. So you know, we're bringing our clients along on this journey at a, at a pace that works for. Because again we do a lot, as I think I mentioned on stage, we do a lot in the pharma space and so tends to be, for very good reasons, a conservative category. And then the rest of this is across agriculture. You know, we mentioned the alcohol brand that we've done it with. So lots of things going in, lots of different areas.
[09:10]
A
Yeah, because if you're a fund manager and you tell your clients like, oh, hey, instead of buying the index funds and you know, things safe like that, we're going to throw everything on Kalshi right now.
[09:22]
C
Probably get a little unnerved, something like that. Yeah, yeah.
[09:27]
B
As you guys are, it sounds like more is being automated and handing over, handed over, excuse me, to machines. Is there anything that you guys want to keep strictly human?
[09:40]
C
Yeah, I mean I think the evaluation of insights is largely human. And I mean when we talk about the future, I always struggle a little bit because people will ask questions in the direction of like, is this something that will always be done by a human and always is a long time. But certainly I think the campaign insights in terms of what we're going to do next and where we're going to put the money. Well, those, those might be started ideas might be generated by the agents in these cases, but those are all being evaluated by people all the time. And so when I mentioned performance monitoring and insights, that's as you said, to automate some of those processes that historically have been a little bit manual. We can do things like, hey, this is underpacing here or our CPL is a little bit higher here than expected and that's arming our traders in the case of the programmatic campaigns with a list of things that are kind of helping them start on the third rung of the ladder. But the actual evaluation of those and execution of those, I don't see a near, even frankly medium term where there's not a human involved in, in those decisions for the most part.
[10:51]
B
What helped you guys determine that? Right. Because theoretically I can handle a lot of things, if not most things when it comes to like the media buying process. Was there something that happened during the test that made you determine this is the stopping point, we've got to have a human here?
[11:06]
C
Yeah, I mean, I think it's some of the outputs that we see. And so I'm not sure I completely agree with the statement that AI is capable of doing it. I guess Maybe in a technical sense AI is capable of doing it, but AI is not there yet in terms of making those human like decisions at scale that we are looking to make within media campaigns. I think I mentioned before being a media agency and stewardship being our prime directive if you will. We basically tolerate a zero or very very near zero. My boss always likes to remind me we're never going to have a goal that's less than zero or that's greater than zero percent error rate. Of course that's what we do and we see stuff in the things that we get back from the models. It's like eh, that's just a little bit off. And it's a little bit off to a degree, whether it's a budget recommendation, an audience recommendation. That's why I say I don't see a really near term or probably even medium term future where we can let that stuff go.
[12:15]
A
How do you determine what those thresholds are or should be? Because it sounds like they can vary depending on what specifically we're talking about. So I don't know if you can have hard and fast rules of if it's 0.3% there's the cutoff.
[12:33]
C
Yeah, I mean that's really hard I would say. And again that's why we're not ready to let machines let the agents make a lot of these decisions and why there's human oversight really into everything that we're doing today. But as I said before, I think the threshold for an actual error with this stuff is zero like our clients really expect. And we endeavor to deliver completely error free work. I mean we're spending money, we take that very, very seriously on behal of our clients. You know we do have thresholds in terms of like what kind of percentages of budgets we want to apply to different things and so we can set more I think hard and fast guardrails within the agentic frameworks that we're using, which is helpful too. But for me the really exciting thing has been what this has done is given that human trader that's working on these campaigns the ability to make much better decisions and to make them much faster. And so people talk a lot, which I totally understand about. Well, is this going to replace the job of that human trader? And you know I'm, I've been in the business long enough to remember when, you know, I need to stop telling these stories because it just reminds people how old I really really truly am. But I mean I remember recording, I worked in radio for a little while early in my Career recording things onto reel to reel and digital audio tape, you know, and then moving into a, into a world where we can create radio spots and email them around. And I think that's the kind of revolution that we're seeing is that the, the work that that trader is going to be able to do is just going to be amplified so much. They can do things faster, we can have planning cycles happen faster, we can get to insights faster. It's not all going to be gated by what, you know, a single human being can do in a day. That human trader is going to have at their disposal an army of agents who can help them do more and do all of it much faster. And so expectations on the client side are going to change as a result,
[14:39]
A
of course, going back to. So now at this point, across all of the different sites, you started in ctv, then you did, I believe, Yahoo's online display, you've now done online video streaming audio. Are there any major important differences when it comes to the different channels when you're using agents? Or is the idea like the bones are the same, the infrastructure is largely the same and so from a purely. When it comes to the agentic considerations, it becomes channel is less of an important distinction.
[15:17]
C
Yeah, that's a great question because we think about, and I just had this conversation with some folks internally here yesterday, we think about like what is our AI vision and strategy? It's like AI is too big of a technology to have one vision associated with it and agentic in terms of its impact on media is too big to have one thing that it does or even one set of things that it does. So when you talk about those different tests that we did, which, you know, I'm glad you mentioned the Yahoo one again because yes, it was displayed, but it was also more an endemic single partner display campaign versus a programmatic campaign which is very different. And we actually today have different teams that are working on those across the organization. When you think about our more direct and endemic team sponsorship team versus the programmatic traders, I think those use cases are extremely different with those I mentioned before, optimization and performance insights that we're pulling out of some of the more programmatic things that we're doing. When I think about the direct and endemic work, we're talking also to our friends over at Google Ad Manager about buying directly into the publisher side ad server using agentic technology, which is a very different use case from the ones that we just mentioned. And so in that case we have media buyers who are RFPing a number of publishers who are getting those RFPs back, who are evaluating them, who are waiting for a rep to get back from vacation, all of those things that have gone into that process now we can have. And this is all very nascent and we're not up and running with it yet. But the idea is that on that ad server side, each publisher represented in Google Ad Manager on the ad server would have their own seller agent and our buyer agent would be negotiating. And what used to take a few days, maybe a week again, depending on how much back and forth or the size of the campaign, in theory, could happen nearly instantaneously. That's an agentic use case. That's more like a negotiation, like a real supercharged, super fast negotiation process. And one of the reasons that planning cycles are what they are today in the media world are because these things just take time. You can't do one of those plans every day the way that we've always done it in theory. And again, I don't want to get too ahead of where we are with the technology, but in theory that should shorten planning cycles from annual to quarterly to monthly, potentially to weekly. And if we can get to those insights faster, I think you can pretty quickly see a future where as soon as we have an idea on performance, even on the direct side, we're able to set our negotiating agents back out. Buyer agents talking to seller agents to refine plan. And then it's kind of where we start is what is and is not programmatic. I think those lines are blurred. Is it using rtb? Is it real time? Technically speaking, no. But it is leveraging technology to make this stuff happen much faster and to get to performance and insights much faster.
[18:41]
B
I want to ask the same question that Tim asked, but about the guardrails. Do they change when it comes to is it client by client, is it channel by channel? What determines the guardrails that you guys have around the agents?
[18:57]
C
Really, really important question too because we have always had and Butler till specifically, I mean, we deal almost exclusively with enterprise level clients. So each one of those we get to know very well. We have people who work on them and each one of them has their own unique set of security concerns, privacy concerns, regulatory environments. And as I mentioned before, we work in pharma, we work in med device, we work in health insurance, we work in alcohol, we work in financial services. So the governance on each of those is different and it's really, really important. And each one of those clients has a different risk tolerance and profile for very good reasons. So we Were just having a conversation yesterday. I think the future is going to be an enterprise layer of AI in general, but specifically with Agentic and those environments are going to have to be controlled on a client by client basis. Today we're doing that essentially with like there's a team working on it, that team understands that specific client and can set up the Agentic framework to align with that client's risk tolerance and specific concerns. But ultimately as we scale that and it gets bigger and that, you know, low single digit number I believe is going to become double digits pretty quickly. We're going to have to have infrastructure that allows us to manage that at an enterprise level and set up very, very client specific guardrails and instances inside of our data environments, inside of our AI environments that are going to give clients the confidence that we are in the same way that we talk a lot about human reasoning at scale in the same way a human would look at that and say, is this on compliance that we will have the right frameworks in place to do that in this environment as well?
[20:57]
B
Are there any clients that are, I guess hesitant? Right, because when it comes to AI and you talk about like a pharma client or something like that, there's concerns over data and whatnot. So what's the client appetite here?
[21:12]
C
I mean the short answer is yes, there are clients who are much more and I obviously won't mention any by name but like we've got some categories that just seem to be a lot more conservative and risk averse. And it's not even necessarily, I think people think of pharma as a category that is a little bit conservative, but it's really, we've seen more and I don't want to say hesitant, I would say maybe more cautious in the financial services area just because, you know, these are companies that do risk management as a core part of their offering. And I think a lot of them in my experience and conversation are taking a lot more of a cautious approach with AI overall. But certainly then that applies to the way that we're doing media buying on their behalf. And they want to have a lot more reviews along the way of like what's going to happen here? What's going to happen here? What data are we giving to this model? How's that data going to be used? Is it going to be shared? And if so, how, how is it anonymized? There are very different approaches depending on the client and the category for sure.
[22:24]
A
You mentioned earlier using Claude skills. Have you all consolidated your agentic work around The Claude model or are you using different AI models?
[22:35]
C
Different AI models. I would say Claude, I mean, CLAUDE was the first one that we used. I think in the agentic world right now in the. The use cases that we're exploring for us and I'm sure others are having a different experience, but for us, Claude has been the most useful. But no, we're also experimenting with ChatGPT, we're experimenting with Gemini, we're experimenting and we use a lot of Copilot here too because we're a Microsoft shop as an agency. So we're exploring a bunch of frontier models. But definitely Claude has been the one that's really at the forefront for the agentic work that we're doing.
[23:19]
A
How does that affect governance? Not even just managing different models, whether it's Claude, GPT, Gemini, et cetera, because those can be. You have to work with those models differently. I hate Gemini just because it doesn't care about the instructions as much as some of the other. It's not as obedient. But even with Claude, now there's Fable 5 and there's Opus 5. And one thing that changed with, I think it was Fable 5, maybe it was Opus 4.8, but you kind of need to be less prescriptive with the model for the model to be as effective. It's more about telling the model, here's my goal, kind of figure it out. Because it can do longer term reasoning and longer term thinking. And so whereas a year ago, even maybe six months ago, it was more. You had to write a 500,000 word prompt and tell it exactly what you wanted it to do. So how does governance change when you're using different models? Or is it the kind of thing where you start with, I think Opus 4.6 was, or maybe I think it was 4.6 was state of the art when you all were doing the first CTV test in December of last year. And so is it like a Microsoft Windows thing where if you're starting out on Windows 95, you're not trying to upgrade to XP because that's going to break every part of your workflow?
[24:50]
C
Yeah, I would say based on the work that we've done so far, we haven't found that kind of. The incremental model upgrades have made a huge difference in terms of what we've already put in place. It hasn't broken. Hasn't broken that stuff. I feel like I should be knocking on wood when I say that the next model is sure to break everything that I've said that out loud. And on tape here. But I think that goes back to what we were talking about before where I think having the human oversight is going to continue to be really important because as these models become more sophisticated, there is longer term implication for what we're doing. And to your point, just like with a new employee, you might need to be very, very prescriptive with like, here's how you do this, here's the next step, here's blah blah, blah, blah, blah. But then when you have your next meeting with someone on your team and it's someone who's been doing programmatic trading for five years, it's a very different conversation. And I think that that's kind of been our experience with the models as they've continued to be more sophisticated. Like the, the context is much more sophisticated and built in. But to me that's even more reason why it requires human oversight is because those inferences can still be as great as these things are, as amazing as they are, as much as when I use them, it still blows my mind every time. They are still unpredictable. And there are still things that happen within the inferences that are making that you're like, that is not right, not something that we would want. And it's really interesting things too. Stuff that you wouldn't expect. Like you have it create a PowerPoint for you which not related to agentic media buying obviously. And you're like, why did you make that logo look different? There's just stuff that happens that sometimes doesn't make a lot of sense. Which is really interesting and fun for us frankly because our team gets to tinker and experiment and see here's what it's coming back with and figure out why did it think this or why did it do this? But again, just more argument that I don't think anyone is anywhere near ready to set up an end to end agentic only media agency or media buying shop. It's a lot that's going to have to go into that.
[27:16]
A
And what about the seller agents? Because obviously you all can do a lot of the governance when it comes to your buyer agents. That's where you have the most control, that's where you have the most visibility. The seller agents though, there's a lot of trust that goes into those and obviously those are the responsibility of the sellers, of the publishers or the platforms. But how they interpret things that are said to them by your buyer agent can meaningfully change how a buy gets executed. Are you able to vet the seller agents? Are you able to see the system prompts that sellers have given as instructions for their seller agents.
[27:54]
C
Yeah, I would say overall, the work that we've been doing right now is so kind of early and collaborative that the answer to that is yes, they're very open and willing to share. Like this is how we're prompting our agent. This is what the agent is doing. I do see a future where that's going to change because at the end of the day it's a negotiation and they have their interests and we have our interests. And we've seen what kind of the proliferation of technology to wield greater negotiation power has, has warped and done weird things in our space over, over the years. Programmatic being, I think probably the most prominent example of that. So I think there's so much left to be written about the norms and how that's going to work. My fear is, and I think this is one of the main things that has caused issues in the programmatic space is what I just described, which is like the buyer has a certain amount of information, the seller has a certain amount of information. And there's a lot of trepidation about information sharing because you fear that it's going to be used against you in a negotiation. And just in a very. I'm not talking about people having ill intent or anything like that, just a very basic sense of like, I'm selling something, I want to get as much money as I can for it, I'm buying something, I want to get it for as little money as I can. And so it's not an environment that naturally lends itself to open conduit of vulnerable information. I think that is a really interesting construct as we think about agentic because the seller agent really needs good things like performance information. Obviously we're never going to share sensitive client information or things like that with, with a seller agent. But to me, I think we need to move towards standards where we can share as much information as possible to improve not just the buy, but the entire ecosystem which again, not to get too much on a soapbox, but part of the reason that we've seen such a proliferation of low quality content is the only signal that the buyer sends to the seller through dsp, SSP and all of those protocols is I want this for less money. That is not going to result in high quality content. It's not going to result in the monetization of things that I think are good for the ecosystem, frankly, not to get too heady about it, but good for society as we've seen. I think if we can come up with standards and ways of working that are going to allow us to. I would love to a future where through that kind of sophisticated agentic media buying model we can communicate to the sell side of the market. Like this is what our clients care about. So this is a financial services client. We don't want to be on some like low quality site with you know, eight ads loading and playing. We want to, we want to make a connection with a real person over here by by creating a value exchange that's really, really meaningful. And I want you seller and I'm obviously anthropomorphizing this to the extreme but like to make more of that stuff, I want the signal they get to be leading them in a direction that creates a better open Internet. And I do think agentic, not to be too maybe overly optimistic about it, but I do think agentic has the capability to do that because we can move from a hyper, hyper limited sophistication of more stuff for less money to here's what we're really looking for and here's the value that we're looking to create.
[31:37]
B
I think it's fair to say based on what you just said, the governing bodies of this industry that be have their work now cut out for them. They're listening to this episode as to setting some industry standards for this because
[31:51]
A
you're right and agreeing on those standards
[31:53]
B
and for them to be adopted.
[31:54]
A
Multiple governing bodies when it comes to hsd.
[31:58]
B
But nonetheless they've been taken to task based on what you said. And I guess we should be checking in with you at our next programmatic marketing summit and our next conversation given you guys are knee deep in the work.
[32:13]
C
Yeah, we've got, as I mentioned, more than half a dozen tests across formats, across categories, across platforms running right now. So we're excited, we're excited to see what this is going to, where this is going to point us next and what we can share with the industry because we think it's really important. And as I said, I tend to be an optimist. So I think that if we all kind of work together and share this stuff about what's working and what's making things better in our space, we could perhaps create a little wave of improvement instead of just racing to the bottom. So,
[32:53]
A
Scott, thanks so much for coming on the show. Really enjoyed the conversation.
[32:56]
C
Of course, it's a pleasure. Thanks so much for having me.
[33:06]
A
Thanks for listening to this episode of the Digiday 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 week day morning. Visit digidaire.comnewsletters to sign up.