Amazon Ads offers full-funnel advertising solutions to help businesses of all sizes achieve their marketing goals at scale. Amazon Ads connects advertisers to highly relevant audiences through first-party insights; extensive reach across premium content like Prime Video, Twitch, and third-party publishers; the ability to connect and directly measure campaign tactics across awareness, consideration, and conversion; and generative AI to deliver appropriate creative at each step. Amazon Ads helps advertisers reach an average monthly ad-supported audience of more than 300 million in the U.S. across Amazon's owned and operated properties and third-party publishers.  

Marcus Johnson: Today's episode is brought to you by Amazon Ads, recognizing excellent in advertising innovation through Amazon Ads Partner Awards. Discover how award-winning technologies are helping brands achieve success across solutions, including Amazon Marketing Cloud and streaming TV. Visit advertising.amazon.com/partner-awards to learn more. [00:00:30] Hey gang, it's Tuesday, December 23rd, and welcome to a special EMARKETER Podcast Miniseries, AI-Driven Media Management with Gigi, made possible by Amazon Ads. I'm Marcus, and here is episode two of this two-part series with our senior director of content, Jeremy Goldman, and Adam Epstein, co-founder and CEO of Gigi. For part one, check out last Thursday's December 18th episode. If you have already, then you are more than ready for part two.

Jeremy Goldman: Hey, Adam, thank you so much [00:01:00] again for making the time to continue this conversation. Really appreciate it.

Adam Epstein: Thanks for having me. This is great.

Jeremy Goldman: You know, what's not great is me, I did a bad job last time we spoke. I forgot to actually give you a little bit of a formal recognition of that award that you won, the partner award from Amazon. So maybe at a high level, you can talk a little bit about what does it mean to get that type of industry validation from the strong work that you guys are doing at Gigi.

Adam Epstein: Yeah. So last [00:01:30] month at Amazon's annual advertising event, unBoxed, Gigi won the Global Technology Innovation Partner Award, which was fantastic to see. And I think it speaks more broadly to Amazon's commitment to its partners and its partners, I would probably say on the ad side, it can be in two buckets, agencies and technology providers. And I'm relatively newish to ad tech. I've only been in this industry for eight years. And when I began, [00:02:00] many people warned me of working... And I've been working with Amazon Ads for the entirety of my eight years here. And many people warned me that, "Hey, it's really challenging to build companies on the back of large companies like Amazon." It's saying horror stories of previous large technology media companies that they built companies on the back of, and those companies weren't particularly partner-friendly, when I think it's the exact opposite with Amazon.

I think Amazon is very rightly identified that the best way to achieve the optimal [00:02:30] experiences for their advertisers and customers is through partners. And a lot of that approach I think is learned from AWS. So there's a tremendous partner ecosystem with AWS in which Amazon has historically built API first, meaning they allow their partners to get access to new features and functionality before actual customers in Amazon's UI. And Amazon has had that approach for almost a decade on the ad side. And it's allowed for companies like Gigi and many other technology [00:03:00] companies to deliver a lot of enterprise value and delight to customers in ways that create a really harmonious partnership, for lack of a better word, amongst all of us.

Jeremy Goldman: So one thing that I also think is really interesting is, from our coverage, Amazon has released a lot of new feature functionality at unBoxed and really, throughout the whole entire year. So what does it mean to essentially partner with a group that keeps on leveling up their [00:03:30] learning, I imagine, from their partners and being inspired by their partners, but at the same time, you get to follow their roadmap and get some ideas hopefully from being part of that broader partner network, if you want to say.

Adam Epstein: Totally. I think it's a great question. So first, obviously, Amazon is making big bets in AI. There's a lot of really, really smart people at Amazon. And if you're a large technology company that's not going all in on AI right now, that would definitely be very yellow and red flaggy. [00:04:00] So of course, Amazon is doing this. We can't view their AI initiatives as competitive for us because AI is simply the way in which we interact with software right now. But I think if you look at many of the announcements that Amazon made associated with their Ads Agent in particular, the messaging around the Ads Agents is fundamentally different than the way that we present our product to our customers. So [00:04:30] I saw a variety of quotes from the VP of Product, Kelly MacLean at Amazon that said the Ads Agent is effectively an easy button that's democratizing access to the Amazon DSP for more new advertisers, which is great.

It's a easy way to get access to some really powerful tools that Amazon has built and is using agentic AI as the mechanism for doing so. So in effect, the Ads Agent is raising the floor on advertisers in the Amazon TSP, when we're doing the exact opposite. [00:05:00] We're raising the ceiling on the most sophisticated advertisers. So within agentic AI, we no longer buy software. We hire software, we train and we manage it just like we would a team member. And many of the most sophisticated advertisers that we work with have very prescribed and high conviction ways with which they want to operate the Amazon DSP that needs to be completely tailored to them, the same way that they would train and manage a team member in operating the Amazon DSP. And it's our job, the technology company, to give our customers the tools [00:05:30] to train their AI agent in exactly that manner.

So whereas Amazon Ads and their AI efforts are going to build one to all or one too many, we're very much building one-to-one in that we are not building directly one-to-one, but we're giving our customers the tools to make it a one-to-one unique bespoke enterprise level experience. And so, we actually believe that Amazon's initiatives with agentic AI is a tailwind for Gigi because now, everyone [00:06:00] is recognizing that they have to have an AI strategy and they have to leverage AI tools to achieve the outcomes that they hope to have for their clients. But the challenge is that many of the most sophisticated advertisers need something that's bespoke and unique to them the same way that their team members that they hire onboard would be unique and bespoke to them.

Jeremy Goldman: That makes a lot of sense. I know another thing that you've kind of helped clarify to people a bit, and I definitely want to talk about your background because I find it really interesting. [00:06:30] But you said that there's a lot of confusion around what a quote unquote "vertical AI" agent actually is. And in your view, maybe you can talk to us about what are the core building blocks. We can get a little nerdy if you want again-

Adam Epstein: Sure, sure.

Jeremy Goldman: ... like models, APIs, all the different nuts and bolts that matter most when you're trying to ship real innovation these days.

Adam Epstein: Yeah. So an interesting thing is everyone now has AI. They're an AI first company. And I think from a buyer, [00:07:00] whether you're selling to agencies or brands, the buyer needs to understand where is their true AI or what are the companies that are more marketing versus substance? And I think about it in a variety of different ways. So one, we were talking to a large HoldCo leader and they mentioned, they're asking me for guidance, like how should I think about using all of these agents that are now pitching us agents for this, agents for that? And I [00:07:30] think we've mentioned this, people hire Gigi to be a media manager for the Amazon DSP. And so, other agents might be a report analyst SQL generation agent. Others might be in audience planning like a holistic planning across media channel agent. Generally speaking, these are roles that humans perform at your company.

And if the agent that you're hiring is able to enhance that role that [00:08:00] requires you to hire less of those people in those seats, then that's a true vertical AI agent. So I would use that heuristic for making those sorts of buying decisions. And so, the role that we have is many agencies right now have tons of open roles and it's difficult to hire folks that have X number of years of experience navigating the Amazon DSP or other enterprise DSPs. There's a lot of attrition in this industry. And so, what we say is, don't hire those two roles. Give Gigi a shot first and [00:08:30] then see how it can enhance the existing team that you have. So that's one heuristic. The other is just on pure token usage. So another way that people can assess things is if that company isn't coming to you with a business model that's aligned to token usage, then they're likely not actually offsetting as much tasks to AI.

Enrich their use of LLMs is likely very limited because token usage comes with costs and that's a natural to the business model that [00:09:00] people just need to have on a consistent basis. So those are two heuristics for making buying decisions within vertical AI agents. And then, I think the building blocks of actually building a vertical AI agent are threefold, which is one, obviously, the models are the models. And so, when we began this journey with Gigi a little over a year ago, we just used GPT5, the last version of GPT5 across the board for consistency and reliability [00:09:30] and uniformity and speed to market. But we've quickly learned that relying on a single model and relying on out of the box frontier LLMs is not actually optimal for designing the use cases that we want for our customers. And so, we now use a constellation of models across frontier LLMs like GPT5 or we've begun to fine tune open source models.

And in the future, we actually might begin to build our own models specifically designed for media buying and measurement. The other building block would [00:10:00] be a RAG, Retrieval-Augmented Generation. So RAG is content and data that doesn't live on the open internet that allows you to customize a vertical AI agent beyond the corpus of general intelligence that an LLM will have. So what are examples of that? We at Gigi have created thousands of little bite size RAG documents that we train Gigi to be able to understand how to read certain metrics or execute certain tasks within the Amazon DSP. [00:10:30] In addition to the RAG that we produce to train the Gigi agent with a baseline level of knowledge, we get all of our customers to say, "Hey, what are the training docs collateral process doc standard operating procedures that you would send to a new team member and how could we begin to train your Gigi agent to be unique to you?" And that's all through RAG.

And then the next part is tool calling. So similar to a handyman or woman with a set of tools, if I ask you to put a nail on a wall, you need a hammer to do [00:11:00] that. And so, an LLM needs access to tools. And generally speaking, within the lens of media buying and measurement, the tools that the Gigi agent has to call are API endpoints or metrics within Amazon's ad tech that we can then instruct Gigi to, "Hey, if a customer asks you to do this, you need access to this tool." A lot of people have questions around MCP. And so MCP... Was that the next question?

Jeremy Goldman: Well, it's funny because I was going to say, I [00:11:30] think RAGs and MCP servers are going to be... Everybody will know what they are within this industry a year from now, right? Because it's like that these are cores to unlocking additional growth.

Adam Epstein: Totally. Anyone working within AI needs to understand nuances of RAG and RAG orchestration. And then a good way to just think of MCP servers is rather than us integrating with every single API endpoint within Amazon's ad tech as a tool call, [00:12:00] we can use... And Amazon announced early version of their MCP server. And so we can say, "Hey, Amazon Ads Agent, these are the problem that we need to solve for our customers. Can you demonstrate which tools we should call to achieve the outcome that we're hoping to achieve for our customer?" And then the MCP server will expose the appropriate tools for that outcome rather than us having to deterministically call a certain API endpoint for that tool. So those are generally [00:12:30] speaking the building blocks.

The other big piece with it was associated with RAG is, customer advertising data is obviously RAG as well. That doesn't live on the internet and allows for someone to customize their Gigi agent. So across customer best practices and customer data, that's how a Gigi agent gets customized for each individual enterprise that we work with. And a really important design decision that we've made that everyone asks us is like, "Are you taking all this data and best practices and laddering it up to a master model?" And the answer is emphatically no. While we understand that that could have probably [00:13:00] been beneficial for the overall efficacy of Gigi's general intelligence, we think it's really important from both a functional and legal perspective to ensure that all of our customers' RAG, whether it be data or content, is unique to them and bespoke to them.

Jeremy Goldman: Yeah. And I think that that's obviously a differentiator, right? You have organizations that are going to want to fine tune and they're going to have different business rules and a lot of different things that they're obviously going to demand it, right? So it becomes a little bit of a selling point to keep those things separate. [00:13:30] But when you were talking before, I want to go back to something about the complexity of everything that you went through. And I was going to joke and say, "You sure you've only been in ad tech for eight years?" Because you mentioned your own journey when we were talking offline from lawyer to somebody who's pretty fluent in these tools. I'll give you like a 98 out of 100. I mean, what advice would you give leaders or teams who don't come from technical backgrounds, but need to build or use AI [00:14:00] systems, which let's face it, is most people these days.

Adam Epstein: I was chatting with a friend, one of our investors who mentioned how I'm enjoying being a founder, and actually this is my third time being a founder. And there are trade-offs being a founder. We're at an early stage. We're certainly not killing it by any means, but we're doing well at a very early stage. And I said, "You know what? I'm enjoying it. I'm really passionate about solving this problem. I love our team." But more importantly, I feel like it's such [00:14:30] a gift that we've been given, our respective Gigi team to be able to build at the forefront of all of this because we're just learning so much.

Whenever we hire a new team member, we like to say, "You're going to get a black belt in AI by becoming a member of the Gigi team." And I think one of the things that I've really enjoyed and I think is one of the things that's allowed all of us on the Gigi team to be successful is that we just have this insatiable curiosity to [00:15:00] learn. And we began this journey, as I said, we knew we needed to go on on the AI. We didn't know what that meant. And it's really been an incredible amount of work across our 14-person team to go from AI novices using ChatGPT to getting a 98 score from you, Jeremy, in our understanding of technical AI terms. And so, it's just a lot of hard work and curiosity.

Jeremy Goldman: I want to give you room to grow. I don't want you to get a swelled head.

Adam Epstein: [00:15:30] I appreciate that.

Jeremy Goldman: But no, in all seriousness though, given where you are in terms of sitting within this ecosystem and growing and getting the recognition through the partner awards, I mean, I wanted to get your perspective on this. Looking into 2026, what's next A, for Gigi, but also for the broader AI powered media world, let's just say, from Amazon DSP advancements to obviously, data collaboration continues [00:16:00] to be something that's entering the lexicon, the average marketing practitioner is now very well familiar with these terms. So what can you tell us? Do you have any bold predictions? Maybe we can start with Gigi and then kind of lens out.

Adam Epstein: Yeah. Really simply on the Gigi side, we want to expand aggressively to other forms of media buying and measurement. We very extensively built our product to be, while the RAG orchestration [00:16:30] and tool calling that Gigi has access to right now is all associated with the Amazon DSP, we certainly want to expand to many more mechanisms of buying media on other enterprise DSPs as well as onsite retail media. That would be the natural extension of our product, and we're working very hard to achieve that. I'm going to get back to the AI piece and on a bold prediction and not necessarily a bold prediction, but what I'm looking to work towards is, we began this conversation of understanding [00:17:00] the tasks that someone in a media management role at an agency has and how we automate that. And so, if we talk to someone, the number one task that Gigi has not been able to solve is deck creation, automatic deck creation.

The amount of time that people spend building decks for their clients is probably the biggest time suck that people have right now. And we sought to solve that problem in 2025. We worked really hard. One of our top engineers worked for a couple months to try [00:17:30] and solve that. So we have all of this advertiser context, we have all this data, we have all of the inputs that they've given into Gigi, and how can we automatically create a deck based on the template that they've given us? And you know what? The technology wasn't there. We tried really hard. This speaks to the evolution of the models and what's unlocked and the type of work that you can begin to evolve in the work that you're assigning to agentic AI.

But I've spent the past 10 days or so on a Nano Banana rabbit hole, [00:18:00] which is Google's AI image generation model. And I think that we're getting really close to having automatic deck creation and having phenomenally produced decks at human level deck creation that could offload a ton of work. And so, we're really excited to hopefully provide that to our customers. It's a hard problem to solve. And again, we tried in 2025. And for 2026, [00:18:30] I know that this is very in the weeds, but deck creation speaks more broadly to something that every single person in many companies offers. And I think that once the technology of deck creation, image creation becomes commoditized, and it gets back to all of these inputs that a vertical AI agent can uniquely do and have, to build the best possible deck for their verticals.

Jeremy Goldman: It's been a time suck for me in the past. So I co-sign, I think that's a terrific [inaudible 00:18:59]. I think it's [00:19:00] a lesson to everybody listening to this, right? Which is, you need the vision for the thing that you want to create and AI might not actually be there yet, but it's moving so incredibly fast that you need that vision and then the AI will probably get there at some point in the near future. But listen, this was a great conversation. I can't take up any more of your time because I want you to get started working on that deck creation for all of us at home. But Adam, this was really fantastic. So thank you so much again [00:19:30] for making the time to speak with us.

Adam Epstein: Thank you. Appreciate it.

Marcus Johnson: That's it for today's episode. Thank you so much to Jeremy and Adam for the conversation. Thanks, of course, to the production crew. And thank you so much to everyone listening into the special EMARKETER Podcast Miniseries, AI-Driven Media Management with Gigi made possible by Amazon Ads, and of course, happiest of holidays to everyone.