How AI Is Automating Ad Agencies

**AI's Impact on Ad Agencies: A Deep Dive with Misha Leybovich of AdSmith.AI**

(0:05) **Introduction to AI Insiders and Misha Leybovich**
Welcome to another edition of AI Insiders by AI Marketers Guild, part of the Market Media family. Today we have a guest I have known a while. I was introduced to him in 2014 back in my agency days. Misha Leybovich was introduced to me by a mutual friend, Mick Darling, and I have been following his journey for a while.

(0:35) **Current State of Ad Automation**
He was talking to me about what is going on on the ad automation front. It is a topic we get to here and there, but I do not think often enough. I was curious to hear some of the latest of what can be done, what cannot be, what should and should not be done, how screwed agencies are, all these kinds of questions. Misha, welcome.

(1:02) **Misha's Welcome and Background**
Thank you. Glad to be here.
Great to have you. I would love to hear in your own words what you are working on these days and dive in.
Absolutely. Everybody, I am glad to be here. My name is Misha Leybovich. I am the CEO of AdSmith.AI. I am going to be talking about my product a bit. This is not a pitch. This is about what is going on in the world of applying AI to advertising and what we are seeing. Given this group is very intensely curious about AI and marketing, it is about what we are seeing, what we are building, and what we think it means for the future of this. To give you a little bit of background on myself, I have been an entrepreneur since 2012, with a four-year stint in corporate. Before starting AdSmith, I was at Google for three years. There, I was in the Marketing Works team. I was building internal tools for Google's own marketers to market Google products. I saw and was supporting the engineering team for campaigns of hundreds of millions of dollars. I learned some stuff there. Before that, I was a marketer for my own startups. I had a couple of startups before that. It was mostly me and a bunch of developers, and as the CEO, I was always responsible for everything else, including marketing, including advertising. I was able to develop some strategies that allowed us to punch way above our weight in terms of actual success in the market for pretty small budgets. The summary of that was efficiently aggregating the longtail, advertising everywhere in every language and every country where it was relevant. I was able to get some of my apps to the top of the charts in 134 countries, up there with Instagram, TikTok, Snapchat. Little Flippy was hanging up there. Based on the way that we were doing things and combining my experience doing my own advertising as a very small business with seeing how it worked at the very large level with Google, here was my main takeaway, and this is going to sound pretty reductive, but I think it is true: all advertising is guessing.

(3:27) **Advertising is Guessing**
Everyone is guessing. Nobody knows what is going to work. The ones that are great at it, the ones that we pay extra money to and get hired more, the great advertisers get it right, and by "right," I mean the campaign performing according to business expectations, get it right about 50% of the time. That would be a great batting average. Most mortal humans do much worse than that, and then you think advertising does not work. Given that, here was my insight about a year and a half ago. I saw that the capabilities of the models were going up and up. A year and a half ago, everyone said everything has six fingers. Give it a second, bro. It is going to get better. Images are pretty much there. Video is just about there as well, at least good enough for advertising, which is disposable art. Most of these things are not going to win big brand awards. It is not its job. Its job is to earn the conversion. The capabilities were going up and up, and the time and cost to produce relevant, quality, credible "shots on goal"—the time and cost would go into zero. It is almost there. An amazing image will cost you 10 cents. An amazing video cost you a dollar or two. This is fundamentally way different than it has ever been before.

(4:59) **AI's Advantage: Outperforming Human Guessing**
If you combine the insight that all advertising is guessing—we are guessing too—with the understanding that we can now produce orders of magnitude more guesses, to me, it only made sense that if I could build an experiment and learning machine which also is guessing but about a thousand times as fast, we could outperform any humans doing the guessing. Let me explain the way that we do that. In any final asset that you see as a part of any ad unit, we are talking text, images, videos, keywords, any actual assets. Ultimately, some prompt went into making that. But then what went into that prompt? There are a bunch of different decisions that go into that, what is the persona that we are targeting? What is our messaging? What style are we going for? What kind of story are we trying to tell? How creative is it? What kind of assets go in there? There are all of these different decisions that a human making assets makes implicitly in their head. It is a complex thing, and I love the creative process, but ultimately it results in a series of decisions that results in the human doing a thing and producing some assets. This stuff looks great. What I tried to do is say, let me automate all of those decisions in a step-wise fashion, going down from what is the organization, what is the offering, what is the audience, what is the persona, what story are we trying to tell, the creative brief, and automate all of that.

(6:43) **Automating Creative and Campaign Experiments**
The advantage of doing it this way is not only do I have this fire hose of ad content that I can produce—my system can produce credible campaign experiments for a couple of bucks in a couple of minutes versus weeks and thousands of dollars if we are talking about an agency doing it. Not only do I get way more shots on goal, and the advantage there, let me skip to the punch line here: every single one of our campaigns so far outperforms every single one. I am not saying this to brag about my company. I am saying this about the approach that we are using, and we are not the only smart guys out there that are going to figure this out. Right now, we outperform every single time versus a credible head-to-head versus a human doing or human agency or in-house or whatever. Why do we outperform every time? Why do we do it in industries that we do not know anything about? We do not know about it, but the AI knows about all of these things. How do we do that? We are not marketers by trade. We are engineers trying to optimize every performance lever available in these advertising platforms.

(8:08) **Flooding the Zone with Credible Content**
For example, if you are doing a Google Performance Max campaign, and sub to that is an asset group. A Performance Max campaign can contain up to 100 asset groups. An asset group can contain up to 20 images, 25 captions, 50 keywords, 15 videos, six sitelinks, all of that. When we flood the zone with credible shots on goal for every single slot, it means that statistically something that we do is going to work. Even our small company now, I think we have to be in the top 1% of advertisers just using all of these things available, using all of these slots, because our logic is that these advertising systems have gotten to the point where it is about what they offer effectively as algorithmic targeting. It is not about selecting dropdowns anymore. I want a new parent who likes basketball and has a white-collar job. The ad platforms are removing more and more of those selectable dropdowns, some for privacy reasons. They are narrowing that down. They explicitly call those suggestions now. They are not even targeting anymore. They are just kind of like, "Hey, Google, Meta, go look in that direction." What it really is is that humans, we are all more complicated than a series of dropdown boxes. What the platforms want, Meta says this explicitly: "We want creative diversity." They say, "Give us a bunch of content, and we will figure out what to show to whom, and this kind of experiment is going to work with this segment, and this is going to work with that segment, and let us figure it out."

(10:06) **Algorithm-Driven Decisions and Statistical Learning**
We believe that the algorithm is always going to make better choices. I think it is reflected in the data. Part of the reason that we are outperforming is just the arbitrage currently that we are taking orders of magnitude more shots on goal with a system that can produce a lot of credible volume of content that would all work. I tell my customers, I do not know what is going to work. I could not tell you in advance, but I know that something is going to work. We are going to learn from that and do better. Let us talk about the learning part because this is where it gets special. I mentioned we automate 200 inputs by creative and business decisions that go into producing any given asset. Those are not locked in a squishy human brain, that some creative made those decisions. Those are fields in a database. That means that when I get performance data, and I see of all these experiments that I am launching, what is actually working? Let us say I put out 10 experiments, six of them do not work. Who cares? It cost a couple of bucks, took a couple of minutes. It does not matter because no one could have predicted that these four work and these six did not. Anyone that tells you that they can predict that is not being honest. The stats do not bear that out. From those four out of 10 that did work, what can we learn? Those 200 inputs now can be matched with the outputs of what actually yielded conversions.

(11:42) **Reducing Cost Per Conversion and Predictive Guessing**
To give you a sense of this, our average—we mostly optimize for cost per conversion. We do not care about impressions, click-through rate, or clicks. We care about conversions because that is all that matters at the end of the day. A conversion is going to be different for every business. Sometimes it is a lead, sometimes it is a sale, sometimes it is a free trial, whatever it is. We typically bring down the cost of conversion by 30% to 50%. We connect to Google and Meta. That is where we connect right now, and we are adding more channels as we go. We bring down that cost of conversion by 30% to 50% already. What is interesting is now knowing which experiments worked there, and I mentioned there are 200 different decisions that went into that. Now I have a 200-vector space that I can train a model to be for this customer, for this audience, for this persona, for this offering. What are the next set of these 200 inputs that we should guess? Yes, it is guessing, but it is increasingly informed guessing. If I took that same four out of 10 experiments that worked and I showed it—we do not do landing pages yet, but I want to get there because that closes the entire loop on the thing. I want the landing page to match the ad. But we are not there yet. If I showed these to a human and I said why did these four out of 10 work? The human, because we all want to look smart and make our best guess, they are going to look at it and say, "Oh, these ones worked because it had a yellow background, and there was a woman holding a friendly dog, and it had the text in the upper left. That is why these worked." The reality is this is a statistical space with 200 different vectors in it that yielded these outputs, and there is a statistical answer, and there is signal among a ton of noise that with enough data we can make better and better guesses as to what works. Our goal is for the next batch of 10, five of them work, then six of them work, then seven of them work.

(14:26) **Automating Campaigns at Scale**
Ultimately, what we are trying to do—the scale of this is going to sound crazy to anyone operating an ad agency with a process that runs at human speed. When I talk about a Google Performance Max campaign that can contain up to 100 asset groups, that represents one experiment for us that we are over time for every customer filling all of those slots, every single one, and killing losers, leaving the winners, and filling it all in until it is all killer no filler. Everything is working and driven statistically until an experiment does not work anymore, and then we replace it with a new one. The goal is to have dozens of experiments launching for every customer every day and to automate this entire thing. That is the gist of what we do, and we fashion ourselves as an AI ad agency. I am sure you have all seen the phrase among investors, "It used to be software as a service. Now it is service as software," where the budgets for services dwarf budgets for software. As AI increasingly enables us to provide these services, more and more of those dollars are going to move over to those services being offered by software. An ad agency is inherently a services business. They may have tools on the backend to enable them to do those things, but ultimately it is services.

(16:25) **Scale of the Ad Agency Market and Customer Complaints**
To give you a sense of the scale here, there are about 400,000 ad agencies in the world, and the spend on these is about $400 billion per year. There is a lot of budget here to go after. The spend on actual ads itself is about a trillion, but the spend on the services to make those ads happen is about 400 billion. I am sure some of you hear agencies, and I am sure they are wonderful, but it is also pretty cliché to hear customers complaining, "I tried this agency, they made all these promises, they tried these things, but ultimately I did not get results." A lot of money is spent, and the economics of human labor to do that not only means that you can try fewer experiments, but you are also pricing out a lot of businesses from getting professional marketing services in the first place. Of all ad spend, about 40% comes from startups and small to medium businesses. These are businesses for whom the economics of hiring an agency and the cost of human labor does not work. It is too large a percentage of their budget. It does not work for the agency because the spend is too small, and they do not have enough to spend on the labor. We see an enormous opportunity here.

(18:10) **Agency Hostility and AI Disruption**
To be honest, when I started this, my thought was, we might serve some customers directly, but also maybe we will work with agencies, and agencies will be B2B to B. Agencies will use us as a tool for their customers. I have to say that the feedback and response from most agencies has been hostile. I understand that we are explicitly going after their margins, and we are trying to offer a better experience for lower cost and higher performance. We all know there is going to be a lot of disruption with AI, and we are going to sort this out. I am long-term optimistic in humanity and our glorious future, but it is going to get messy in the meantime. I am not here to talk about, "Oh, this is going to supercharge agencies." Anyone that wants to work with us, we will be happy. Anyone who wants to use my tool, I do not care if they are an agency or a business or a customer themselves. I think a lot of the rhetoric right now, because people are coping with the massive disruption that is happening, is, "It is going to be fine. People are going to use these tools, and they are going to be better." I am not here to share that message. I am here to share the message that I think—maybe I am naive, maybe I am selling my own book here, take my motivations with a grain of salt.

(20:01) **Validating AI's Disruptive Potential**
I want to share a quick thing to support your point, which coincidentally I was reading in their newsletter this morning. They sent this an hour ago, and I am like, it already stood out, but now, based on what you are saying, it is clearer to me because it talks about Horizon Media, the largest of independents, building a single command center for programmatic buying. The new Horizon OS platform sits above the adtech stack. I will send a link to the newsletter, but I want to share their opinion on this, and this is eerily in sync. Michelle, I love your reaction to this. The U of Digital opinion: "This is a step in the right direction for agencies, but it disrupts the traditional agency model, which is charging for services to do this work, not tech. They will have to evolve their entire positioning and pricing in order for this to work."
Yes, I think that is still, in my opinion, a pretty soft-handed approach, because at the end of the day, when a system can provide every single one of those services, what exactly is the agency doing? What value do they still add in the chain here?

(21:42) **AdSmith's Goal: World-Class Marketing Services through AI**
My goal with AdSmith is to provide world-class marketing services around advertising, but we may expand to other types of marketing after that. I went to advertising first because that is where the money is. World-class service to companies of any size around the world, done instantly, and at the highest level of performance. I know these are all big claims. Let me talk about the steps of the services that an agency provides and talk about how we are approaching automating every one of these things, and then I will get to the questions in the chat.

(22:31) **Why Agencies Exist: Complexity and Abstraction**
Why do agencies exist? Advertising is complicated, particularly now. I have been using these ad platforms for 15 years, and I still find it confusing. There are all kinds of knobs and levers. There are all kinds of different choices you can make. There is the amount of content that you need. It is very difficult to do every step of it. Then we even have things broken down into creative agencies, strategy agencies, brand agencies, performance agencies, media buyers, all of these things. It makes sense why businesses that are good at making tennis rackets or whatever they do say, "We do not have these internal capabilities, let us outsource this stuff." That makes total sense. Ultimately, an agency is a compendium of jobs, jobs to be done. It does those jobs in a package, and it abstracts the complexity from the end customer. Until this point, you absolutely needed humans to do that. There are so many subjective things and judgment things, and all of that made total sense.

(23:51) **Automating Agency Functions with Coordinated Agents**
We are getting to this point now where, particularly with conversational agents as the interface between the ultimate advertiser and the ad platforms. Conversational interface, the ability to have a series of agents. I think the word "agent" is highly overused. I am long-term bullish, but right now I think there is a lot of hype. I do not believe in some single agent that can do everything and has all the tools at their disposal. I think there is too much opportunity for it to go off the rails currently. Ask me again in a year, but currently, I do believe in a team of coordinated agents where this is my keyword agent, this is my media buyer agent, this is my image review agent, and then a controlling agent to know how to dole out the work. I do believe that that works. So you combine the conversational interface with the ability for the sub-agents to use a series of tools that are prescribed on how to use it and delineate every edge case possible. The data schema that holds these 200 decisions that I talked about. The ability to hold those somewhere in a very structured way. I say that our schema is one of our greatest assets because the shape of your data and the opinionated way that you think advertising works is one of the things that makes our campaigns outperform every time. The ability to then have these platforms—I want to credit the platforms like Google, Meta, all of that—what we are doing, even if we could produce all of this, all these experiments, and we sent it to Google and Meta, we rely on them to do the not just A/B but A through triple Z testing for us all the time, and to do the algorithmic targeting. That is a critical part of what we do that is in place. Then the ability to have a feedback loop to make better decisions. All of those things, all of those services, can now be done by agents.

(25:59) **Automating Intake and Strategy**
Let us walk through it: Intake. You get connected to an agency somehow. You are a customer. You are selling your tennis racket. I need to do some advertising. I need to grow my business. I do not know how to do it. Let me talk to somebody. What happens? You get on a call. You have a conversation. They ask you about your business. They ask about your offering. They ask you about your target customer, the advantages of your product, your creative preferences, all of those things. Guess what? Now these conversational, voice-to-voice, even avatar-to-avatar, it can look like me talking to you right now. It can look that good. Those are all available now. So, the intake process, that is the easy part. Getting all the information can be automated. My goal is to get someone landing on AdSmith, entering, and getting them not only intake but all the way to a published campaign within five minutes. That quick. We are still a little ways away from that. We are still around a couple of days, but we are reducing bottlenecks one by one. Next is what is the strategy? What is the offer? What kind of message do we put out there? Again, guess what? Our conversational agent can talk to you about your business needs, your preferences, store all those in a list of ideas of things to try, and execute those. Another service automated.

(27:26) **Automating Targeting and Creative Personalization**
Targeting. We are talking about audiences in terms of locations and languages, the combination, and then personas because there are all kinds of different flavors of people. Before, let us say you are selling life insurance. You could do a generic life insurance ad. But then, let us say you did a special version, one for people who like basketball, and one for people who like sushi, and one for people who like action movies, whatever. An ad has two jobs: it has 0.1 seconds to stop the scroll. We see 4,000 ads a day crossing our retinas, and we barely register most of them. So, you have to stop the scroll and then earn the click. The landing page, to your question, Conor earlier, the landing page has to seal the deal. So, we want to get into that later. But it has 0.1 seconds. In that 0.1 seconds, if I am trying to sell life insurance, is it some non-zero percentage more likely that I am going to catch someone's attention if the imagery contains a basketball or some sushi or something about an action movie, perhaps? Because as humans, we are a complex array of interests and likes and things that catch our attention, and the basketball is going to be a little bit more likely to stop the scroll than something else. Before, it would have been completely impractical to launch campaign experiments for that niche of audiences, but now we can.

(29:05) **Targeting Approach: Conversions Over Specific Demographics**
Can I ask one question about all this? For the work you have been doing here and elsewhere, is it better if someone is coming to you and saying, "These are our targets, these are what we know. We do not want to target 20-somethings, we do not want to target boomers. We are a stay-at-home dad brand versus a Fortune 500 CEO, working mom brand"? Or are you better when someone says, "As long as someone buys the bleeping product, we do not care who you are. We do not care where you run them. They could be on an X-rated site at 3:00 a.m., and if they have a craving for my mac and cheese, serve it to them"?
Absolutely. Honestly, every advertiser is going to have to evolve into the latter, because increasingly, even at the biggest brands now, we will talk about brand safety in a moment. At whatever size company, the business owner, the CFO, the CMO is going to increasingly have to justify why they are doing a thing that is less efficient and less profitable on their ad budget than other things they could do. So our approach to it is as follows.

(30:33) **AI Reliability and Creative Constraints**
People think AI is unreliable. It is going to produce crazy things. This is not the case with the current models. We also have humans reviewing everything, but it is never going to accidentally insert something crazy into your ads. That said, we ask every customer, "What creative constraints should we set?" The AI is very good. The creative space from which you can produce a given ad is infinite. There are infinite ways to arrange a series of pixels within some rectangle to produce something that tries to convince someone to click on it. AI is very good at exploring all the different combinations, a high-dimensional space that we cannot even imagine in our heads. We can constrain that. If the brand says, "We want it to be only these colors, only these messages, only these styles," no worries. If you give that input to the AI, it absolutely can do the job of staying within that.

(31:44) **AI's Role in Regulated Industries and Performance Focus**
I am excited about regulated industries because there are a lot of rules about how you can advertise. No worries. Put all of those rules as an input to every generative call when the AI is doing things. Then you can add it as a check afterward to make sure that we reject anything that for some reason strays outside of that. You are going to catch pretty much anything that does not meet those, and you can stay within whatever constraints, be they regulatory, be they brand-based, whatever it is. Stay within these boundaries but be as creative as possible within those boundaries. That is what gets interesting. David, to your question, the brands that tend to work with us now are the ones that are like, "I just want conversions." I may have my opinion that this is a stay-at-home dad brand. That is cool. We can certainly create ads that meet that persona. No problem. I would not say that it is all randomness and chaos. If your brand is a stay-at-home dad brand, advertising to teenage girls is probably not going to work. But my question is always, "How do you know?"

(33:09) **The "How Do You Know?" Question and Purchase Influence**
By the way, my 12-year-old probably has 80% of the purchase influence, maybe 90%, in my household. Advertising anything to her is the best way for her to say, "Dad, we need this."
Absolutely. If your brand is selling adult diapers, you are probably not selling to teenagers much.
Sure, absolutely. There are some bounds of reason here. We are not going to be—it is very unlikely our system would create ads showing teenagers independence. But within that range, sometimes it is the person that needs it. Sometimes it is their adult children looking for it for them. To take that example, we only judge by what gives us the best cost per conversion. That is it. We will make our creative to fit within whatever bound you want as the customer, but ultimately it is going to be what works, and that is what it is based on.

(34:19) **Performance Advertising: No Room for Humans?**
That is why I think, and this is part of my scary message for agencies: believe me or not, I think that—I do not know how long it is going to take, five years, ten years—but performance advertising, I am not even talking about brand advertising, that is a different thing, because I am talking about anything where you get a ton of shots on goal and a clear signal, "Did this work or not?" through a conversion. For performance advertising specifically, I do not see room for humans in this. I do not. To provide the inputs, sure. Each brand can decide how they want to show up in the world and give that guidance, but then to actually execute and provide the services, I do not see where humans add value in this chain or how they can outperform. Ultimately, what matters is how they can outperform what the AI can do. I think there is just going to be increasingly clear based on the numbers.

(35:25) **Automating the Creative Layer**
We talked about the service layer. We talked about strategy, targeting. Now, let us get to creative: keywords, copy. People say, "It needs the human touch. It needs the human experience. It needs to have the lived experience of the empathy of being a human." Yes, that is true. But the AI has all this by virtue of all the things that we as humans have already created. The AI has internalized what the human experience is by the content that we can produce and ultimately can produce stuff that is increasingly going to be—there are a million different ways that we can sell this product. There are a lot of different messages we could use. Let us try all of them and see what resonates. Put it this way: in our system where we show every time, "Here are the keywords we are going to use. Here are the captions, all of that on the text," it is extremely rare that a customer says, "No, I do not like that." They could exit out, they could change it, but it just does not make sense because what we want to do is have this cloud of constant things that we are trying and develop a statistical understanding of what words resonate more.

(36:39) **Copy for Robots and AI-Driven Production**
So, copy that is just for the robots, a concept. When we talk about images or videos, yes, there is a prompt that needs to go into that. To produce a single ad, you do not need me for that. Go to Gemini, go to ChatGPT, go to Midjourney, go to Ideogram, whatever you like. Enter a prompt, you will get an amazing-looking ad. You do not need me for that. If you want to run thousands and thousands of these, if you buy the premise that it is a statistical understanding that matters, then you do need some sort of a statistical approach. Coming up with those concepts for every one of those prompts, I am a creative guy. I can come up with maybe 15 ways to sell a sandwich, but I cannot come up with 300 ways to sell a sandwich. Having the AI come up with the concepts and varying these concepts matters because humans get stuck around local maxima, around our own creative preferences and beliefs on what works. The AI does not care. The AI explores anything that is credible. A creative strategist is not needed. I am sorry to say. The production of it—images, videos, all that—I have zero creatives on staff. Zero copywriters, graphic designers, video producers, none of that. It is not needed, especially if you are going to take this approach of a statistical, thousands-of-times-more experiment, which I think is going to prove out to be the only approach that works, not only for my company but other companies will do this too. I think this is going to be the approach that works. Another service thing: creative agencies. We are talking about brand design, different story there, you get one shot on goal, but if you are doing performance advertising, you have thousands of shots on goal, different story.

(38:29) **Automating Conversion Data and Campaign Settings**
Let us talk about the replacement. One thing that we have experienced is that 0% of our customers we work with closely have had their wiring—the actual way that we take conversion data or conversion events, define them, send the information back to Meta or Google, add value rules, add enhanced conversions, all of these optimization levers—optimized. Everyone has something messed up. This is very difficult to do, and my team is good at it because we do it all day, all the time, and this is our specialty. With all of these things, we are building these playbooks for computer-use agents to then go and for any advertiser of any size, be like, "Great, give us access to your system. We are going to take over your browser. We are going to make sure that everything is set up perfectly," because that is free performance sitting right there on the table, just for free to pick up point here, a point there. Campaign settings, very confusing. Which one should we use? We have particular opinions about that that we see work best. So we automate using nodes. Optimization: you have to appeal things on the platform. We can automate that to do the various optimizations that they suggest. All these can be run through APIs.

(40:00) **AI for Media Buying and Budget Rebalancing**
Rebalancing. Let us take media buyers. An advertiser does not care about, "I want to spend $1,000 today on Google and $2,000 on Meta and $3,000 on Reddit." They do not care. What they want is to spend X amount of money. That is their budget, and to get the best results from that. A media buyer is going to look at that, maybe rebalance once a week, once a month. They are going to make their best guesses. We have a prototype agent already that can do that automatically, and all you do—it is amazing what these tools could do now—you take a CSV of all of the performance data for the past 30 days or 60 days, or however you want to do it. Give it to the agent, say, "Hey, for all of these different objects at the campaign level, the experiment level, across channels, how should we rebalance budget, and which experiment should we kill?" It looks at that, and the recommendations it gives, I am telling you, are right every time. It is just numbers. This is another part of the service layer that AI can do.

(41:09) **The Hard Part: Handling Messy Human Inputs**
Every optimization possible. My goal is to make it so that anything with a human in any part of that just cannot have the throughput and the speed to compete. Now let me tell you about the part that is hard: the part that is going to take us the rest of this year and probably most of next year too. The service layer, even though everything that I am saying about conversational agents handling that, extracting learnings, and feeding that to agents to do the right thing—humans and customers are messy. We are all messy, and we say things in different orders, and we contradict ourselves, and then we say we want one thing, but then shown it again, we change our minds. It is messy, and there is no getting around that. Sometimes the smaller the advertising, the less sophisticated, the more opinions they have, and the more that changes because it is emotionally driven, whereas our thing is entirely statistical. How do we handle that? That is a thing that we think about a lot.

(42:25) **Designing for Human-Agent Interaction and Bottleneck Removal**
The thing that is going to take us the longest is to make it so that no matter how we design this—on the left sidebar, you have this agent that is always there, and you can talk to anytime to always figure out—they start a conversation. It might be from we do not understand what context they are starting from. They might switch topics midstream. They might say something that contradicts something before. Ultimately, all of those need to turn into fields in databases that are learnings that we can feed into generative AI calls for all the things that we are going to do and how to sort through a bunch of messy conversation and turn that into the right structure and then implement that correctly and have the thing they said now supersede the contradictory thing they said before. These are not trivial things. That is going to take us the longest. Our approach is to automate the backend, the production layer. We are working on right now. On top of our service, which is 10% of the spend, our service fee, we offer a white glove service for an extra thousand bucks a month. We offer that with humans right now. We are using that as the opportunity to understand every edge case, and we are nowhere near the bottom of that list yet. There are so many edge cases we do not understand yet. But as we do that, and we are like, we have seen all these things before, we are gradually removing bottlenecks.

(43:52) **Example: Image Review Agent**
I will give you an example. It might take my account manager for the white glove service 15-20 minutes to produce an experiment, and the bulk of it comes from looking at the images and saying yes or no to different images because even now, with AI models being amazing, you might reject half or even three to one. They cost 10 cents, so who cares? But still, that probably takes three-quarters of the time to do that. Guess what? We can produce an image review agent that passes every image once it is produced to an agent that says, "Hey, here are the 30 things that can generally be wrong with an image. The logo has a problem. The text is weird. The human has three arms, whatever." It looks at it and then says yes or no. I am pretty sure we can tune this to be not too restrictive, not too permissive, just the right amount to be efficient with this. With that, I have reduced the time for her to produce an experiment from 20 minutes to 5 minutes. That is a series of things. All of these things we cut time until an agent handles everything. That is our ultimate approach. I have been monologuing for a while. Maybe I will get to some questions here if anyone has live or I can answer the questions in the chat.

(45:18) **Q&A: Adam's Question on Pricing Model**
Who has some? Okay. Got a couple of hands coming up. Adam and then Emily.
Misha, first of all, thank you for all this really cool stuff. As a fellow ad tech nerd, I love the way you are thinking about all this. I wanted to push on that last mention of the 10% of media spend model because it has always been the case that when you go from 1 million to 20 million in spend, it is not 20x the effort. Now that astoodic thing is just going to make you more vulnerable to SaaS platforms that come along and say, "No, this is just a license fee." Maybe it is tiered a bit on spend levels, but I wonder if you could unpack that a bit.
Sure, absolutely. As spend goes up, we certainly have in mind to lower the percentage as the spend goes up. Let us talk about the reason why we are doing that model in the first place.

(46:29) **Pricing Model: Performance and Value-Based**
It was an easy one to slot into because that is the industry standard. More to the point, with AI systems, you want to pay based on results and performance of outcomes. My ideal is to charge some percentage of the profits that a customer makes based on our ads. That is ideal. Then it is a no-brainer.
If you are running the full loop and the measurement piece for them, you could.
Yes, but here is the problem: I have to get pretty embedded in their ERP systems or whatever it is to know what the actual profit or sales are. We are not at that level of sophistication yet. I absolutely want to get there. But for now, the spend is a proxy for that, because if it is working, they are going to keep spending, and they are happy to pay my 10%. If it is not working, they are going to stop, and I am not going to make any money anymore.
Also, if I were your client, I would drive so many other costs through and crush the synthetic profit.
Yes. Ultimately, we want to get as close to the value as possible. This is our answer for now, and that will evolve over time.

(48:15) **Q&A: Emily's Question on AI Pushback**
Thank you. This is fascinating. I am coming from a very different angle. I work in nonprofit marketing and communication. I am trying to figure out how to potentially leapfrog over these absurd advertising and agency costs and just get to the point of trying to get our messages out there. I am more curious about any observations you have on pushback to AI, culture-wide but also from your clients. They obviously are not opposed to AI if they are coming to you, but also the people that receive the ads. The ads and the quality of the content are getting better and better, and people are not differentiating between human-made content versus AI-made content. I am curious if that is a factor at all in terms of how you are thinking about eventually automating everything and what that means for the content and the clients that you are working with.
Yes. As you said, no one is coming to us—we are very explicit about what we are—so no one that is not in AI is coming to us. Fair enough.

(49:15) **Consumer Perception of AI Content**
Put it this way: I am sure there are consumers that look at something and say, "Oh, that is AI," and they can reflectively have a knee-jerk reaction. There are all kinds of flavors of people, people are entitled to their opinions. I would say, number one, that delineation of being able to tell, "Oh, that is AI," is increasingly non-existent. It is going to get less and less of a problem if it even was. At the end of the day, I just look at what is the cost per conversion, and enough people are fine with it that it is still lowering cost conversion by 30% to 50%. Maybe we are ruffling some feathers. Maybe some people do not like it. Maybe they skip past it. But that is no different than the other 4,000 ads they skipped past this day. Ultimately, the only way to have a statistical approach and do this at scale is with AI. There is no other way to have that volume. I had enough conviction this was not going to be a problem even a year and a half ago, when it was a problem sometimes. But I was like, it is not going to be. I think that is what we see playing out. But you are right. Some people are not going to like it, and then those ads will not work. But I think by and large, most people have no idea. My mom sends me cat videos all the time of cats dancing with top hats, "Is this AI or not?" I am like, "Come on." I think it is not a huge problem.

(50:39) **Nonprofit Ad Spend and Adam's Follow-up Question**
By the way, Emily, Google offers—you probably know this—$10,000 a month of free ad spend for nonprofits, and we have an interesting way to take advantage of that. So hit me up if you want to talk about that. Adam.
Hi. Yes, I have a number of questions. And a couple of comments. I think, first of all, I was glad to hear you talk about the difference between what you are doing and brand advertising because what you are doing is performance. It is short-term. I think most of us here have been around long enough that we see the need for a balance between some brand awareness, education, things like that. Performance marketing is very quick turn.

(51:24) **Historical Data vs. New Explorations**
I completely agree with you. There is a lot of AI that can handle this. There is a lot of AI that should be in this ultimately. I think a hybrid model is what we will end up with. I was wondering from early in your talk, you do all of this stuff algorithmically that is backwards looking. It is all historical. How do you get new stuff in there? How do you get, "Gosh, maybe we should try this market, maybe this other kind of advertising will work, maybe this other channel will work"? Do you account for that in your algorithms?
Absolutely. Let us take it in terms of channels. Yes, right now we are connected to two channels, the two big ones, Meta and Google. But ultimately, I am trying to get 30-50. If there is an API where I could send campaigns to digital advertising displays in Omaha and get performance data back, we will absolutely use that in any way that we can. I think a lot of the platforms are overlooked for various reasons. If we are only optimizing based on performance, any channel, we are always exploring that.

(52:30) **AI Creativity and Data Feedback Loop**
In terms of concepts and things to try, AI is extraordinarily creative when asked the right questions. The way that we handle that is basically for everything that we do, we feed back in, "Here are the things that we have already tried, and here is how they have done, so let us try some new stuff." It inherently, given that our goal is to sample the infinite creative space as effectively as possible, it behooves us to always be trying new things and then see how those do. That is tempered by it is not all just random shots in the dark. That is where a feedback loop comes in to try to balance here is what has worked, and so let us try things in the neighborhood, exploring from these new local maxima that we find, and so where we can find even better local maxima. Yes, it is—everything is inherently backwards looking because you need data to then see what happened and make better predictions about the future, but those predictions are always informed with a bias towards trying new things that are more likely to succeed than other new things. That is the only way that we know, and that is how we outperform as well because we just try so many things, something is going to work.

(53:50) **Continuing the Discussion**
And Misha, do you have a hard stop?
I am good. I am good. Keep talking.
Okay, for those who need to go with the hour, that is fine. You can always catch the replay, but we will keep this open a few more since there are still some questions coming in. Adam, did you have more, and then we can go to Jay-Z.

(54:16) **Adam: AI as a Commodity, What's Next?**
I have one more, which is that eventually, because this does work for performance advertising, this goes everywhere. Eventually, you are not going to be that unique. Eventually, this becomes the commodity. So what is next?
So, in any kind of software business, there are two kinds of defensibilities driven by network effect: people and data. In both of these, your system gets more valuable with either more people added to it or more data added to it.

(54:55) **Defensibility: Data Moat and Platform Independence**
I am very clear-eyed about as a builder, this is an amazing time. I have built the most sophisticated, useful, economically valuable system I have ever built for a tenth of the cost of what it—and I am a guy that has been around, but I am not unique in that regard. So our approach to this is that ultimately, I talked about—I say our data schema is one of our biggest assets. The way that we store all those different 200 creative and business decisions that go into every experiment we do. What that means is that the 50th time that we are running an ad for dog food, we are not making random shots in the dark. We are knowing what has worked for this customer, for other customers in this industry, and in general, and what is working right now, because it is constantly changing. Our bet is that the data moat that we can accumulate over time that allows us to make better and better predictive guesses than the next company. Someone can copy my interface pixel perfect, sure. Someone can even—it might be harder but—reverse engineer my schema and figure it out and copy that. It might take a little bit longer, but let us even say that. What they cannot do is then catch up on all of the learnings that we have had so far. So when a customer goes to them versus to us, we are going to perform better.

(56:17) **The Platform Challenge: Fiduciary Responsibility**
Also, in practice, most of these markets evolve into a handful of winners, and then new entrants might try, but most of them are, "What is their story for how they are going to win versus someone that is already established and has a bunch of data?" One other big thing I think about defensibility is the platforms themselves. Certainly, Google and all of them want to bake this right in, vertically integrated, a one-stop shop. Zuck says so much about Meta. The reason I am not afraid of that is that, I mean, they have this right now, and it is not very good, even though we are using the same models, and I have all kinds of opinions about that having been on the inside and understanding how it works there. But even if they make that excellent, as good as mine, Google will never tell you to go spend 30% of your budget on Meta. They cannot do that, or vice versa. Somebody needs to be the marketing fiduciary on behalf of the customer that is only focused on optimizing their spend. I think that inherently has to be a layer outside the platforms that is using the platforms most effectively in concert. That is where I see our play. That is what agencies are doing now. I am just trying to be a better agency with better performance and better cost direction.

(57:33) **Jay-Z's Question: AI Replacing ICP/Persona Identification?**
I do not want to dominate the conversation, but I have a couple of other things I would like to talk to you about. I will reach out to you all.
Fantastic. Thank you. Thanks, Adam.
Hey, Misha. Thanks for walking through all the agents coordinating. David, Adam, Karen, it has been a minute. It is nice to be back in the session.
Good to see you.
Yes, I miss you guys. I work with a lot of early-stage startups and do user research, who they are selling to, and even what features to build from a product perspective. For example, they say, "I think everyone will buy this." And I say, "No, I think only HR is going to pay for it, but maybe all team members want a meeting AI that tells them if they are being a good collaborative partner or not." So they say, "Why do not we just spend some ad dollars and see who clicks?" I am curious if you think your solution replaces finding an ICP. The AI will find your ideal customer profile for you, and you do not have to do the hard work of prioritizing them in the beginning.

(58:47) **Persona Generation and Embracing All Credible Options**
When we get inputs into what this organization is about, what this offering is about, and we feed that into our persona generation agent, it does a pretty good job. It does as good a job as a human would with the same inputs. There is not one persona. There are a hundred different personas to sell this exact same protein bar, and how you are going to appeal to them. I think the old way used to be, "Let us get our heads together about our ideal ICP and market to them." But I am like, why not everything? Why not spell it all? At the end of the day, we do not know what is going to work, and different things are going to work for different people. I am like, let us try all things that are credible and have a reasonable chance, and let the performance data tell us.
It is like machine learning with feature expansion, right? If you are doing your job right and better than the existing systems, you should have some unsupervised discovery.
Absolutely. Our agent that says, "Great, we are going to do another experiment. What are all the choices down the chain? What persona should we try this time?" We have tried this persona a handful of times, it is doing pretty well. Let us throw another experiment in there. Here is another persona. We have not tried this yet. Let us try that. Again, it is all just shots on goal. I want to give a lot of credit to these ad platforms. The way that it works is I can have a campaign. A campaign is the object that holds budget and learning. The budget is important because all the things that are sub to that campaign, be it asset groups in Google or ad sets and ads in Meta, they are all competing for that same budget every day. I can say, "What do I know? What do I know if it is stay-at-home dads or soccer moms? I do not know. It could be any one of these things. Let me put them all out there, and let me let them compete against each other to see who wins the budget." Ultimately, what do I care about? My lowest cost per conversion. As long as they get that, that is cool.

(61:09) **Brand Constraints vs. Performance Optimization**
Now, a brand can say, "For whatever reason, we care about only doing this persona." Great. No worries. Just have that be an input, and we will absolutely stick to that for you. It is probably not going to be as statistically good performance as if you let us try anything. You will never be a replacement for some ICP or persona stuff because even in a perfect world where you are jacked into the ERP and you are able to optimize towards those types of profits and stuff, there is always going to be some e-commerce people. That is cool that you get us these people, but they are not as high as an LTV.
Great. But then feedback, and we know that this person—I see where you are going with it all, and I agree with that. But Jay-Z, I do think there is still a role.

(61:55) **Value Rules for Optimized Conversions**
But let me speak on behalf of the robots here. Honestly, how do we handle this? One of the levers that we optimize with these ad platforms, I think a lot of people do not even do that they should, is value rules. Maybe a conversion is not all created equal, and we feed back in, "Sure, but then this person is going to make us $10,000, that is going to make us $4,000." So it is going to train these models to find more of the people that make us $10,000.
I agree. I agree with you in a reductionist sense, but I will give you a great example. Elective healthcare, aesthetics, med spa: $300 Botox, $3,000 breast surgery, $5,000. We were working helping a PE roll this all up. We were like, "Great, just feed us. We just need to know which ones end up converting to the $300 versus the $5,000 because it is night and day on CAC and LTV ratios." And the practice management system was such a mess. You are assuming perfection, and for a lot of industries out there, they will be able to feed that value back into your loops.
I know it can be a mess. We have had to work with customers to craft somewhat elaborate formulas that end up spitting out some sort of volume number that we can feed back in. Right now, yes, that is in the realm of human judgment. But as I look at it, and I think about, again, we catalog every single one of these learnings, and I think about, "Could an AI have come up with this same thing given the same inputs with the right prompts, and can it ultimately with a computer-use agent or an API-use agent be able to implement this?" I do think so. It might take me until the end of 2027. But I do not see any of these things that an agent cannot do. The numbers will always be the judge. That is why I am in performance, because it just—

(64:02) **Inputs, Data, and Collaborators**
I think Adam is saying with unsupervised machine learning, you still need some inputs. Misha, you are saying you do not need any inputs.
You need inputs initially. What is your product? What is your offering? Who do you think your customers are? But after that, it is about what the data is telling us. It is not unsupervised. It is supervised by the data. In a world of performance advertising, I agree with that.
Question for you: if you are showing up in a different way as an agency—you are showing up as not software as a service, but service for software—are there different types of collaborators than an agency would have? We are fractional CMOs. He loves talking to agencies because we are running RFPs all the time as the fractional CMO. I do not know. Different world of collaborators, maybe.
Yes, somebody at the end of the day needs to be a human that visits AdSmith.ai and has a conversation with a thing. Currently, I am not even there yet. Currently, somebody needs to be hands-on keyboard using the tool.

(65:10) **Targeting Least Sophisticated Advertisers**
I do not care if it is a fractional CMO or an in-house or an agency. I do not even care if it is an agency if they white label us, as long as I am getting paid. We are all good. I do think that my goal is to have it be that the business, the unsophisticated at marketing business owner themselves, can just come and talk to our system the same way they would talk to an outsourced option, be it an agency or whatever. Sometimes they do not even want to deal with that, and sometimes they want to delegate that responsibility to a fractional CMO or something else, and that person will be the one doing it. I am trying to optimize. I see the biggest opportunity in helping the least sophisticated advertisers because they currently—that allows me to expand the market of marketing services. It is also realistic because what I found when I go to talk to bigger brands, at least today, if they are big enough to have an internal creative team, it is a no-go. They hate us.
Yes.
As they should. I get it. I would too. If I am sorry, it is either me or somebody else. Somebody is going to be doing this. We are going to have to work our way up to the bigger brands until they are comfortable, until they just cannot argue economically why not to do this. I like serving the smaller brands at first. If we can do that scalably, we have a lot of longtail to scoop up.

(66:57) **Typical Client Budget and Contact Information**
The last question that I need to pop: what is the typical budget for this? What are your clients spending on a monthly basis?
We say the minimum—this is not a hard rule. Our system will work with anything, but we tell customers you should spend at least $100 a day for a given campaign. So $3,000 a month. Then it is going to be, in that case, $300 for us on top of that, and then maybe $100 additional for the creative. So maybe $3,500 a month. That would be the minimum. You can go less than that, but I warn my customers, "You may not be buying enough impressions for us to have statistical enough learning to learn from." Because we are guessing too, but we are just guessing in the most statistically efficient way. Some people spend way more than that, but if people cannot spend that much, then I am like, "Do not waste your money on advertising. Do something else."
You need to throw enough against the wall. Something sticks.
Something. And again, not to imply that your stuff is…
No. And some of it is going to perform like, for sure. Anything does, but everything is credible. This could work. Shot on goal, and that is what we aim for.
Do we have your email? I will put it in the chat here. Just misha@adsmith.ai. I also have my—I know Dave sent earlier, but here is also my LinkedIn.
All right. Thanks everyone.