The Autonomous Marketing Harness

Julia Nimchinski:

And next up, for our opening keynote, we’re super excited to welcome Gil Allouche. Gil is founder and CEO of Metadata, the world’s first autonomous marketing platform. And they raised $50 million, and they hold 4 issued AI marketing patents. Gil managed to be 50 steps ahead all the time. How are you doing? What’s new? And let’s get this started.

Gil Allouche:

Hello, and hello everyone. Thanks, Julia, for having me. Yeah, I’m super excited about this talk. I’ve also listened, before on a regular, and then on the Zoom to Martin and Russell, and I couldn’t agree more that, we’re… well, first of all, AI is going to become less and less technical. And two, that marketers are becoming more and more technical, so that those two are converging, and and I’m here to talk about that, and what is possible today with, with Agentic go-to-market, Let me share my screen. And… I’ll walk us through it. And I’ll leave 10 minutes at the end, or even more, to ask me questions.

Julia Nimchinski:

Amazing.

Gil Allouche:

You looked at my screen.

Julia Nimchinski:

Yep.

Gil Allouche:

Okay, cool, just move the Zoom. Panel? Alright. Cool. Well, it’s nice to meet you, everyone. My name is Gil Allouche, I’m the founder and CEO of Metadata, and we started about 10 years ago. My personal background, I’m a software engineer, I’ve done AI and robotics, published a paper, and then did a few years working as an engineering manager. Then I completed my graduate school and worked as a marketer. Out of everything, I wanted to get closer to the business side. And so I… I wanted to, you know, the closest thing without being sales was for me to generate the meetings, and I actually ended up doing some sales, some inside sales, to really get a sense of… okay, we know how to build great products, and even today, it has become even more of a commodity to build products.

Anyone can build a product. But the distribution, the marketing, the sales motion, the activation, the onboarding, that’s where the closest thing to Amos lays right now. And, Honestly, I said that to a lot of people, I feel like I’ve been waiting for this age of AI to come. And this is… this is our error, you know? This is the time where there are no more bottlenecks, no more limitations, anything can be done, and almost instantly. And so you have to choose what you want to put your time into, where… what products should you build, because… not just because you can, and how should you do marketing in this new age?

And so I’ve been running a… I remember the first time I did any Agentic go-to-market, which is about a year, and some time ago, a year and a half ago, maybe, where I plugged in HubSpot API, Google Analytics API, and there were maybe 3 or 4 other MetaRata MCP, and a few others, and a few other systems. And I got more answers in one day than I got in the year prior. And I was also able to execute and take action on top of that, like add form fields, form fields to my HubSpot demo request, and do progressive profiling, and really do pretty good attribution, in plain language.

So. What I’m going to show you today is something that is the next version of that. and how I’m running the go-to-market motion for metadata. And this is kind of a new… there’s a new motion I recently, started working, maybe, like, 3 weeks… 3 weeks long. It’s just in the beginning, and it’s showing a nice, hockey stick, but we already generated about 300… 30 meetings, sorry, from the… from our named target accounts, which is great. We have, you know, we’re starting kind of, like, taking it easy in the beginning, and then, pushing it up, and I’m here to share what works, what doesn’t work, what does it look like.

Gil Allouche:

So, first things first, what is a harness? What is… when we talk about the JT Harness and the go-to-market harness, what does that… what does that mean? And so the first thing is. Having, a ledger, which… really, everything that you’re about to do needs to be written ahead of time. There is no, like… you know, hoping that the outbound email is gonna be personalized. There is no, letting an email or an ad come out without first reviewing it and knowing exactly ahead of time how it’s going to look like. That’s one of the most important parts, because this is what differentiates people who are just running the show with AI without any governance or any review system, and sometimes get burned.

And once you get burned. you are very hesitant to go at it again. And so that’s the thing I really, you know, leave you with, is being able to do it really good from the first go is going to give you a lot of tailwind to keep going, and do even more, and be… let AI be more autonomous. The second one is gates, and and not gates without… you can do… you can do different types of gates, and gates are essentially rules that you agree on and direct the artificial intelligence, direct the LLM, to follow.

And you can do it in two ways. you can do it agentically, which I do not recommend, or you can do it deterministically. And the difference between the two is that the second one is predictable. And so if, you know, if you go into ChatGPT today and you ask it the same question twice, you’re gonna get two different answers. And… but if you are running running the same question through a function, through some sort of deterministic code, you will get the same answer, and you can predict ahead of time, and you can verify and confirm that this is the right answer.

And this is very important, because if you have gates, and gates are essentially rules to ensure that no hiccups are happening, you better have those hiccups predictable, deterministic. And so that’s… that’s the second most important thing about a harness. And then last but not least. I, you know, I’m not a huge fan of human loop. I feel like humans, we are pretty much a bottleneck, for… for Genetic systems. Like, they can do a lot more without us being and saying, yes, yes, approve, approve, I agree, this is a good idea. And so… but the way to do it, without, again, getting burned. is having a confidence bar.

So, every decision, before it’s being made, should have some sort of a confidence rate. And if the confidence rate is very high, and by very high, I mean 99% and above, then you’re fairly comfortable. You’re not gonna make zero mistakes, right? If you’re moving to running your company. using agents, they are going to make mistakes, like humans do. Like when you drive a Tesla, it goes into accidents, like humans do as well. And so, the same thing applies here. The question is, where do you interject and ask. the AI to pause, ask you whether this is the right way to go, and go back, and what other, guardrails can you use to help it so it doesn’t make those mistakes.

So these are the three things that… that we use. In this particular go to Washington market, we’re doing everything you can imagine, and we’re doing it from the console, so that you have gifts, direct mail being sent, LinkedIn posts being published, outreach on LinkedIn with invitations, and outbound landing pages personalized and created per account, per prospect. All of the data, all of the targeting is enriched and verified with with zero bounds and hashed keys, so that you can target them using ads. And of course, we’re using all the analytics that you can imagine, like all the website visitors, and all the HubSpot information, and the Google Analytics, and Warmly, and G2, to really understand Where is the market right now, and how can we act on it quickly?

So we were able to, in about, in a few weeks, generate around 30 meetings from those target accounts. These are hard to get to CMOs, and so I’m not running here something… that is spray and pray, but really, we’re focused on T1 accounts, and we didn’t… the T1 accounts are, like, about 400 of them, and really going into decision maker in those accounts.

And doing it using multi-touch, so that they’re… we’re running the ads, we’re running the social outreach, we’re sending the emails, we’re sending the gifts, so on and so forth, and the prospect on their end, when they see, and I’ll show you in a second how the landing page looks like, is personalized to them, the screenshot of the product is personalized to them, the ads, the ad creative that metadata creates. are already generated for their brand, on their brand guidelines. The audiences generated are their audiences based on their case studies and list of customers and competitors.

So everything is done in a very thoughtful way, but autonomously, completely. So the old world are a lot of people, you know, lots of people in the mix, a lot of data, a lot of silos, you know, one person working on creative, one person building the product marketing, one person running the ads, one person running the SEO. Many times, they don’t talk to one another, they’re already overflown with work to do, so there’s a lot of silos. And, the second thing is. it’s, you know, other than the silos, there is a limit to how much execution a human can do, especially when it gets to technical, repetitive, mundane stuff, and so those tasks are fully automated today.

In the new world, you really can handle maybe 100x that you could before, it’s truly the multiplication is there, but you have to do it in an intelligent way, so that, again, you don’t get burnt, you don’t… it doesn’t start sending emails and set up ads with the wrong budget, or out of brand, so on and so forth. But you can accomplish a lot with just 1, 2, 3 people. You can go and manage a spend of $5 million with that small of a team. And in fact, it’s going to be spent better, because there are less… there’s less syncing, less meetings, less confirmations, and kind of forced data sharing that has to happen.

Gil Allouche:

Here’s an example of the landing page that it automatically personalizes for the prospects. You can see everything is customized. There’s nothing here that is generic. Even the logos that are chosen are chosen based on the prospect. And, there is an ad commercial that was generated using Metadata MCP so that customers can see we talk about autonomous systems, but what is actually possible? Can you generate a commercial for my brand? for, for ZoomInfo, for, for Notion. In 20 minutes, in 30 minutes? And the answer is yes, and we prove it here. the commercials, the ads, include… they are… they are running after, a full research is being done.

So you go into their website, you understand their brand guidelines, their guide… their case studies, their resources, their product screenshots, so on and so forth, and you build it based on… based on that. Same thing for the audiences. It goes and looks into your best customers, into your list of customers, into your competitors, into your G2 categories, and understands who is buying those solutions, and generates lists for you. Different segments, competitive segments, buyer intent segment, firmographic, target account, so on and so forth. In this particular one, I’m using Metadata MCP, and I’m really only using four skills.

One is the landing page generation, one is the commercial, one is the ad creative, and the fourth one is the target audience creation. All of these can be done programmatically, which is how I… I launch them using Claude Code. Or codecs, whatever you’re comfortable with. And every day, it’s quite fascinating. I have… I have this running for a few weeks, as I mentioned, this particular motion, and every day, there is a lot of actions to be done, and at first, it wasn’t as smooth as it is today, and I’m sure it’s gonna be even better in the next upcoming weeks, but there is a lot of actions to act on.

So there is, like, some sensing happening. Okay, someone is visiting your website. 30 of your… out of your 300 to 400 Tier 1 accounts have visited your website. okay, what are you gonna do about that? And then actually schedule it and make it happen. While I was doing this presentation, for example, in the last 15 minutes, it sent about 30 outbounds to 30 of those accounts from our target list who visited the G2 category with different messages, depending on what page they visited. They’re also seeing us with ads. And that level of, like, being able to act on it so fast within, you know, half an hour, an hour, or the latest, like, a couple days, that is a big… a big difference than, kind of having a manual motion.

So just to give you some numbers, you know, when I build that list, including all the clothes lost that we have, churn customers, target accounts, Tier 1, 2, and 3, Bombora, G2, website visitors, so on and so forth, I had a pretty big list, then I started shrinking it, and based on, can I find the right contacts within those? Is there something else I can learn about them? Do they currently have something else that they’re running, is it… are they even running ads, and what’s the budget that they’re running? So on and so forth. So I can use all these insights Of course, to personalize the outreach and the ads, but also to understand and rate them and rank them, and see which ones are most important to focus on.

And this is an example. So, there are different journeys based on the different segments. So, G2 intent data, for example, they get a Convo ad, then they see a gift being sent to them via Loop and tie, everything’s through the API, as I mentioned. Then there is a static ad, a regular ad that you see on social, and when I say social, I mean there are 8 different channels. It’s actually social plus search plus LLM, so… They see our ads on ChatGPT, on AdWords, on Bing, and of course on LinkedIn, Facebook, Meta, X, Reddit, So on and so forth.

And then you have the ability to also, in this particular motion, they’re also getting emails, of course, which is the classic touch, but at this point, they already saw our brand multiple times, it’s not a surprise. And the resolution here is pretty good. Like, we’re getting responses, and people are pretty surprised that they visit the G2 category, and they get an email from us within a day, you know, or within a few hours. taking the action is really the most important part, and it’s, again, it’s not going to be perfect the first go, so you should be comfortable with making mistakes, but then use the guidelines that I presented to minimize those mistakes, and make sure that the mistakes are not tragic.

It doesn’t spend a million dollars of your budget out of nowhere, or it doesn’t send some kind of a false language In the ads, or what have you. In this particular recipe, I’m showing you how we’re using a different tool for the different motions, like the LinkedIn outreach, I’m using Goji Berry, the gifting is Loop and tie, the outbound is being done with instantly, the Bire intent, the metadata MCP for the ads and the target accounts. and optimization of campaigns. I’m using Calendly, and I know exactly at any point of time, who is getting which meetings booked.

Did they happen? Did they not happen? I get all of that from the API. And last but not least, Agent Force was here. We have Salesforce, which is the CRM of choice for the majority of the world. And we use that CLI to… To see how much… how many opportunities are actually being generated, what’s the amount? This is just one example, the gifting. The gifting one is one of my favorites because I think it’s a soft touch, people really like it, and it’s nice to do. It gets the attention, and so we’re really going after Tier 1 targets.

We’re also using G2 and Bombora to give us some of that additional intent. And we’re doing a bunch of different, different things. We’re sending them the gift, we’re sending them an email, we’re showing them ads, we’re connecting them on LinkedIn. We send them a note from Loop and Thai after they redeem the gift, so on and so forth. All of that is automated, and the results are pretty good for about 10 who redeem in about 2 weeks, two and a half weeks. Five of them already have meetings, and so it’s a healthy conversion that we’re only going to optimize and improve over time.

I’ll show one more example here, which is the… the ads and commercials. So, one of the things that people say all the time is that video, you know, there’s zero touch, there is the dark funnel, and then there is the video format, which is the most engaging these days. And it’s true that It’s true that video is very, very engaging, so is document ads and kind of old-fashioned static ads. But, in this particular example, we are showing how we’re generating all of those, you know, personalized ads, an ad per account. Not per person, but per account.

For privacy reasons, you’re not able to do the… usually the personalization on the people… on the person base, but you can do it on a company base, and you can do it on a job title base, and so it almost self… qualifies. When someone clicks on, are you a VP Marketing, it’s, you know, running a $2 million plus budget, click here. If they click there, there’s a good re… there’s a good reason for it. Usually, they self-qualify themselves into that segment.

Just to give you an idea of the depth of work, just one of the playbooks The ads and commercials. is running ads on all of these channels, and each and every one of those channels have many campaign types, but you have the document ads, the static ads, the video commercials, thought leadership, boost, conversation ads, message ads, the video ads, so on and so forth. And, the idea is to run many experiments, not too many, to spread yourself too thin. you know, choose a chunk of experiments that you’re going to run, give them the full budget, give them the full time, 1.5x, whatever it is, the budget, give it the time to succeed or fail, and then choose the 20% that succeed, and then move on to the next set of experiments.

That’s the way Metadata does that, that’s the way I recommend to do it, so you don’t spread yourself too thin, but you also identify arbitrage opportunities. If an ad works really well, it’s one of the 2 out of the 10 who are running really well and generating all the demo requests you want for the cost that you’re interested in. you can take that ad now and replicate it to other channels. It’s the same target audience, maybe the same ad, or more or less the same ad, but you’re now expanding the umbrella to have a lot more coverage.

I mentioned before the confidence rate. This gives you an example. This is Metadata 1. you know, recommendations. So, campaigns that we run with metadata are different than campaigns that we run with… directly native on the channel, and the difference is that when they’re managed, we can set them up on the channel, but when they’re managed or optimized by metadata, it identifies issues or, optimization potentials, and it recommends it. You can click Accept, or you can click Accept All, or even set it up on autopilot. And, it… the depth of that… imagine doing 5,000 optimizations in a month or so.

That is humanly impossible, but it really brings you… brings down your customer acquisition costs and increases your conversion rate. Instead of going up, you can see the customer acquisition going down over time. And that’s the place where Human in the Loop is has an advantage. If the confidence rate is not there, and it’s not 100% sure, it asks you, and you should build it in a way that it asks you for your confirmation.

Gil Allouche:

Some lessons learned from mistakes as well. So the enrichment was one of the things that we saw is really hard. We’re using, like, we’re using our own audience graph, we’re using 64, and CrossData, and Appify, and a list of others, and it’s still not easy to find You know, all the keys, all the emails, the phone numbers for the people that you want, and confirm that they’re actually working. Gifting is very interesting as well, you know, we’ve tried different, different mechanisms. We’re trying three… two different mechanisms this week, so we’re still kind of measuring what gets us the highest ROI while creating a really good impression, good first impression.

And this is, as of last night, I have number one here, but you can see that the LinkedIn, motion is really not working yet. It’s running, but it’s not working yet. I haven’t… it might be affecting other… other outcomes, but I’m not seeing a direct response as much as I’d like to. That’s it. I went through it really fast, because it’s quite comprehensive. If I showed you the 25 skills, they’re long and they’re complex, but once you have them, you really do have a go-to-market motion that is fully Agentic and autonomous. And in next month, I’m going to run a workshop for about 2 hours.

And I’m going to, at the very least, you’re going to leave that, you know, that workshop with the ability to understand how to run the first, you know, at least one, one playbook out of the one… out of the eight that we’re running. with your systems, with your stack, with your CRM, with your marketing automation, your target accounts, it will start with the definition of the target accounts, and of course, with the execution and actions against them. And so, if you’re interested, you’re more than welcome to join, and and it’s gonna be about October 7 or October 8th.

That’s it. I’m going to open it up to see if there are any questions, or any comments that I can answer.

Julia Nimchinski:

Phenomenal, Kino. Thank you so much, Gil. People get butterflies thinking about autonomous marketing, finally. Let’s address a couple of questions here. The first one is quite technical. Do you have an extensive monitoring and alerting system built into your marketing harness?

Gil Allouche:

Absolutely, yeah. I have a good amount of gates that are always on. For example, for the outbound, there is one check that is being done for the formatting of it. The other one is the language, the other one is the, you know, the… in the ads, for example, we’re making sure that it looks at the ad after it generated it. Is it on the brand guideline of the target account? So on and so forth. So we do have a lot of them, and one of the hacks I’m using to really check the LLM itself, because the LLM hallucinates and agrees with whatever you say sometimes.

And so one of the ways to check it is to use this concept called the LLM console, which is where you go to… you ask the LLM to go to OpenRouter, which has all the models, you give it the API, and then you tell it, hey, for this particular question, I’m not 100% sure you got it right, so go and ask 10 other models. And it’s, like, a little bit like a jury of LLMs, and they debate, and they either reach a unanimous vote, or they don’t. And if they reach a unanimous vote, your confidence level just went through the roof, because if 10 different models agree that there is one direction to go, the likelihood that this is the right direction is very high.

Julia Nimchinski:

And that’s actually the second question. What is the rate of failure where hallucination result in incorrect asset?

Gil Allouche:

I mean, the more you do… I mean, there are a few ways to answer it. It’s getting better and better, but it’s still pretty high, meaning you have to put a lot of guardrails in really good environments. So, for example, I have a knowledge vault connected to my LLM, so it knows about all the projects that everyone is working on. So it’s not like I’m gonna type a project one day, and then the next day it’s gonna come up with two different ideas, it’s gonna go into the directory, it’s gonna learn what it did, what was successful, what was wrong, what were the mistakes, and not repeat them.

The second one is the LM Console that I mentioned. The third one is, there are skills out there that are already vetted. Like, for example, Firecrawl has… The developer index, I think it’s based on 70 million, applications or projects that were generated, and so it really knows how to clean up your… your deterministic code, so when you tell the LLM, build me this go-to-market motion, you run it by this skill, it’s another hour, another $30 of API costs. But it will find you all kinds of bugs and issues that you weren’t even aware of, and even the LLM console didn’t catch.

Julia Nimchinski:

Awesome, the next question here, how do you… how do you protect against a mythos or hugging Face crisis?

Gil Allouche:

What do you mean by crisis? I guess this would be the question I have. What, what, what, what…

Julia Nimchinski:

I guess security risk.

Gil Allouche:

Oh, security, I mean… I’m having the opposite problem with Mythos, so Mythos Fable 5.1, right? And I’m having issues where it is cautious about things to do, because it’s worried that there’s some malicious code happening. I’m not currently experiencing anything that Makes me, wary about it.

Julia Nimchinski:

Yeah, Gil, you mentioned that enrichment is one of the biggest challenges. I’m curious, how did you solve it, or solving it? How do you approach it?

Gil Allouche:

So I’m doing it by having both a good waterfall of audience providers. So I start with the metadata, you know, audience graph, it has about 20 of them, and then it goes into the ones that we don’t have yet, like 64, like cross-data, and there is a bunch of others. And on the other end, I run the results. first of all, I have, of course, an LLM doing a smoke test, but also I’m running, like, a zero balance against them to see that, are these emails actually working? Can we send them emails? Can I match them against an audience on LinkedIn, in Facebook, etc?

And if they pass all of these gates, I have pretty good confidence that we’re there, which is why the number is not 100%, right? Like, I have, you know, 5,000 targets, 15,000 contacts, but really, out of those, about half. are perfect. The other half are not yet there.

Julia Nimchinski:

Definitely. Changing gears. Curious your vision on the future of this category, and how do you define this category at all?

Gil Allouche:

Who knows? It’s hard to tell, it’s changing so fast, but I will say this, people, I think many times practitioners… the biggest limitation, the biggest blocker is human psychology and change management. That’s the biggest hurdle, for adopting Agentic go-to-market or any kind of autonomous operation at this today. The technology is already here. there is no way… there is no need to wait for the perfect AGI, or what have you that you have in the books. Like, you already have the technology today, and you can do 100x of what you were able to do a week ago.

And so, I think that more and more people are going to to gradually get to that place, and I think many are gonna build their own Agentic systems, and they should, and my advice will be to set up those guardrails. Ahead of time, so that the benefits are high and the AI slop is low.

Julia Nimchinski:

Agreed. And last question. How do you see the future role of a CMO? Is it an architect? We kind of discussed it previously. Gong Chief Revenue Architect, how do you envision the marketing alternative?

Gil Allouche:

Yeah, well, there are two aspects for the CMO. I think there is the execution, and then there is the, the taste, the… the communication skills, and so I think the CMOs are going to… the best CMOs in the future are going to be to have more of those, you’re talking about Gong CMO, like, having more of those soft skills, knowing how to engage an audience, put those pictures that attracted engagement. put the right, you know, like, Adam Robinson’s type of posts that attract thousands of comments. Like, this is how you really build a mode in a world where the execution is the easy part.

And then the second part, of course, is that, like you said, the architect and the designer. You can today take a piece of paper. draw with a pencil how you want your go-to-market look like, and then upload that picture to Cloud, and have it built the actual motion for you with the APIs and the MCP server, and that’s something that I recommend strongly.

Julia Nimchinski:

Amazing. Thank you so much again.

Gil Allouche:

Thank you, it was a pleasure being here.

Julia Nimchinski:

You invited our community to your event, I imagine that’s the best way to re-engage.

Gil Allouche:

Absolutely, yeah, this is a good way to… to… to get your hands… hands in the actual practitioner mode.

Julia Nimchinski:

Awesome. Thank you.

Gil Allouche:

Thank you, John.