The Negotiated Future: Ethan Settel on Newton Research, Agentic Media Buying, and the Reinvention of Media Operations
- 2 hours ago
- 46 min read
What happens when AI stops assisting media teams and starts acting on their behalf? In this episode of Signal & Noise, we talk with Ethan Settel, Head of Sales & Accounts at Newton Research, about agentic media buying — one of the most significant shifts in advertising today.
Newton Research is building AI agents that don't just summarize data, but connect systems, run analytics, build forecasting models, and execute media decisions directly on platforms. The vision: teams of AI specialists collaborating to plan, buy, and optimize campaigns with speed and precision no human workflow can match.
We dig into Newton's work with NBCUniversal, FreeWheel, Yahoo, and Locality — where buy-side and sell-side agents collaborated on premium video buying across linear TV and streaming, offering a glimpse of a future where software negotiates inventory directly with software.
Ethan explains why measurement and trusted data are the foundation of any agentic system — without them, automation just amplifies errors. Along the way, we tackle the industry's hardest questions:
How AI agents differ from generic tools like OpenAI ChatGPT, Anthropic Claude, and Google Gemini
Why data normalization has historically consumed most of a data scientist's time
How agentic systems can democratize advanced analytics for planners and buyers
The role of protocols such as MCP and AdCP in enabling agent-to-agent communication
Whether DSPs and SSPs become strategic platforms or simply "dumb pipes"
Where liability and accountability sit when AI begins making media decisions
Why human oversight remains essential, even as automation accelerates
His take on jobs: agentic AI won't replace media professionals — it frees them from repetitive work so they can focus on strategy and experimentation. If Ethan is right, the future of media won't be defined by better dashboards. It'll be defined by intelligent agents negotiating across the buy- and sell-side in a market more automated and data-driven than ever.
The negotiated future has already begun.
Watch the full episode and join the conversation.
🔑 What We Cover💡 Key Takeaways🎯 Why This Episode Matters
Read the full transcript below.
Brett House (00:01)
Hey everybody, welcome back to Signal & Noise. I'm Brett House, joined by my co-host Riel Longacre. And today we have Ethan Settle joining us, the head of sales and accounts at Newton Research. And we were just joking is that we typically have a cardinal rule not to invite salespeople to our podcast ⁓ due to the fact that oftentimes in our career experience, yeah, we made the exception because Ethan is an exceptionally smart guy. He's yeah, ⁓ he's not, he's promised not to shill Newton Research.
Rio (00:06)
There.
Ha
We made an exception, right?
And he's a friend of the pod.
Brett House (00:30)
But we have new research.
Rio (00:31)
Even though we do like the company
a lot, is why we, besides you being a good guy, we invited him on, and we know, but go ahead, Brett, I cut you off.
Brett House (00:36)
Yeah, it is
one of the more, you know, some of y'all do some shilling for, for, for Newton research. It's one of the most, you know, more interesting ⁓ companies in advertising right now. It's pushing past the normal sort of AI productivity story. Right. ⁓ It's not just about writing reports faster, summarizing dashboards. Newton is building basically marketing analytics agents. And correct me if I'm wrong, they connect data, run analysis, build models, create plans, support buying decisions for the actual media buying, agentic media buying process.
measure outcomes, that's a lot of coverage in terms of end-to-end media workflows, right? And then it interacts with other regions outside of that workflow, right? So some really interesting stuff that does not, from all of our research, ⁓ reek of AI slop or... ⁓
Rio (01:25)
Yeah, it's not like,
we can write a brief for you faster. It's actually doing real things, connecting real activities in media, which, and even some stuff, Ethan, you showed me at, you know, last time we got together, I was super impressed. So we're like, this is going to be fun.
Brett House (01:37)
Yeah, thanks for joining us, Ethan. definitely interested to hear your perspectives on both Newton research and the industry as a whole. And introduce yourself to the audience.
Ethan Settel (01:37)
Great, yeah.
I appreciate it. Thank you, Brett and Rio. Who has to do any shilling when you all do all the shilling for me? ⁓ No, I really do appreciate you all having me on. I'm really excited to participate. So yeah, I'm Ethan Sattel. I've been working at Newton Research now for two years. ⁓ I worked with the founding team at their last company as well, Data Plus Math, another company very focused on what we think of as the fun and exciting problem of media analytics.
Brett House (01:54)
You
Yep.
Ethan Settel (02:19)
And we were acquired by LiveRamp a few years ago before the team saw a chance to start it back up again. I've spent most of my career in media analytics. so finding another chance to do it and to bring AI agents to the mix has been really exciting. And just to talk about the challenges and the exciting things beyond even what Newton is doing in this space, I'm thrilled to have the opportunity.
Rio (02:43)
Cool. Well, the timing for this is really good, Ethan. ⁓ We know you've rolled out a bunch of new things that I think are super cool. But the timing is just generally when you look at the industry and where things are going, right? Like, there, are we going towards this agentified or agentic trading or an agentic or AI powered future in paid media and media buying? What will happen to RTB automated auctions? What will people do? mean, as they navigate these mazes of tools that, know, brief generators.
bid management tools, data pipeline. There's all these different tools that are dropping in that are AI based. They're actually kind of amazing. Like what really happens? Like what's the role of people? Will agents really be able to take a, let's say a goal from a person and then run across all these different solutions, build audiences, them, measure them, optimize them. Will this really happen? How will these...
How will communication work across them? It brings up a lot of really fundamental questions. And then what happens to all the DSPs, to SSPs, to all the different tools that kind of sit beneath this, right? There's a lot there. So this is going to be fun. ⁓ Excited to have you on here and let's, let's go into the questions, Brett.
Brett House (03:53)
Yeah, it's a packed thesis as you can see, but I think it's all interrelated. But you joined Newton early and I think you guys also, the founders, ⁓ sold, bought and sold integral reach as well, right? As well as Data Plus Math, which I've heard about a lot for a long time. Data Plus Math was active in the industry for what, 10, 15, 10, 12 years? Right before exit.
Ethan Settel (04:14)
Yeah, active for a while and then acquired by LiveRamp and kept as a name. think we actually, until this year, I believe this is maybe the first upfront that Newton Research has, or sorry, that Data Plus Math has not appeared as a logo in at least one TV Network's upfront presentation. we've been, yeah, it's been around for a while. It's got some staying power, that Data Plus Math brand.
Brett House (04:18)
Yeah.
Yeah.
Yeah.
So, yeah,
so what, I mean, outside of the founders and kind of pretty solid, very solid track records, right? What was meaningfully different, ⁓ you know, seeing that we're facing kind of a wave of AI tools and capabilities and promises ⁓ at Newton Research that kind of brought you on board?
Ethan Settel (04:53)
Yeah, I think part of that grounding was a lot of what drove me to want to lean in and join Newton Research early. You know, that team, the founding team has such a deep understanding of media and media analytics and really like saw firsthand in trying to deliver analytics to the media industry. Some of those challenges and like the breakdown between getting really great, impactful, powerful, unique analytics and measurement and reporting, but then turning that into actual
downstream optimization and changes to media planning and buying. You can have the best analytics in the world, but feeding that into that process of like, what do we do with this? Always felt like a hurdle. So I remember, yeah.
Brett House (05:35)
Yeah, it was always a promise,
Always a promise, but rarely delivered, right?
Ethan Settel (05:40)
Exactly.
Like you can have the best analytics in the world ⁓ and like just do nothing with it and sit on this gold mine. And when I saw the prototype of Newton, it was almost like three years ago now, like it floored me. I couldn't believe that how much of an issue this would solve for marketers and advertisers, the clients I worked with at data plus math at live or even before that, ⁓ you know, taking all of this hard work and this brilliant effort from smart data scientists and turning it into something real from like a downstream impact.
Rio (06:09)
How did you do it so quickly? Cause I remember like this is just three years ago. I remember like shortly after you joined, we took a look at it. I was surprised how far ahead you were. Other people were just messing around with like prompting, right? And prompting basic prompting, There was not a lot, right? You actually had a product. How were they able to do it so quickly? And what about their understanding, I guess of advertising and media enabled that.
Ethan Settel (06:30)
Yeah, I think it comes from a nerdy data scientist bias that we think the hard part of media is analytics. We're very biased, I admit that. But I think by starting with that, when the continual rising tide of AI skills and agentic skills and knowledge, to your point, the ability to take natural language and turn it into like a visualization or an interesting story is almost like table stakes now. But the ability to like get at the
Brett House (06:57)
yeah.
Ethan Settel (07:00)
consistency and rigor of, how do I take the output of an analytical model and turn it into a scenario planning exercise or forecast the impact of media and then actually go make that that's really challenging.
Brett House (07:10)
Yeah, or
update my segmentation schema and activate those segments across, you know, XYZ platforms. Yeah, that's the part that breaks often, right?
Rio (07:18)
And
to do it in real time with any frequency at all, right, is very challenging.
Brett House (07:22)
Yeah.
And it's amazing how many companies just at the very most, at the most basic level, even struggle to get into a tableau output, right? That's trustworthy. ⁓ You know, and so to think that we're, this is just light years ahead of that. This is saying not only are we trusting the output of the, of the analytics exercise or data science exercise, but we're, now connecting this to other work streams, which are critical to planning, scenario planning, budget allocation, media buying, all that stuff. And is that
So how does that work? I mean, how does that work within the platform? And what take us through that those kind of hops?
Ethan Settel (07:58)
Yeah, absolutely. It starts with that grounding of analytics that I've belabored plenty already. ⁓ It's ensuring a consistency, right? To your point, in order to get to that hop and that leap of, I'm going to go make downstream changes, the teams, both the analytics and the media team, need to feel very confident that it's going to give you the same answer when you ask it to do a workflow over and over again. The last thing you want is for the same question to say, hey, great, move 20 % of your budget here.
Brett House (08:03)
Yeah.
Yeah.
Ethan Settel (08:24)
it there, it's going be better. And the next time you ask it a minute later, it actually 10 % from here over there, five over here over there. That's not going to get any confidence. So there's that piece of it. And then the other is enabling downstream agentic connections to different media providers and also the relevant data environment. agents that are able to speak different languages can speak to get technical, an MCP server with publisher number one.
Brett House (08:32)
Yeah.
Ethan Settel (08:52)
to be able to go call an API with data provider number two, and then also have an agent to agent connection with a walled garden over here and a clean room over there. I think establishing that interconnectivity is also really crucial because most people aren't planning and buying and analyzing media in one central platform. They're doing it 20, 30 different places. So that connectivity point was really important to establish as well.
Brett House (09:11)
Yeah. And there's an agent that's
built, that's got skills that are built to handle or maybe a swarm of agents that handle that specific workflow or that whatever the individual workflow might be. You get different agents specialized in speaking to an SSP, speaking to a, you know, a data warehouse, whatever it might be. Is that right?
Ethan Settel (09:29)
Exactly. We think of
it like a team of specialist agents with a coordination layer or even an orchestration agent who's going to ensure that each step that each agent takes is done consistently, that work is handed off properly, that each agent has a set of skills that it can apply to each of those different environments and each of those different workflows as well. So it's really like a multi-agent platform. And then to make it work well for clients, they need to feel like they're bringing in their own individual expertise.
and programs, then getting data in a way that aligns to their specs, their decisions. So having skills with each agent that can do that and then have them hand off work consistently to one another is what we spent a lot of time building from like the orchestration layer in our platform and our product.
Brett House (10:13)
Yeah, the, I had a question. just wanted to interrogate an earlier part, right? With the analytics output, because I think it was an important point to make, because it ties to me this notion of like a memory system that's repeatable. And it's kind of the problem with LLMs. If you're not training in LLM, especially a public LLM on kind of a core set of a rule book, a data dictionary, know, MD files, et cetera, you're going to get a different answer to that same question every time you ask it. And then if you move that same...
Rio (10:16)
Go for it.
Brett House (10:43)
scenario over to another LLM like Claude from GPT, you're going to again get different outputs that aren't aligning with anything that you think you might have prompted previously. So tell us about that memory system, because that to me sounds like it's your proprietary tech. You're building this probably on premise or within the client's ecosystem, right? And it's learning and developing and improving and you're building a kind of a battle tested system that can
Rio (10:46)
You'll get a radically different answer, yeah.
Brett House (11:12)
ensure that your outputs are consistent based on whatever the client asks for. This is total supposition on my part. I don't know. that?
Ethan Settel (11:17)
Yeah. No, you're,
you hit the nail on the head here because we almost think to break it down into two categories where there's our skillset, what we bring into Newton, we bring into our agents, which is really that kind of as exhaustive as possible compendium of marketing science and media knowledge. It's like almost like the marketing science playbook.
Rio (11:38)
And this
is what Newton is bringing specifically, your corpus of knowledge you're bringing to the solution.
Ethan Settel (11:41)
Exactly.
Brett House (11:42)
Yep.
Yeah.
Ethan Settel (11:46)
Exactly.
Inclusive of like examples of the exact sort of code and exact sort of models that if I'm an expert data scientist with six or seven years of industry experience, I can pull out of my back pocket almost at any time. So having that corpus of knowledge there, but then also ensuring there's a place for each individual, whether it's an advertiser, whether it's a publisher, whether it's even a third party involved in ecosystem, they have a chance to bring their own knowledge base, their own approach. Like if everyone's taking the exact same analytical approach to every single problem, it's like
Brett House (11:58)
Yeah.
Ethan Settel (12:15)
borderline media analytics slop. So like having the expertise and the specialization that any given data scientist or any given media buyer or planner can bring to those agents and making sure that lives somewhere that's proprietary to them, but is equally referenceable by Newton's agents for their work. That's critical to making sure it does something unique, different and repeatable.
Rio (12:34)
You know, it'd be interesting to maybe talk to the viewers about, um, cause some of them, some of them are probably very deep in AI and are using tools and have, and have their own agents that are running, running tasks for them. But some people might be a little less versed in that. So it might be interesting maybe to take a moment to talk about, well, Newton has these agents that are connecting data. They're writing code, they're running models, they're running analysis, like different agents are doing, everything. Brett, to your point, are they have different, different sets of skills they're bringing? How would this be different from someone just
a normal chat bot or let's an AI interface or people are prompting it. Maybe, maybe spend a moment on that for people who are a little less deep in AI.
Brett House (13:08)
I'm sure this is something you have to pitch to clients all the time where you're like, why can't I do this on my own? I've already got Gemini, I've already got Copilot, I've already got ChatGPT, we've got Teams or IT teams set it up.
Ethan Settel (13:10)
Yeah. yeah.
Rio (13:13)
Why can't you actually BT do this for me? Yeah.
Ethan Settel (13:19)
Yeah,
it's a great case. I think I'll use like an example of let's say somebody wants to do an exercise where they've run a media campaign, they'd like to do some sort of analytics, they'd like to use that analytics to inform an optimization for the next campaign for the following quarter, you name it. If you go to a chat, GPT, a Claude, a Gemini out of the box and say, hey, here's a whole bunch of data. I want you to go tell me what worked with didn't and use that to inform my next plan. It's going to go pull from a broad
large knowledge base that's got hundreds or dozens or hundreds of thousands, hundreds of millions of different data points that are completely irrelevant to the question at hand. It's going to lack that contextual knowledge of all that work. It's also going to lack any sort of specifics unless you spend a lot of time feeding it in to any one of those out of the box agents on the backend. ⁓ Whereas with Newton,
Brett House (13:56)
Yeah.
Yeah. And the
contextual places, that's a really important point to double click on that. Building that almost like a context map, which is your blueprint for how it should behave to, you know, it's what the semantic layer does. It's what MD files do, markdown files. They give that instruction booklet, right, to encode how it should be doing these things and what it should be looking at and how it should be interpreting those things, right?
Ethan Settel (14:27)
Exactly.
Yeah, there's almost like multiple different semantic layers or different reference libraries where it's that that reference library of like all the things I could do in the media analytics space and constraining that there's the reference library of, hey, for this specific data set that you need to pull from to do this analysis, there's a lot of nuance and there's a lot of potential pitfalls you might take in how you write and generate code and build and train a model where if you aren't really tidy about how you think about the parameters you're bringing in, how you're normalizing the data before you bring it in, how you're
Brett House (14:38)
Yeah.
Ethan Settel (14:56)
referencing historical benchmarks, you're going to write and generate a bad model. And then not to mention also that consistency layer that a given advertiser will want to bring and say, ⁓ but it has to look at this prior, it has to ensure that this specific media element or this channel, this creative is treated in a unique way because a client cares about it or because there was some nuance in a campaign. So all of those pieces have to be taken into account. And Newton is able to look at all those things and build models.
Brett House (15:17)
Yep.
Ethan Settel (15:24)
execute code in a way that is consistent across each of them and repeatable again every time you want to go run them.
Brett House (15:30)
Yeah, no, that's good.
Rio (15:31)
Who would be using
this like at an agency or let's say a brands team? Like who is the target user of this? And then I think also like who would you be, might be different person, who would you be selling this to?
Ethan Settel (15:43)
Yeah, it honestly is a bit of a mix. We typically at some point or another, like we come very tight with the analytics org at an agency or media analytics org at a brand measurement team. It's going to be their expertise that really feeds into Newton's approach for any given workflow for that client. But then also it's really crucial and important that, let's take the example of an agency, the media planning team, the media investment team are also brought in because one of the benefits of
Brett House (15:53)
Yeah.
Rio (15:55)
Measurement team, yep.
Brett House (15:56)
Yeah.
Yeah.
Ethan Settel (16:13)
good agentic solution, I'll paint with a broad brush, not even Justin Newton, is that you're democratizing access to some of these more advanced technical workflows, in our case, analytical workflows, to a media practitioner, someone who historically would have to go wait in a queue to ask a data scientist for a question, can now press a button, inherit all of those data scientists' best practices and models, let it run and let it output something in a way that speaks to the next action they have to take. It's going to output a change to a plan, or it's going to go directly into a platform and pull a lever.
Brett House (16:21)
Yep. Yep.
Ethan Settel (16:41)
to change a bid factor, to adjust a creative weight, to actually go shift dollars between different media elements. That's the piece that a practitioner needs to be able to do. And if it has to go take a pause, pull it out of the agent, bring it somewhere else, put it there, then you're still bringing in another manual element that's right for error and distraction.
Brett House (16:58)
Yay!
Yeah, and especially if you're trying to make ⁓ relatively quick, we won't use the word the trite term real time, but you're trying to make relatively quick changes to budget allocation, to media, to audience targeting parameters, et cetera. You need inputs that are relatively quick and trustworthy. And to your point, the media planner is not going to be the one pulling out those analytic outputs, right? But it sounds like from what you guys have built, to go back to that topic of what you've built as a core proprietary tech that incorporates
kind of key elements from the brand, right, is really kind of well founded in governance, right? Model governance, data governance, all that stuff because that's what's ensuring, one, it's a moat for you guys as an organization, but you really can't, ⁓ it's ensuring that your outputs are consistent and aligned to the client's needs so that they can, when they do actually take an action, like a budget allocation, it doesn't come back to,
to haunt you because it was the wrong action. do you guys have, it's a constant optimization process, but do you guys find that it's driving performance upgrades and improvements, just not just in speed and efficiency, but also in actual outcomes?
Ethan Settel (18:17)
Yeah, no, absolutely. And we see that really just when you're able to take a broader view or more, I'll say both broader and more granular, you're able to look across more campaigns more often because you no longer have to wait multiple hours or days to get a data scientist to actually go and dig into something. It means you're able to more rapidly respond to changes in campaign performance, changes in the media landscape overall, right? Things.
Brett House (18:36)
Yep.
Ethan Settel (18:45)
that the ground is going to be constantly shifting under you as a media planner and a buyer. And so if you're not able to respond and act nimbly, then you're going to lose opportunities to, you know, capitalize on moments of Mars. Exactly.
Brett House (18:56)
Competitive challengers or whatever. Yeah, it could be economic factors
that are impacting some sort of campaign performance rate. You need to be able to shift fast, right, in order to take advantage of something that's happened in the econometrics world, right, the seasonal world, whatever it might be, these exogenous factors that impact, you know, ad spend and ad effectiveness, yeah.
Ethan Settel (19:14)
Yeah.
Yeah, 100%. And just by sheer nature, being able to do more of that and respond more quickly to that alone, you're already seeing performance gains. And then when you even layer on, certainly happy to spend forever talking about advanced data science, causal models and things like that, you kind of get into like Greenfield of, these are things and approaches and methodologies and models that no one's been able to run because it's just, there's been an upside down economic impact of like, don't want to invest so much data science time and compute cost.
Brett House (19:43)
In compute time,
Ethan Settel (19:44)
time
Brett House (19:44)
yeah.
Ethan Settel (19:45)
on all of these things, but now all of sudden, like, wait, can do some really interesting multivariable stuff here that I couldn't have done before for a given media campaign. And we're only just scratching the surface of that, but that stuff gets my data scientists really excited.
Rio (19:45)
Get out y'all.
Brett House (19:54)
Yeah.
Well, I think we
got to be going to make a call it here is that Ethan is proving that he is not just an average seller. Like you are a like, yeah, you know, you know, consultatively, like what this stuff actually means and how it works, which is it's super impressive. And it's funny because there's a core story that I'm hearing.
Rio (20:04)
I vouched for you,
Ethan Settel (20:04)
Yeah.
Brett House (20:15)
once you get, Noreo and I interrogate the shit out of these AI companies that we talk to, right? You get below the surface and you're like, okay, what does this really mean? What is this founded on? Where is the governance layer? Where's the underlying data layer? ⁓ Et cetera, et cetera. And it's a common story on how agentic works in general. ⁓ So you guys are doing some really interesting things with, there was just a recent announcement with NBC Universal, RPA Freewheel.
Right. That was described as the first of its kind proof of proof of concept for cross platform, agentic buying across linear and digital premium video. That sounded pretty interesting. How can you? Yeah. Yeah, exactly. Like bridging the gap that everybody's been trying to bridge for decades as linear. Yeah. Linear declines like apply everything that we just talked about in terms of the marketing analytics discipline, like the the the ⁓ the
Rio (20:52)
Which should be kind of amazing, right? I mean, to actually do that.
For the last couple decades anyway, yeah.
Brett House (21:10)
the diligence or governance and then the activation connectors so that you can act on these outputs quickly. How does that fit into this announcement that you guys made?
Ethan Settel (21:22)
Yeah, it's, all the, all of the analytics grounding is a footnote in that announcement. But like, I think as you drill into it, it's, it's a critical piece because it's, it's that analytics that then feeds the, all right. I've analyzed, I performed, I understand what I'm trying to do here. Now, how do I harness the power of an agent to connect into, this case, NBCU and freewheel. I think Yahoo and Locality were also part of that announcement as well. So it's.
Brett House (21:31)
Yeah.
Ethan Settel (21:50)
from like the most premium of premium inventory and live linear sports through to streaming through even like local linear and streaming and come more programmatic streaming and premium video as well. Have an agent that can go speak to each of those individual, whether it's a seller agent and API and say, all right, well, here's what inventory is available or here's what's been negotiated by you already that you have flexibility to move in and out of. Let me take all of the insight and the output I've now done on that footnote of
Brett House (22:08)
Yep.
Ethan Settel (22:19)
I've done all this analysis. know what's working. I know what's not. I've got some user, some planner input of like, this is how things are changing. This is what the brand, the client wants to see and expects to see. Bringing all of those things together and then establishing a much more seamless agent to agent. This changes. This is the inventory that's available. There's obviously a lot of hand holding along the way for some of these earlier kind of first mover ones. Everyone wants to be very careful that the agent is going to do things correctly, that it's not going to go say,
Brett House (22:43)
Yep.
that it doesn't hallucinate.
Yeah, yeah.
Ethan Settel (22:48)
Yeah, it's
not gonna say, ⁓ I'll take a spot here that I'd allocated for somebody here in live sports and just throw it somewhere else. Yeah.
Rio (22:52)
Yeah. Yeah. Being cautious is good. And I think,
you know, people have also heard so much about hallucinations and problems. So I've been thinking people are cautious, maybe overly cautious, but yeah, sorry, cut you off. Keep going, Ethan.
Ethan Settel (23:03)
No, no, it's fair. I'm overly cautious, especially when there's millions of dollars at stake here in important television and streaming buys.
Brett House (23:11)
Yeah, and what
do think the core problem that speaks of them? it sounds like to me you're solving like a core fundamental fragmentation issue with that media buying process because these places, they live in different platforms ⁓ and different ad delivery systems when you...
Rio (23:24)
Well, in fact, fact,
that gets to a question I wanted to bring up, Ethan. This is interesting. So I remember like, so you guys started a measurement, which makes a lot of sense to me. Like that's your background. you have a bunch of people, deep market analytics, data science, marketing science background. And I thought what you, what you initially rolled out, like I said, a few years ago is really impressive. It was quick. It was quick to market. When we met last time, I think it was at CES. It might've been right. I remember we had, you quickly demoed me some of the new features. You were actually showing me a Prisma file goes up here, which is, the media.
Ethan Settel (23:27)
Yeah, for sure.
Brett House (23:32)
Yeah.
Rio (23:51)
planning file, right? And you were showing me actually, you know, but I saw a bunch of rows and these are each row was in activation on a different, on a different platform for different DSP. All the bid information was there that I mean, so this is really seems like it's a much more, but so much broader solution than you initially had love. You can co I don't know what's ready. What's not what you can talk about now, but love to understand like art, like it sounds like you're expanding into that activation space and Brett, some of the stuff you were just talking about. Sounds like it's, it fair to say that you're going in that direction?
Ethan Settel (24:20)
Yeah, absolutely. I think it's a big piece of what we're doing now. Again, I'm going to caveat my next piece by saying I'm once again very biased in that I think the analytics is the hard part and the tricky part. But starting with that to ensure that when an agent is making buying decisions or taking a buying or media plan from a client and executing it, that it's doing it with the right sort of grounding, think is the harder and the trickier part. think we've exactly.
Brett House (24:31)
Yeah.
And the part that needs education, right? Client education,
yeah.
Ethan Settel (24:51)
Yeah, ensuring that the clients understand that, hey, when you go and make these trades and you make these decisions and changes to your media plan, that you're doing so ⁓ with intelligence, with purpose, with expected goals in mind so it can be monitored. Yes. Yeah. And then we have then spent a lot of effort recently as well to ensure that the different connectivity points, the different ways you have to interact with a seller agent here or an API there, an MCP server there, that you're also
Brett House (25:03)
with guardrails governance yeah
Ethan Settel (25:18)
dealing with the realities of still how some of these inventory sets are bought and sold, which is definitely not programmatically for the likes of linear and much of high quality streaming video. ⁓ And so ensuring that you have ⁓ an understanding and a grounding in how to go interpret a Prisma file, for example, and how to go pass that into a different type of seller agent at every kind of relevant endpoint, and then bounce those line items, those creatives, those et cetera against each relevant
agent or endpoint, and then orchestrate that by in a way that is consistent, accurate, efficient, ⁓ then recurringly monitored.
Brett House (25:56)
And is
it also in a simple way outside of just a pure managed services layer, which would be person communicating with other person on the other side of the client. Is there a way that it communicates exactly what's happening, deconstructs?
⁓ how this is being rolled out and the reasons behind it, you know, so that, you know, whether it's a readout, whether it's some sort of communication that's integrated into Slack so that buyers know without having to get somebody on the phone, or is that more of a managed services play to help people understand what the heck was just done by those multiple agents?
Ethan Settel (26:30)
Yeah, okay.
Yeah, one of,
yeah, a key piece of like any good, I think any good AI tech or any good agent tech right now is transparency. I think there's a, there are plenty of companies out there that ⁓ are going to like obfuscate a lot of the reasoning and the logic behind what an AI agent is doing. And we've from the start have been like on the other end of the spectrum of full transparency and not just like to your point, a readout of like, this is what I'm going to do. This is why I'm doing it.
Brett House (26:42)
Yeah.
Yep.
Ethan Settel (27:03)
by the way, also, here's the code I wrote and the analysis I did that led me to that decision. And here's the code I'm going to write and the message I'm going to pass to each endpoint. ⁓ You don't have to look at all that. It's far too much in a terminal frame to look at,
Brett House (27:14)
Yeah, but that might be particularly
useful for the data science team to say, oh shit, something broke, right? And that's the wrong, or there's something wrong with that code and they can actually troubleshoot in real time because they're seeing that. But the media buyer might just want the actual contextual readout of what the heck's happening. I want to know, can you deconstruct the linear piece? Because that seems to me from an ad serving model compared to streaming channels of all types.
⁓ is different through different pipes, right? You know, the ads are added and inserted in advance, like during the content sort of build process. How does that work on the linear side of television?
Ethan Settel (27:54)
Yeah, I think it's somewhat about meeting the publishers in the industry where they are. So you're spot on in that a lot of linear buying is not necessarily agentifiable, to use a terrible mishmash of words. ⁓ And so it's like bringing in agentic skill sets and knowledge and workflows where they can. breaking or accelerating some of the communication between what's available.
Brett House (28:08)
Yeah. Yeah.
Rio (28:10)
No, why
not?
Ethan Settel (28:22)
⁓ and what you want to buy, ⁓ to translate budget line items, to translate creatives that already exist inside a planning tool or inside ⁓ an environment. We're not necessarily reinventing a lot of that, but we're again trying to meet each side of the equation where they are and then streamline where we can.
Brett House (28:37)
Yeah.
So for linear, would be like you're basically giving a recommendation for the next buy because they have to plan for, whereas digital, you can make those changes much more quickly, right? You can do creative swaps, can do budget allocation swaps, all that sort of stuff within the digital pipes or programmatic pipes, right?
Ethan Settel (28:57)
Yeah, I
think it's part of the grounding of our agents is know what you're talking about. And for one, right, don't go say, ⁓ you know, I suggest a reallocation here. Hey, that, you know, Sunday Night Football spot worked really well. Let's buy 10 more of those tomorrow. Like it doesn't work that way. Like, let's make sure you understand what you can and can't move. ⁓ And then like, let's also like understand the levers you have though, right there, especially when you get more and more of these on the linear side, these publishers building their own seller agents, building their own interfaces that are going to start.
surfacing more and more things like here's where there's maybe remnant inventory available. Here's this that or the other here's where you might have some more things you can do from different audience based planning like that's just going to open up more and more opportunities for advertisers with good agentic tools at their disposal to then start taking advantage of that and being more nimble with how they plan and buy even more static type of inventory.
Brett House (29:47)
Yeah.
Rio (29:48)
Yeah,
no, it makes sense. mean, like, you know, linear may be declining, but it's still tens of billions of dollars per year, right? It's still a lot of money goes into it. But looking at measurement, I remember if this is five or six years ago, I remember like I was like at a consulting firm. were trying to client an agency client had wanted us to build this meta DSP. Like we started at measurement. I remember we, looked at and we were even just looking at retail media. We weren't even looking at DB 360 or the trade desk or even open web stuff, right? This is just stuff where the inventory is fairly consistent. Even
coming up with a common data model based on their APIs to even report in one place was like so difficult because like everyone calls things, they're roughly the same, they call them different things. They slice and dice them different ways, because that's their secret sauce, right? It was almost impossible, I remember. It took us like months and months and months and we ended up just having different screens for each one of the platforms. We had like an Instacart one, an Amazon Ads one.
Brett House (30:45)
Yeah.
Rio (30:46)
a Kroger's one because it was just like, it kind of broke them all. How does like, did AI make that easier? How did you solve for that? I'm very curious.
Ethan Settel (30:54)
Yeah, I also want to respect that you clearly, you know and understand the measurement challenge. It's sometimes a little hard to like get people who aren't involved in measurement to know like exactly that issue where it's really not as easy. Like you think something ran, I should just know what happened, right? There's a lot of places where that can break down. Yeah. And so there, you know, there historically has been kind of a need for massive amounts of manual work. I early, early in my career, I think a lot of the work I was doing as like empowering analytics.
Brett House (30:58)
No.
Rio (31:07)
Not that easy.
Brett House (31:10)
Yeah.
Ethan Settel (31:23)
for measurement was like that rectification of all of the different things that could go wrong about how you've named and labeled the campaign across all these hundreds of different platforms, then all the different grains, there's no common denominator against which they're giving you data back. Historically, someone has had to do that manually. ⁓
Brett House (31:41)
⁓ yeah.
Rio (31:41)
Well, people wonder why do agencies give like
a monthly report in Excel? Like that's why, right? Because it's so freaking hard in this manual. Sorry, cut you off. Keep going.
Ethan Settel (31:46)
Yes. No,
exactly. like there's, there's, will say there's really great ⁓ ETL tooling, ⁓ both agentic and non, but I think even better now with some of the agentic tooling that exists that it helps and solve that just that problem alone. But we knew that was not always going to be the case and not every company and advertiser we work with is going to make that level of investment to ensure that there's some normalization process. So ensuring that our agents and are also really nimble and good at like.
doing that on their own as a starting point, but then being able to reference and use any existing knowledge base that the customer has at their disposal. So again, maybe they're getting some little page somewhere from publisher that says, hey, by the way, this is how we named your campaigns and this is the difference. ⁓ Being able to point an agent at that, bring that into its corpus, bring that into its knowledge base specific to that data set, reference it only when using that data set and then understand that you're to have to cut the other data you get from.
Brett House (32:29)
Yeah.
Ethan Settel (32:39)
an Instacart a different way, from Amazon a different way, from retail media network B a different way. And so even where there isn't pre-normalization done, meet the client, meet the advertiser where they're at with whatever state their data is in. And then bring whatever level of normalization you can before you start bringing that analysis. But then not having to repeat that every time. It's awful experience in advertising. Every time you go and want to do an analysis, have to re-normalize the data and figure out how to make sense of it.
Brett House (32:44)
Yeah.
yeah.
Yeah. And that's what's taken them in so long in so many cases, right? Is that data normalization process, right? It's it's, you know, you could have multiple data warehouses, different, different types of data. and it just to bring that all together and to match it and, get predictable outcomes was a very FTE heavy process that took months and months. And I've seen it in, you know, I've seen a multitude. So we had a really good conversation with that.
Mike Finnerty is a guy that I worked with at New Star and we were running their MMM practice. He's now the US president of Mutinex, which is an agentic based ⁓ MMM tool that basically is taking off the FTEs and replace them with kind of agents and agent swarms to do different tasks. Are you guys, did you guys, that where you started and how do you, with that in mind, knowing that there's this kind of next gen, non-people based, non-managed services type approach to the new.
to MMM today, where do you guys fit into that mix and are you guys a disruptor? Do you take place of it? Do you compete with it? Or is it just an output that goes into those kind of tools?
Ethan Settel (34:13)
Yeah, it's a fair question. think we feel really good about our MMM capabilities because it is a, it's an interesting topic because there are people who are very passionate about the way in which they have built their media mix model ⁓ or that have a really good long standing consistent relationship with a third party that has done media mix modeling them for years. And they reference that model and every run of that model is a comparison point to the previous run. And so,
Brett House (34:39)
Yeah.
Ethan Settel (34:40)
If you go in and say, don't worry, I'm going to build a new one for you. It's going to be great. I'm no, it's not. I've got 10 years of experience with this model. And if you run a new one, I have no reference framework for it. And so we've built capabilities or agents where if you want to build a model from scratch, our agents can walk you through that. have different approaches. It's a Bayesian methodology. It's methodology. It's you want to use like more of one of these open source packages like a Meridian or a, I always forget the other name. I apologize. I figured if that was the matter of the Google one ⁓ or a pie MC package. Great. can use that.
Brett House (34:46)
Yeah, exactly.
Rio (34:46)
Right.
Brett House (34:49)
Yeah.
Yeah.
Ethan Settel (35:09)
Or if you've got like your own existing model that you really trust and you really love, great, like show us the saturation curves. We'll take those, we'll take the existing kind of parameters of the model that you've already built and Newton can just accelerate the iteration of that and he can incorporate other more grander things. Like we'll take the incrementality testing that we're doing and then use that as another signal to improve the model. So there's that formation of the model, the inputs of the model and then the output. How do I actually take the model and feed it to something meaningful? Because I think that's still the biggest.
Brett House (35:30)
Yeah.
Ethan Settel (35:38)
disconnect even with those really strong MMM providers and the agile providers. It's like, how do you then take that and then feed it into scenario planning, into forecasting, into optimization at even a more granular level than at the channel?
Rio (35:51)
Look at it. FDF T is that you mentioned a second ago, Brett. So Ethan, I'm curious, like how are you going to market with this? And I guess like the reason why I'm asking this is like, you could go to market as okay, we're going to augment what, people are doing, like, you know, agencies, cause some of them may be worried about, mean, they, want to have show clients that they have savings from AI, but obviously having their model disrupted can be a little scary. But then I talked to, I don't know you've ever ran into Adam Epstein, he's the CEO of Gigi, right? They have this, they have this really cool solution that's been completely redone and a gentrified.
In fact, used to, another podcast I was on, ⁓ Gloria, who's his head of product. She was my co-host for a while. Like they're great people and I've been following them. Adam was saying, this is wild. His pitch was he would look for agencies that were hiring, let's say traffic, traffickers, right? And he would call the agency, Hey, I saw your hiring of an ad for like 10 traffickers. What if I, what if I told you this solution could do the job of five of them? And he said, that's how we actually.
trades his leads as he looks for ads and then pitches the head of the agency or the head of programmatic. I'm curious to how are you going to market? you, it like a labor replacement? Is it augmentation? I'd love to hear how you're doing that and how the market's responding.
Ethan Settel (37:03)
Yeah, no, it's I feel like it's the age old age old three year old question with AI. I'm like, to what extent is this just going to take people's jobs or is it going to just make them more efficient? You know, I think the reality is that getting AI agents that are really good at analytics that are really good at the work was we talked about, like they're going to streamline and make things a lot faster. And they still need some kind of guidance and a human in the loop to just make them
better individually for any given agency. ⁓ If I take the agency use case there, Again, I kind of made the point earlier, if everybody's doing analytics the exact same way and using the same agents, then there's no real competitive edge. But if you've got agents and technology that say you've got some really, really smart people on your team, you've got really smart practitioners or skilled, talented practitioners in your analytics org, in your trading org, ⁓ you name it, to bring that skill set and that experience that they have from doing this work for a long time and then
Brett House (37:45)
Yep.
Ethan Settel (38:03)
building it into a specific knowledge for these agents to draw from. Now all of sudden, you're bringing them to more and more campaigns, more and more clients. For some folks, they may say that negates the need then for me to go hire a bunch of junior people to do that work. But for others, well, now I can just do more. We have this whole concept that we tell, it's almost like unlimited analytics. What if you could bring unlimited analytics to every client if you're an agency or every campaign if you're an advertiser or a publisher? ⁓
Brett House (38:19)
Yeah.
And
yeah, and you have this continuously learning system, right? is, Rio and I've been talking to George Musi, who's, who's doing some fractional work as a chief strategist for her office, for her company called Mad Sense. And that's exactly the story that I'm hearing from, from like the true applied AI companies is that you put it on the client side and you're building not only IP around all those inputs from the data analytics people, the data science people, right? That builds over time. It's not going anywhere because you're, you're, it's, you're, hosting it, managing it within your own firewall.
whatever, right? And it spans time, right? So there's a temporal play to it so that you've got this knowledge base that helps improve processes, improves outputs, improves quality or fidelity of the models, all that stuff, right? Is that where you guys are really leaning heavily to build that IP for the companies that you're supporting?
Ethan Settel (39:19)
Yeah, it's building that IP. It's building a functional layer for them to apply it and scale it. I think both of those pieces are critical.
Brett House (39:25)
Yeah.
Yeah. This is applied AI. is what, know, cause, cause people like, it's almost like, how do we define AI? And it sounds like you guys are doing this where it's like, this is going to have not only short-term improvement gains or efficiency gains, but it's going to have this long-term impact because the more you train it, the more it learns, the more governance that's built into the system, the better the outputs are going to be over time.
And suddenly, to your point about the fear that you're gonna lose employees, ⁓ you're just gonna make everybody operate at a much higher level and the company operates, your clients operate at a much higher level because of this IP.
Ethan Settel (40:02)
Yeah. Yeah. It's a
thesis of like, if I'm a big complaint from the data scientists at the previous org that we all worked at, at Data Plus Math, was you spend the 80-20 rule, right? It's everyone's great analogy. Everyone loves it. But you spend like 80 % of your time as a data scientist is just fighting to normalize data, fighting to fix changes in models. And like, if you flip that equation, 80 % of your time is spent iterating on new models. And I kind of mentioned like having more firepower to do more interesting, more granular modeling work.
Brett House (40:14)
Yeah.
Yeah, totally.
Ethan Settel (40:31)
Like you start opening up new ways to investigate media performance, new ways to think about optimizing media performance that you couldn't bring to every campaign. I think that's the most exciting thing.
Brett House (40:35)
Yeah.
Yeah. It's a level of
innovation versus the kind of junkyard data management stuff that you have to do for most of the time, right? The 80 % of the time. So I'd love to shift in real time if you think this idea of like agent to agent collaboration, there's been a lot of talk about protocols in the ecosystem with MCP, a model context protocol, add CP. Yeah. Can you tell us like what's the most important thing that we need to understand around the sort of protocol layer?
Rio (41:02)
ADCP, UCP, there's a bunch of them, yeah.
Brett House (41:12)
when it applies to applied AI. ⁓
Ethan Settel (41:15)
Yeah,
I mean, I think we definitely feel, I would say I feel self-important, like individually, the importance of having like good protocols, a set of standards and language so that like agents can talk to each other. And you know, don't have to know how to speak five or six or seven different languages anytime you want to go place a media buy, analyze a media buy, things like that. ⁓ You know, that's kind of the approach we've taken at Newton where it's like we're, and we're part.
Brett House (41:37)
Yeah.
Ethan Settel (41:42)
of AdCP, we're also part of IB and their like AMP work. We also like built a lot of MCP server connections and the different agent to agent protocols that are existing. Like that stuff to me feels almost like, everyone's going to have to know how to interact and interface between different agents. And, you know, it should be table stakes when you're building agents that they should be able to go interoperate across all of these different protocols. And yeah.
Brett House (42:04)
Yeah, it's reducing friction. It's reducing friction across
these different jumps that the agents inevitably have to make.
Ethan Settel (42:09)
Yeah, exactly. And whichever, if there's one predominant one that ends up winning, so to speak, and becoming the one protocol, I think the industry will quit to snap to and adopt to that. I think the agents themselves are very, very good at getting even better at just reading the protocol documentation, reading the API documentation, reading MCP documentation, and knowing, not quite instantly, but very quickly, how to then write to that, how to speak to that. And so it feels like it's, again, I'm very biased.
It's almost like not the battlefield. It's like the biggest and most important thing to fight over. Yes, agents need to be able to speak to each other and speak a common language. I don't want to diminish the importance of that, but which language they need to speak to each other feels like, you know, the agents are pretty good at speaking a whole bunch of different languages for lack of a better analogy.
Brett House (42:53)
Yeah, that commoditizes
Rio (42:54)
No, they're at that.
Brett House (42:55)
quickly as long as they're trained on those protocols. It'll just seamlessly operate underneath the surface, right?
Rio (43:02)
If an NBCU example, I thought this was kind of interesting. Like Newton had buy side agents and then NBCU and Freewheel, had their sell side agents, right? On the supply side. And I mean, do you think this is the future? Do you think this is like, this is how things are going to play out? And if that happens, like what happens to all the tools underneath, right? Do they...
Do they remain there? Do they start to go away? ⁓ If all the planning is happening, right? It seems like some functionality in DSPs is potentially migrating out of it and being gentrified and going to Asians. What happens there? Do you think that's the future?
Ethan Settel (43:42)
Yeah, again, someone in the thick of it, I'd be, I think I'd be for a long time, I said like, no, I don't feel that way. Like it definitely feels like that's the direction the media buying and transactions are moving. It's like from agent to agent. I still think every agent needs to be informed on the buy side by, I've belabored the point from like performance of campaigns by what buyers want to actually buy. the south side, there's like,
all of the nuances of like, what's actually available? What's the cost between what's spoken for? What's not? If it's an audience component to it, where do I actually expect that audience to show up? Like there's all of these things that need to feed into an agentic knowledge base on both sides of the equation. Right now it still feels like there's very much a space where every player on both sides still has a way to and a value to add to feed that to the agents. I think that gets flattened a little bit, right? Like I don't know if I need.
Brett House (44:34)
Yeah.
Ethan Settel (44:36)
four to five or six different intermediaries to tell me where audiences are available. And then also like where the different publishers exist and how they're surfacing those audience. also like what image like maybe that kind of consolidates a little bit at some point. I wish I could tell where.
Brett House (44:50)
You heard it here, Ethan's calling for the death of the DSP. ⁓ And the
Rio (44:53)
An SSB, right? Yeah.
Brett House (44:56)
SSP.
Rio (44:57)
Yeah. Well, because you think about it, it's come up a lot.
Brett House (44:57)
No, but yeah, I think that's a common theme. That's a common theme. I think flattening is the best way to put it, right? That there's gonna be a simplification of that for basically direct data communication or agentic communication versus having to go through a bunch of intermediaries to get that same information. It doesn't make sense. It's more friction in a system that should have less friction.
Ethan Settel (44:59)
Yeah
Brett House (45:17)
and less cost associated with it because there's less friction.
Ethan Settel (45:21)
Yeah, I still think there is... Yeah, please.
Rio (45:22)
Yeah, well still some questions, right? Because you
have like, because you have like, know, between, I mean, DSPs do serve a lot of purposes, Between identity and frequency caps and, the bid management, some of these things, you know, the bid management with us migrating out. Some of these other things, and maybe you can argue you don't need identity that maybe doesn't play as big of a role or maybe it does, I don't know. So I think, I think it's still an open question, like whether, you know, do they, do they become dumb pipes? And if so, like,
Can you justify 20 % take rate? Probably not. So I don't know if the, I think the jury's still up, but it is interesting. So I'll let you continue, Ethan.
Ethan Settel (45:57)
Yeah, no, mean, again, there's, think they'll focus on where there is a competitive differentiator or then they're bringing somebody to the equation to a point like a bid management system or like a more like a system of record, right? Like where are these agents going to write down a record of like what happened, why it happened? And then how is that going to maybe continue to make the interactions between buy side and sell side smarter? Like I could see very well, like there might be a layer there, whether it's at the DSP level or somewhere else where that's a really crucial piece that continues to play is it's, like
I'm learning from all of these interactions. I'm an intermediary between them and I'm going to streamline ongoing ones and maybe inform both sides of the equation as to like where there's efficiency gains for a seller agent and where there's efficiency gains for buyer agent. ⁓ I'm a good entry point for audiences that maybe don't want to be surfaced to both the buy side of the sell side. there's room in there for lots of intermediaries. It's just, think what specific value they bring is going to tighten.
Brett House (46:47)
Yeah. Yeah. And
on the topic of reducing friction, where are you seeing the most organizational, when you're speaking to clients, because you do speak to clients, friction within the implementation of these agentic systems? mean, is it coming from legal, procurement, IT? ⁓ Where do you see that friction in terms of the actual GTM and commercialization process of these agentic systems?
Ethan Settel (47:10)
Yeah.
It's a good question. Legal and procurement evolved much faster than I typically as a seller give legal and procurement credit for. ⁓ I think the early two years ago when you're trying to, when any organization was trying to buy IAI, it became very scary, very fast. Legal and procurement would, you know, get a whole bunch of document, yeah, whoa, whoa, Mm-hmm.
Brett House (47:19)
Yeah.
Yeah.
Put on the brakes, hold on.
Rio (47:31)
Plus privacy, plus all these other departments
would jump in. InfoSec, 100%, yeah.
Ethan Settel (47:36)
Yeah, I've been impressed by how quickly I think orgs just recognize the pace of change that has to happen in those departments in order to adopt AI quickly. Because if you don't adopt it today, three months from now, the thing you're trying to adopt is almost already irrelevant. So they've moved surprisingly quickly. I think honestly, one of the stickiest points has just been once you're starting to enable these sort of AI tools and putting them in the hands of an analyst or a buyer or whoever, it's making the kind of operational shift into like
Okay, but I've always almost had my analytics team do this. They operate in a silo and then they land it somewhere that then my media planning team picks up and then my buying team picks up. The breaking down of those silos and the almost bleeding in together of the different capabilities that these teams now have. If a planner can do the exact same work that an analyst was doing, because they don't need to know how to write SQL anymore, go build a model in a notebook somewhere, all of sudden,
they can start to ask questions that they otherwise couldn't have asked before. And so like, how does that lead to friction then when, or kind of the person who has a plan, well, I always like to go to this analyst, like you don't have to anymore, but that's what I do. That's what I like to do. I have a rapport. I to get things packed in a certain way. Yeah.
Brett House (48:46)
So there's behavioral change in there. Yeah, there's behavioral change.
It's empowering to the planet because they can do more with less dependency. And to your point about the analytics professionals, the data science professionals, it's empowering to them because they're able to spend less time with data maintenance and unification and more time actually doing innovative next-gen model upgrades, basically.
Ethan Settel (49:10)
Yeah, as an analyst or a status scientist, I'm not going to get called in every time. Well, hold on, the client just said they have a million dollars more to spend. Can you tell me where to put it and pull out of my interesting work? Go over here on the scenario planning. Now it's like, no, now the planner, like, oh, I can do that by myself. I'll go do that kind what if scenario planning with that extra million dollars. And the data science is still off in data science land, of playing around with the new model and then saying, oh, I'm going to incorporate this in the scenario planning tool now because I just figured out a way to bring a
a specific parameter about creative length that I didn't have in my linear model of analysis before, and now all of sudden it's in there.
Rio (49:45)
Well, it's liberating, I bet. I guess looking at liability, I think this is an interesting, potentially thorny question, right? Because these agents are not only running analyses, but they're maybe making recommendations, sometimes even implementing decisions, right? So where does, like, who's accountable? Like, who's liable, right? Let's say it blows through all your budget, it buys those 10 spots, right, on the Sunday Night Football, right? Which is terrible for user experience, but your budget's gone.
Is it the agency? it the model? Is it the brand? Like where does liability sit in this whole thing?
Ethan Settel (50:22)
Yeah, the more automated you get, the thornier that issue becomes. I think we, right now, we have a human in the loop firmly, like in all of those buying decisions, where the agent's going to tell you what it recommends, spell it out, I think, to your other question, but very clearly, like, this is what I've seen, this is what I'm going to change. Like, do you agree? Do you consent? Do you concur? Yes. Do want me to make that change? Like, yes. And it happens. Right? I think it's...
Brett House (50:26)
Yeah.
Ethan Settel (50:48)
Because there's the human in the loop, there's still that element of control and somebody who's like holding the bag at the other end of this for when the change is made. Once that human's out of the loop, then yeah, I'll leave it to our lawyer to of decide who gets liability there.
Brett House (50:59)
Yeah, well, is there someone
managing, you because I'm assuming that these are kind of agentic teams with the agentic workflows with subagents that I'm assuming, just so if I'm wrong, tell me, that are doing different skills. ⁓ You know, but I'm assuming at the agentic teams level, there are, you have basically like overlords, like agentic overlords that oversee and maybe that's not the right way to put it, my dumb way of putting it, but you know what mean? You've got a manager.
Rio (51:18)
overlords.
That's we're all going to have a jet. Brett, we're all going to have agent
Brett House (51:26)
Agents overlord before we know
Rio (51:26)
overlords at some point.
Ethan Settel (51:27)
one
day, sooner or later.
Brett House (51:28)
it, right? But you have an agent that basically oversees and course corrects when things go wrong for those agents teams underneath them. I'm assuming that you've got that layer. Is there a person that's communicating with the layer that's managing the other agents? Right?
Ethan Settel (51:45)
Yeah, there's
only a person who at the starting point is defining like, what is the happy path, right? Like what is that? Agenda overlord need to make sure is happening at each step. And then that is taking place, right? There's some, there's an orchestration, an orchestrator agent who's going to ensure that all those steps are followed one by one. And then still though, building in pause points where there need to be pause points and somebody to step in and say, hold on, like, don't go make that change yet. Like you could theoretically do that like full send like.
Brett House (51:51)
Yeah.
Yeah.
Yep.
Ethan Settel (52:13)
run the analytics through the buy and don't ever ask me to stop and pause. I haven't seen a client yet who's wanted to do that and we would discourage it for now. ⁓ But yes, that is technically feasible.
Brett House (52:22)
Yeah. And there's somebody, there's somebody
probably a data science or data analytics person that's actually kind of managing this system, right? And has this thing operating under them. So there's an accountability that probably goes up to a single person or at least a team, I would assume.
Ethan Settel (52:37)
Yeah,
yeah, typically as a starting point, a data analytics team, whether it's a single data scientist or a number of scientists and analysts who are ensuring that like the first time our agents, Newton's agents are going and running an analytics workflow, they're maybe taking a fine-tooth comb and ensuring that, yep, that all looks correct. The model was trained properly. It was tested properly and then executed and ran properly. But at a certain point, your goal is to bring those people out of it once you feel confident that that consistency is being brought.
Brett House (53:05)
Yeah. Yeah.
Rio (53:05)
the Uber overlord. So, Ethan,
Ethan Settel (53:07)
Ha
Rio (53:07)
we know that
you do have a hard stop in a few minutes. We could probably keep going for a long time, I'm sensing, but maybe we go to Quick Kits real fast. Want to do that, House?
Brett House (53:10)
He's got a client call, Trump's everything. Yeah.
Ethan Settel (53:11)
Hmm. ⁓
it's fun.
Brett House (53:17)
Yeah, sure, sure. So what media task, you because in this process, there's a lot of fear of jobs being lost and all this sort of stuff. I think this conversation has sort of proven the opposite, that it's actually raising the value, the potential value delivery of each function by taking them out of the weeds and out of like the 80 % work that you talked about before. But what media task do you think is most likely to disappear if we had to think about this?
Ethan Settel (53:42)
Yeah, think data entry, it's already kind of on the way out. It can be pretty automated at this point.
Brett House (53:49)
Yeah.
Rio (53:51)
What's the most overrated AI claim in advertising right now? Like the most BS claim? There's a lot of them.
Ethan Settel (53:53)
That
Brett House (53:54)
There's so many to choose from.
⁓
Ethan Settel (53:58)
is a tough one. think like the truly, wholly fully automated cross channel media buy with no human in the loop, you probably sense my hesitation to go there. I think anyone who's telling you they can do that. I question whether you want to trust them or
Brett House (54:11)
Yeah, I think the same thing about the fully automated shopping agents that do all of your shopping without any human agency or control. It's kind of the same concept, but people still have to be involved to set up all the... Yeah, so ⁓ what's the most underrated risk ⁓ in agentic buying?
Ethan Settel (54:28)
⁓ Yeah, think it's, I'm going to keep being biased about analytics. think it's doing it unintelligently, if that's even a word. I think if you're trusting an AI agent to make recommendations, make changes, if they're not coming and showing you the grounding and the foundational analytical knowledge and data science knowledge that they're pulling from, I'd be very questionable and skeptical about the sorts of insights and results and changes they're recommending.
Brett House (54:36)
Yeah.
Yeah, it's the transparency in the black box that's something that you need to be, you know, pay close attention to because that's just not acceptable. It's got to be transparent. so we don't make you late for your client call. don't want to, we don't want Ethan losing a client. This was awesome. And thanks for packing a ton into a short period of time. We normally go for a little bit longer than an hour, but this has been awesome Ethan. We're happy to have you back. But for everybody that made it this far, so thank you one.
Rio (55:06)
Great conversation.
Brett House (55:17)
⁓ for joining us. ⁓ Visit us at www.signalanoise.ai. You can find us on YouTube, Apple Podcasts, and Spotify, ⁓ well as our newsletter. So we will talk to you all soon and thanks a lot.
Rio (55:25)
Subscribe if you haven't.
Thanks again.
Ethan Settel (55:31)
Thank you both.