“Yes.” Building Claude Code for Marketing | Kashish, Hightouch의 전체 녹취록

was the unique insight on the travel product and what was the unique insight when you pivoted? 00:03:54.319 --> 00:04:04.319 Yeah, so I would say um in travel, the unique insight was just that like people will pay money and use your service for like basically delegating their agency. 00:04:04.400 --> 00:04:05.439 It's very similar to LLM. 00:04:05.599 --> 00:04:09.520 Like, if the LLM can book better travel than you, then you will just have someone else do it. 00:04:09.680 --> 00:04:12.719 Um the original high-touch product was actually not an AI marketing product. 00:04:12.879 --> 00:04:16.560 The original high-tech product was an AI customer success product. 00:04:16.800 --> 00:04:25.199 And so we basically took this like a travel agent is nothing more than like a really, really good concierge or like customer support, customer success. 00:04:25.360 --> 00:04:27.600 So we took that but applied it to BDB SaaS. 00:04:27.680 --> 00:04:32.720 And we said, these B2B SaaS companies, they should be doing their entire customer success motion over Slack. 00:04:32.959 --> 00:04:34.959 That was the original high-tech product. 00:04:35.120 --> 00:04:40.879 Um, we realized that customer communications is not that valuable, but customer data is extremely valuable. 00:04:41.120 --> 00:04:42.639 So that took us to customer data. 00:04:42.800 --> 00:04:45.519 And then customer data took us to reverse ETL. 00:04:45.759 --> 00:04:51.920 Um, we found that getting CSMs and AEs to do their work in some new platform is unfeasible. 00:04:52.079 --> 00:05:00.720 But putting better data in Salesforce and better data in gain site is actually extremely valuable because the salespeople and the CSMs already live in Salesforce and GainSite respectively. 00:05:00.879 --> 00:05:06.560 And better, richer data in those tools would just transform what they what value they can provide their customers. 00:05:06.879 --> 00:05:08.160 So that was reverse ETL. 00:05:08.240 --> 00:05:12.079 And how did reverse ETL come to the current product CDP? 00:05:12.399 --> 00:05:14.480 So maybe for the viewers, like we can give some background. 00:05:15.120 --> 00:05:18.000 2020 is when we started working on reverse ETL. 00:05:18.160 --> 00:05:23.439 Reverse ETL is basically you give me a SQL query, I'll pipe that SQL query to your SaaS tool of choice. 00:05:23.839 --> 00:05:25.519 So you can maintain one central database. 00:05:25.600 --> 00:05:29.519 Let's say this one central database has all your user metrics, revenue metrics, and so on. 00:05:29.680 --> 00:05:30.639 You just maintain that. 00:05:30.720 --> 00:05:35.680 I'll pipe that to your sales tools, finance tools, customer support tools, and customer success tools. 00:05:36.079 --> 00:05:42.800 So now you can maintain the same definition of revenue, LTV, whatever you want in all your different tools just by updating your database. 00:05:42.959 --> 00:05:45.600 That was the original value pop of reverse ETL. 00:05:45.839 --> 00:05:52.639

B2C companies loved it, B2B companies loved it, really everyone just needed this abstraction because they didn't want to build these reverse ETL pipelines internally. 00:05:52.879 --> 00:05:54.800 So we did that for the first two, three years. 00:05:55.040 --> 00:05:59.680 After doing that, we found that marketers were the ones willing to pay the most for this use case. 00:05:59.920 --> 00:06:01.920 Because marketers have this cardinality. 00:06:02.079 --> 00:06:06.800 They have like 50 ad tools, a few different customer engagement tools, and so on. 00:06:07.040 --> 00:06:10.399 And the value of customer data for them could be millions of dollars. 00:06:10.639 --> 00:06:16.720 If you sync the data correctly to Facebook, Google, Snapchat, TikTok, you can now stop marketing customers that already bought. 00:06:17.199 --> 00:06:24.399 One of our banking customers immediately saved$6 million by unmarketing customers that already bought, right? 00:06:24.480 --> 00:06:29.040 So they're happy to pay upwards of like$600K to a million dollars. 00:06:29.360 --> 00:06:35.519 Um, because marketing became the number one use case, we started building higher-level UIs and interfaces for marketers. 00:06:35.680 --> 00:06:37.120 That became the CDP. 00:06:37.279 --> 00:06:40.720 And so what we discovered is that there's this market called customer data platform. 00:06:40.879 --> 00:06:49.759 Customer data platform is inherently flawed because all these CDPs said, I want to store your data and I want your marketers to live in my platform and do all the things here. 00:06:50.000 --> 00:06:51.360 In Enterprise, this was broken. 00:06:51.439 --> 00:06:58.000 So we talked to all the large enterprises using Snowflake, Databricks, and BigQuery, and they said, I have a CDP, it's my data warehouse. 00:06:58.160 --> 00:07:01.600 Can you just give me a UI on top that makes it accessible for marketers? 00:07:01.680 --> 00:07:04.160 And then they can sync the data to the marketing tools. 00:07:04.319 --> 00:07:06.959 So the syncing platform we had already built, that was Reverse ETL. 00:07:07.120 --> 00:07:11.600 We just needed to build these um self-serve interfaces for the marketers to access the data. 00:07:11.680 --> 00:07:13.759 And that's what became composable CDP. 00:07:13.920 --> 00:07:20.240 And composable CDP is what took us to like over 100 million in revenue and like really, really high enterprise traction. 00:07:20.399 --> 00:07:25.759 And that journey was mostly 2023 to 2026, where really composable CDP blew up. 00:07:26.079 --> 00:07:28.800 And then how did you launch the other new product? 00:07:29.360 --> 00:07:36.720 Um, after we had captured the marketing team, so now we're doing all insights and data syncing for marketing. 00:07:36.879 --> 00:07:41.120 Now the marketer says the next biggest bottleneck is content and orchestration. 00:07:41.279 --> 00:07:52.800 If you can create the email campaign for me or the ad campaign for me, which is the content assembly task, and you can orchestrate that campaign for me in my campaign tools and ad tools, then you're really doing like the end-to-end marketing workflow. 00:07:52.959 --> 00:07:55.199 Um, and all of it is powered by agents. 00:07:55.360 --> 00:07:56.560 So they didn't know that, right? 00:07:56.639 --> 00:07:58.319 They just said, build the workflow for me. 00:07:58.480 --> 00:08:03.120 At the same time, ChatGPT 4 came out and we realized that actually like agents can do all this work. 00:08:03.199 --> 00:08:05.199 And so those two things happen at the same time. 00:08:05.360 --> 00:08:20.959 And so we started building so many more products at the same time to um give them the like lifecycle marketing um studio and then the advertising studio that really does like 80% of the work they would need to do in order to launch their email and ad campaigns. 00:08:21.279 --> 00:08:25.439 And CDP, like customer data platform, have existed for a long period of time. 00:08:25.680 --> 00:08:37.279 Uh, what was your unique insight that you could replace the existing CDPs or say that the existing CDPs are redundant in an AI era? 00:08:37.600 --> 00:08:49.440 Yeah, so this is um still the core insight, and it's still really working, which is that it's impossible for an enterprise brand to stuff their data into some schema that is rigid. 00:08:49.679 --> 00:08:51.039 They'll have their own data warehouse. 00:08:51.200 --> 00:08:54.399 In their data warehouse, they'll have a customer with multiple bank accounts. 00:08:54.559 --> 00:08:58.639 They'll have a customer with multiple pets, they'll have a customer with children, and they're buying clothes for the children. 00:08:58.799 --> 00:09:01.600 These kinds of things cannot be represented in someone else's TAS tool. 00:09:01.679 --> 00:09:03.919 They can only be represented in your enterprise database. 00:09:04.159 --> 00:09:05.679 And so that was the core insight. 00:09:05.759 --> 00:09:10.320 It was that the semantic understanding of data is heterogeneous between different brands. 00:09:10.480 --> 00:09:22.399 Um, and if you can represent that in your SaaS, then um all the marketers and product people, whoever it is, can come in and actually access that data in a very um customized way for their business. 00:09:22.559 --> 00:09:23.919 Um, this is our core offering. 00:09:24.000 --> 00:09:31.679 It's the semantic layer on your data warehouse and this UI, which is audiences, journeys, identity resolution, and so on on top of your database. 00:09:31.840 --> 00:09:33.840 And so the two advantages, there's three advantages. 00:09:34.000 --> 00:09:36.320 One, on day one, we have all the data. 00:09:36.399 --> 00:09:38.960 There's no six to twelve month process of giving us the data. 00:09:39.120 --> 00:09:41.679 Uh, two, the data is infinitely flexible. 00:09:41.840 --> 00:09:47.440 So if you update your definition of like account balance, now you have a set checkings account balance and savings account

balance. 00:09:47.600 --> 00:09:50.879 You can make that update in your database, we'll just inherit that immediately. 00:09:50.960 --> 00:09:56.080 And now your marketers will inherit that immediately, and they can run their marketing programs based on inheriting that. 00:09:56.320 --> 00:10:06.960 So these insights is what resulted in a much more um purpose-built product for enterprise and a like very high-value product for enterprise. 00:10:07.120 --> 00:10:10.159 So there's many enterprises that actually never bought CDP. 00:10:10.320 --> 00:10:14.799 Um, people think CDP was enterprise because all the CDP vendors convinced them of this. 00:10:14.879 --> 00:10:16.799 But actually, CDP was stuck in mid-market. 00:10:17.120 --> 00:10:20.080 CDP never broke into enterprise because it was not composable. 00:10:20.240 --> 00:10:23.679 We built the first composable CDP and then we started just dominating enterprise. 00:10:23.840 --> 00:10:25.200 And that was the core insight. 00:10:25.360 --> 00:10:31.919 And then lastly, um we can now do this for all like Martic and Adtech. 00:10:32.159 --> 00:10:34.240 So we're not like inherently a Martic platform. 00:10:34.399 --> 00:10:37.519 Um we do this for Martec, we do this for ads, and we do this for product. 00:10:37.840 --> 00:10:43.360 So any in-app personalization you want to build will give you this as well as email, as well as advertising. 00:10:44.559 --> 00:10:45.120 Understood. 00:10:46.639 --> 00:10:48.879 And did segment never have that? 00:10:49.360 --> 00:10:49.600 No. 00:10:50.720 --> 00:10:57.279 So um most of the CDPs they had this architecture, which was an event schema. 00:10:57.440 --> 00:10:58.320 Event schema is what? 00:10:58.480 --> 00:11:02.639 It means that I have customers, and the customers have um it things they can do. 00:11:02.799 --> 00:11:07.840 They can do button clicked, cart checkout, um, checkout clicked, and like refund started. 00:11:07.919 --> 00:11:09.759 So each person can do events. 00:11:09.919 --> 00:11:13.919 It's basically impossible to roll up on an event schema, the canonical. 00:11:14.080 --> 00:11:16.399 Actually, how much money does this have in their bank account? 00:11:16.559 --> 00:11:18.399 Actually, do they still use the product or not? 00:11:18.559 --> 00:11:19.279 And so on. 00:11:19.519 --> 00:11:24.480 So the best way to represent a customer is actually um a relational database, not an event schema. 00:11:24.639 --> 00:11:28.159 So almost all these CDPs that were built on event schema were inherently flawed. 00:11:28.320 --> 00:11:30.240 So I mean that's the simplest way I can describe it. 00:11:30.320 --> 00:11:32.080 For more technical customers, they understand this. 00:11:32.159 --> 00:11:43.759 And as soon as you say this, it's night and day that some CDPs operate on event schema, others on the relational database that are set up by the enterprise brand and have truly like all the data. 00:11:44.000 --> 00:11:46.320 And then you get this big governance benefit too, right? 00:11:46.480 --> 00:11:49.360 So let's say I have some large brand like Domino's. 00:11:49.440 --> 00:11:54.559 Um, Domino's Pizza has to now run North America, Europe, and Asia. 00:11:54.720 --> 00:12:00.159 Each one will have a separate database and separate governance, and we'll just inherit all that governance for their marketing teams across the world. 00:12:00.559 --> 00:12:01.120 Got it. 00:12:01.360 --> 00:12:05.120 So you become the most easy platform to integrate? 00:12:05.600 --> 00:12:06.879 Most easy and most flexible. 00:12:07.039 --> 00:12:11.039 So it's like somehow giving them much more power with much lower integration cost. 00:12:11.440 --> 00:12:11.679 Okay. 00:12:12.159 --> 00:12:15.279 And they don't have to to duplicate the the database. 00:12:15.679 --> 00:12:16.159 Correct. 00:12:16.320 --> 00:12:16.960 Exactly. 00:12:17.279 --> 00:12:21.120 And what's the biggest ROI the customers have seen with high tech? 00:12:21.679 --> 00:12:25.440 Yeah, so basically how a CMO thinks of it, and this is all pre-AI. 00:12:25.519 --> 00:12:27.440 So I'm mostly talking about CDP pre-AI. 00:12:27.600 --> 00:12:30.960 Um, the CMO says, I have all this customer data. 00:12:31.200 --> 00:12:37.919 If my experience, my if my customer experience was informed by the customer data, then the customer experience would convert much better. 00:12:38.000 --> 00:12:38.240 Yeah. 00:12:38.399 --> 00:12:45.120 Because your experience in my app and in my website and in my email program would be completely tuned to you. 00:12:45.279 --> 00:12:53.360 And you would feel so much like I know you and I know what you want to buy and what you're thinking about today, that you would always convert with me versus convert with my competitor. 00:12:53.519 --> 00:13:02.000 So they can like clearly track how better customer data increases the conversion of their campaigns and makes their customer more sticky, higher LTV and so on. 00:13:02.240 --> 00:13:22.159 And so, like large gaming company, large media company, large retail company, all three have like case studies where by bringing in what I would call self-serve data for their marketing team, which is what Hatchet provides, they were able to completely transform like their metrics into higher conversion, lower CAC, higher LTB. 00:13:22.960 --> 00:13:23.440 Got it. 00:13:24.000 --> 00:13:25.039 And this is pre-AI. 00:13:25.120 --> 00:13:29.279 So this is simply I'm a marketing team, my data team has all this data. 00:13:29.440 --> 00:13:31.600 I now have access to the data and I can use it now. 00:13:31.840 --> 00:13:34.879 Post-AI, so you called this out, which was agentic CDP. 00:13:35.039 --> 00:13:39.360 We started by building a composable CDP, which we still is the foundation of everything. 00:13:39.600 --> 00:13:46.080 We then added the ability for agents to use our semantic layer to find opportunities within the customer base. 00:13:46.240 --> 00:13:49.600 So the agent

will say, You have this customer base that plays tennis. 00:13:49.759 --> 00:13:52.559 You've never once marketed them a tennis racket or tennis shoe. 00:13:52.720 --> 00:13:57.759 If you just simply did that, instead of sending them the same weekly email every week, you would make extra 30 million. 00:13:57.919 --> 00:13:58.639 What do you think? 00:13:58.799 --> 00:14:00.879 And the marketer says, that's a great idea. 00:14:01.039 --> 00:14:06.559 Um that insights product that's autonomous and able to write abstract queries on the data warehouse. 00:14:06.639 --> 00:14:08.320 Um, that was not possible pre-AI. 00:14:08.480 --> 00:14:14.960 Post AI, it's one of our most used features and allows the marketing team to discover all these new insights about their customer base that they never had before. 00:14:15.279 --> 00:14:20.879 So now you are giving insights autonomously without even the marketer pinging, hey, can you give me new insights? 00:14:21.279 --> 00:14:21.679 Correct. 00:14:21.840 --> 00:14:22.480 Exactly. 00:14:22.720 --> 00:14:25.519 And it's very different than like the AI products you see on the market, right? 00:14:25.679 --> 00:14:30.000 Usually it's some chat box on the market, ask it a question, it'll give you the objectively correct answer. 00:14:30.159 --> 00:14:34.159 This is saying, we know this marketing team needs to reduce churn. 00:14:34.639 --> 00:14:43.120 Let's build research agents that are constantly scanning the customer data, the email data, and the advertising to see where are all the opportunities for us to reduce churn. 00:14:43.600 --> 00:14:50.320 Um, and then that leads you to the next platform, which is now that you've identified these opportunities, you know you want to run the tennis campaign. 00:14:50.399 --> 00:14:54.639 You then need the content to run the campaign, which is then the other products that we're working on. 00:14:55.120 --> 00:15:04.159 So so the so the market that you are in, like the CDP market, there was never a company that became 20 billion to 50 billion, right? 00:15:04.399 --> 00:15:09.200 So so it was the product problem or the uh size of the market problem? 00:15:09.679 --> 00:15:11.120 Uh product issue. 00:15:11.440 --> 00:15:14.639 So um almost every enterprise I meet, what do they tell me? 00:15:14.799 --> 00:15:19.039 They say that they never bought CDP because they thought CDP was flawed. 00:15:19.279 --> 00:15:26.000 They instead invested in Terraform or Oracle DB, now Snowflake and Databricks, and that became their like CDP. 00:15:26.159 --> 00:15:30.879 And then they invested in a team of engineers to make that connect to all their other IT systems. 00:15:31.120 --> 00:15:36.320 So they could not even conceive like the idea of buying CDP because it was broken for enterprise. 00:15:36.480 --> 00:15:37.919 This is like the largest part of the market. 00:15:38.000 --> 00:15:39.840 That's the part of the market that we're going after. 00:15:40.080 --> 00:15:49.200 And so that I think is why the CDPs that you're thinking about that are kind of stuck in the 100 to 300 million revenue range were stuck there. 00:15:49.360 --> 00:15:54.480 And when they were at our scale, so let's say we're in the 100 to 200 million range, they were not growing 100%. 00:15:54.720 --> 00:15:55.679 They were growing 20, 30%. 00:15:56.080 --> 00:16:02.399 And it's because they were only selling to the mid-market, the D2C consumer brand that was trying to get their data ready for the first time. 00:16:02.559 --> 00:16:06.159 Not the enterprise brand that has so much more data than they can actually utilize. 00:16:06.399 --> 00:16:06.960 That's one. 00:16:07.200 --> 00:16:16.639 Two is actually um CDP market is better than you think because Adobe and Salesforce actually have collectively almost two billion of revenue. 00:16:16.960 --> 00:16:17.759 Just on CDP? 00:16:18.000 --> 00:16:18.559 Just on CDP. 00:16:18.720 --> 00:16:20.320 And this is marketing CDP. 00:16:20.720 --> 00:16:21.840 Um, just on CDP. 00:16:22.000 --> 00:16:23.679 Marketing cloud actually for them is even bigger. 00:16:23.840 --> 00:16:27.360 Salesforce has like five billion on marketing cloud, Adobe has like one to two billion on marketing cloud. 00:16:27.519 --> 00:16:30.080 Just for CDP, they have like two billion-ish total. 00:16:30.240 --> 00:16:35.840 Um, and so that's the part of the market that people can't really see because it's hidden within these suites. 00:16:36.080 --> 00:16:42.320 Um, and so these there's a lot more CDP market size out there than you would think. 00:16:42.559 --> 00:16:51.279 Um sorry, and they so yeah, I I think like we we think for high touch the CDP market's gonna become a billion dollars. 00:16:51.440 --> 00:16:56.720 And then we think the lifecycle marketing and advertising markets that we're working on will become even bigger. 00:16:57.279 --> 00:16:58.639 Billion dollar revenue market. 00:17:00.799 --> 00:17:02.399 Would you enter marketing cloud then? 00:17:02.879 --> 00:17:03.919 That's a natural choice. 00:17:04.319 --> 00:17:05.759 Um we're considering it, yeah. 00:17:05.920 --> 00:17:14.400 So we are very much actively working on what I would call content creation for lifecycle marketing and orchestration. 00:17:14.640 --> 00:17:20.960 Um, I don't want to convince an enterprise brand to switch from Salesforce Marketing Cloud to HighTouch for the marketing cloud itself. 00:17:21.119 --> 00:17:27.359 I'd much rather say, use my agents to orchestrate your marketing cloud and to build the email templates that go into your marketing cloud. 00:17:27.599 --> 00:17:31.039 And once I've done that, I've done the strategic pieces that really matter the most. 00:17:31.519 --> 00:17:36.640 So let's take one of the largest customers that you would have, right? 00:17:36.799 --> 00:17:41.680 So a customer, a large customer in an enterprise would pay you what, a million dollars today? 00:17:41.920 --> 00:17:44.319 Or could it go into double

digit million also? 00:17:44.559 --> 00:17:46.720 Uh yeah, let's say we're between one and one and ten. 00:17:47.279 --> 00:17:47.680 One and ten. 00:17:47.920 --> 00:17:50.960 I'd say let's assume four or five million dollars on average. 00:17:51.279 --> 00:17:56.400 Can you can you walk through the life cycle of that customer on on how high touch integrates? 00:17:56.799 --> 00:17:57.039 Yeah. 00:17:58.720 --> 00:18:00.880 Oh, let me think of some examples. 00:18:02.480 --> 00:18:07.200 Yeah, so I think maybe a good example would be like the largest food delivery app. 00:18:07.359 --> 00:18:14.000 Um, so say they start with um they have to do marketing to merchants, they have to do marketing towards consumers. 00:18:14.160 --> 00:18:16.480 This is two different tables in the database. 00:18:16.640 --> 00:18:21.359 Um in the in the past, this was actually two different CRMs and two different CDPs. 00:18:21.519 --> 00:18:25.839 Um, because the CDPs are so rigid, they can only handle merchants or consumers. 00:18:26.000 --> 00:18:29.519 So the data team comes to us and they say, hey, look, it's just data. 00:18:29.680 --> 00:18:32.559 Can you help me build one abstraction on top of all of these data? 00:18:32.720 --> 00:18:36.880 And then my marketers that are doing merchant marketing versus consumer marketing, they'll live in the same place. 00:18:36.960 --> 00:18:39.599 And actually, they can do cross-marketing and whatever they need to do. 00:18:39.920 --> 00:18:47.680 So that was the first step, um, building them just the CDP for the merchant marketers and the consumer marketers. 00:18:48.079 --> 00:18:53.680 Um, the more data-driven that became, the more the conversion in each of these two verticals for them started improving. 00:18:53.920 --> 00:19:08.240 Then they said, with AI, can I drive insights to figure out which merchants need more marketing or which consumers are about to drop off and what I can sell them, or which like dinner kind of like um marketing programs I can give them. 00:19:08.400 --> 00:19:10.400 So we gave them this insights layer. 00:19:10.559 --> 00:19:14.720 They told us that they tried twice to build this insights layer internally. 00:19:14.880 --> 00:19:19.759 Every company right now is trying to do AI SQL and solve like the analyst job with AI. 00:19:19.920 --> 00:19:20.880 Uh, it's not working. 00:19:21.119 --> 00:19:24.559 The reason our AI agent is better at writing SQL queries is two things. 00:19:24.720 --> 00:19:34.240 One, our semantic layer informs that agent on what the right queries are, what the right tables are, which is semi-obvious, but building a good semantic layer and setting it up in enterprise is quite difficult. 00:19:34.480 --> 00:19:39.039 We have a very self-serve UI where you can see our graph and like set up that layer. 00:19:39.599 --> 00:19:42.319 Second, is we taught agents how to stop hallucinating. 00:19:42.559 --> 00:19:47.279 When you ask Cloud to run a SQL query for you, Cloud will

oftentimes lie that it ran the query. 00:19:47.440 --> 00:19:52.640 Or it'll run top 50 in the query and not return the rest of the table because it takes too long. 00:19:52.799 --> 00:19:55.519 And then it'll just fabricate the results for the rest of the table. 00:19:55.680 --> 00:20:09.119 So there's all these different things that happen in AI that we have kind of solved and built smaller agents that are constantly evaluating larger agents to make sure the SQL queries that are run are um actually answering the question that was asked and actually accurate. 00:20:09.279 --> 00:20:13.759 So we are now able to answer all objective insights for them, and that's like the agentic CDP. 00:20:13.920 --> 00:20:21.440 Then they say, hey, like, can you now agentically write all my emails and run all my ads? 00:20:21.519 --> 00:20:24.559 Um and so for this, we've built what we call content assembly. 00:20:24.799 --> 00:20:33.599 Content assembly means that I will take from your existing assets and existing images all the images of like people eating food, all the damn assets, and so on. 00:20:33.680 --> 00:20:37.920 And I'll teach an agent how to create new collateral using these assets. 00:20:38.160 --> 00:20:43.359 So rather than the rather than the agent hallucinating people that are not real and like food that's not real, it's all real. 00:20:43.519 --> 00:20:46.960 But it's um cross-product of real assets into new assets. 00:20:47.119 --> 00:20:49.519 Um that's like the content generation product that we built. 00:20:49.680 --> 00:20:58.880 And this will create any email you want based on the templates of this delivery company and any ad you want based on previous ads that are on brand. 00:20:59.039 --> 00:21:01.519 So we've solved the brand adherence problem in AI. 00:21:01.680 --> 00:21:05.359 There's no other company right now in the world that can do brand adherence as well as high tech can. 00:21:05.440 --> 00:21:17.680 And so when the marketer onboards, they find like within like a few days, we are now like on brand and able to create production assets that are then doing the marketing, um, doing all their ad programs for them. 00:21:17.839 --> 00:21:26.240 So there's a digital finance company, they've stopped doing ads that are assembled by human creative designers. 00:21:26.400 --> 00:21:33.359 All the ads are, all the brand design is made by humans, and then all of the assembly is done by AI and then launched into market by AI. 00:21:33.519 --> 00:21:34.960 So those are kind of like the journey. 00:21:35.119 --> 00:21:42.400 You kind of go from solving the data problem, then the marketing self-serve problem, then the content assembly and orchestration problem. 00:21:42.559 --> 00:21:48.000 And then eventually you start just building these more autonomous agents that are coming to you with work. 00:21:48.160 --> 00:21:54.559 And they're saying, hey, um, this subset of customers is really interested in Thai food. 00:21:54.799 --> 00:21:58.640

You should just market them Thai food every Saturday or every Friday. 00:21:58.880 --> 00:21:59.759 And you can't. 00:22:00.240 --> 00:22:01.920 Because you have no images for Thai. 00:22:02.240 --> 00:22:11.200 So tell someone to go do a photo shoot for Thai food, and then I'll build you all the campaigns you need for all the different regions in North America to run this campaign. 00:22:11.519 --> 00:22:14.720 But you still need some other tool to execute that campaign. 00:22:15.519 --> 00:22:18.079 Only in email and SMS. 00:22:18.319 --> 00:22:21.119 In ads, we can just push it directly to Facebook, Google, and so on. 00:22:21.359 --> 00:22:21.519 Okay. 00:22:21.759 --> 00:22:24.480 And in email and SMS, all these companies have some tool. 00:22:24.799 --> 00:22:29.599 So personalization has not been like clearly solved yet. 00:22:30.240 --> 00:22:31.279 And it won't be. 00:22:31.519 --> 00:22:32.000 Why? 00:22:32.319 --> 00:22:45.359 Because AI is like gonna be able to do a much better job of personalization than was even possible or like even thinkable. 00:22:46.000 --> 00:22:49.440 So like what everyone wants is one-to-one personalization. 00:22:49.599 --> 00:23:00.079 For one-to-one to exist, you need somehow for like inference to allow me to check the data and the past transactions and all the possible products I could sell to any every individual consumer. 00:23:00.240 --> 00:23:04.799 I think there are like architectures and paradigms that will enable this, but there's a lot of work to be done. 00:23:05.119 --> 00:23:12.640 And then it requires the correct data flow in and the correct like human approval flow or agentic approval flow out to actually enable these campaigns. 00:23:12.799 --> 00:23:14.799 I think we're like similar to coding. 00:23:14.880 --> 00:23:22.559 Um you can have like Devin write 60 tickets per hour, um, or not tickets, sorry, um 60 PRs per hour to solve tickets. 00:23:22.720 --> 00:23:23.839 Is it the right PR or not? 00:23:23.920 --> 00:23:24.720 That's the real question. 00:23:24.880 --> 00:23:28.079 So like I think that's gonna be the bottleneck for marketing very soon. 00:23:28.160 --> 00:23:32.079 Um HighTech will solve the content assembly and like data task. 00:23:32.240 --> 00:23:39.279 Once that happens, there will be like so many um PRs for marketers to look at and say yes or no. 00:23:39.519 --> 00:23:40.720 And I think that's the dream state. 00:23:40.880 --> 00:23:51.359 The dream state is um AI can do so much work that the marketer becomes the decision maker and the creative body rather than the like operational executor of these things. 00:23:51.680 --> 00:23:54.640 Most marketers today are afraid that AI will take their jobs. 00:23:55.119 --> 00:23:56.240 I think it's wrong. 00:23:56.480 --> 00:23:57.599 I understand the concern. 00:23:57.680 --> 00:23:59.519 I think it's like totally wrong. 00:24:00.640 --> 00:24:01.200 Yeah. 00:24:01.519 --> 00:24:07.839 So you're how much how many years are we away from that kind of that level of personalization? 00:24:09.039 --> 00:24:10.160 Um, it really depends. 00:24:10.319 --> 00:24:13.039 I think like some brands could achieve that next year. 00:24:13.279 --> 00:24:16.720 Um, but again, it's like the the bar is gonna keep sliding. 00:24:16.880 --> 00:24:25.119 So like we're gonna achieve like 2026 capability in 2027, but then by 2027, technology has moved so much that actually the bar is higher and higher. 00:24:25.279 --> 00:24:27.839 So I think like personalization is a never-ending task. 00:24:28.079 --> 00:24:35.440 But we should achieve like superhuman personalization this year. 00:24:35.680 --> 00:24:40.160 And it really just depends if people like can transform their teams quickly enough to adapt to it. 00:24:40.319 --> 00:24:46.079 This fear that you just mentioned that marketers are worried about their jobs being taken is one of the reasons why transformation is so hard. 00:24:46.240 --> 00:24:57.279 Um, I truly don't think it's possible for AI to take their jobs because um what's instead going to happen is there's gonna be so many more campaigns we can run and want to run. 00:24:57.440 --> 00:25:02.799 So, in fact, you'll need more marketers, more designers, and more just like people to run all these campaigns. 00:25:03.039 --> 00:25:08.240 Um but the individual output per person will increase like crazy. 00:25:08.400 --> 00:25:15.440 So if like one campaign, one marketer runs a campaign per quarter right now, maybe that marketer runs a campaign per week in the future. 00:25:15.599 --> 00:25:21.759 Um but the appetite from the marketing team goes up like 100x, so even outpaces that increase in productivity. 00:25:22.160 --> 00:25:25.920 So do you sell to the CMO or marketing ops person? 00:25:26.559 --> 00:25:28.480 Both, but it's really the CMO. 00:25:28.720 --> 00:25:28.960 Yeah. 00:25:29.119 --> 00:25:33.359 And I've talked to like almost like 50, 60 CMOs in the last six months. 00:25:33.680 --> 00:25:38.559 And they're all wondering how to create this transformation in their company. 00:25:38.880 --> 00:25:55.200 Um and the reason this transformation hasn't happened yet is because Gen AI was so bad at staying on brand that everyone that tried Nano Banana, that tried Gemini, that tried um any of these um image and video models, they experienced the lack of brand adherence and they experienced last year's models. 00:25:55.680 --> 00:26:01.519 Um like just in the last two weeks, the um improvement on Image Gen is like amazing. 00:26:01.920 --> 00:26:12.880 Can you can you share some of the GTM insights that that are unique to you that worked in the last three to four years to reach to the enterprise and make them customers? 00:26:13.200 --> 00:26:19.599 Yeah, so I guess um, it's I'll say maybe last like three years, and I'll say maybe this year. 00:26:20.000 --> 00:26:29.680 So last three years, biggest insights are mostly just focusing on verticals where you can really change their

life. 00:26:29.920 --> 00:26:38.240 So, like some vertical that doesn't have a lot of first-party data, CDP doesn't really matter because most of their marketing program is not going to be influenceable by data. 00:26:38.400 --> 00:26:45.440 Um, you find instead a retailer or um a media brand or a travel brand. 00:26:45.599 --> 00:26:50.400 They know their customer, most of the customers are logged in, they can really influence the customer experience with first-party data. 00:26:50.480 --> 00:27:06.960 So convincing our team to only go after the right verticals, um, hiring the right team, um, really, really being in the details myself of like what are we talking to customers about to get them to convert and then convert at higher ESV ACV. 00:27:07.119 --> 00:27:13.039 These kinds of like really tactical things, honestly, is what grew the go-to-market function and the like revenue so quickly. 00:27:13.200 --> 00:27:15.519 Um, and then just making sure you're going after the right market, right? 00:27:15.599 --> 00:27:17.359 So we never went after SMB. 00:27:17.759 --> 00:27:21.359 Everyone else says go after SMB and earn the right to go after enterprise. 00:27:21.519 --> 00:27:23.119 We went after enterprise on day one. 00:27:23.200 --> 00:27:28.160 Um we were starting to close Fortune 500, at least Fortune like 1000 in the first year of business. 00:27:28.400 --> 00:27:28.720 Why? 00:27:28.880 --> 00:27:35.920 Because we admitted to ourselves that if we're gonna be a serverless later on the data warehouse, we're not fit for SMB, we're only fit for enterprise. 00:27:36.079 --> 00:27:37.920 So we just did enterprise on day one. 00:27:38.079 --> 00:27:42.640 Um and then more recently, what's been working is running a modular company. 00:27:42.880 --> 00:27:47.279 What I mean by modular company is that you have different businesses within your company. 00:27:47.519 --> 00:27:53.519 And instead of them competing for resources and sales headcount and marketing headcount, they have their own budgets and they can just run. 00:27:53.839 --> 00:27:55.599 I think Rippling does an amazing job of this. 00:27:55.759 --> 00:28:00.400 Rippling runs their HR business, IT business, other business quite independently. 00:28:00.559 --> 00:28:05.920 And their GMs are incentivized quite independently to make their business succeed or not succeed. 00:28:06.160 --> 00:28:09.039 And so pre-AI, we didn't need to do this. 00:28:09.200 --> 00:28:10.480 We were mostly a CTP business. 00:28:10.559 --> 00:28:25.039 But in the last two years, we've now built these other two businesses that have their own PMs, their own engineers, soon their own salespeople and their own marketers that will be able to kind of run and not get blocked by approval or like convincing of other folks. 00:28:25.680 --> 00:28:26.480 Got it. 00:28:26.720 --> 00:28:31.839 So tell me more about building a modeler company and when should a founder think of building one? 00:28:32.319 --> 00:28:33.440 Yeah,

probably never. 00:28:33.599 --> 00:28:35.279 I think most founders probably should not build one. 00:28:35.359 --> 00:28:35.519 Right. 00:28:35.680 --> 00:28:39.359 Um, it's much better to have just a coherent strategy and one thing that really works. 00:28:39.519 --> 00:28:49.519 So if you can have like um, let's say you're a ramp, you have so many products, fine, but the corporate credit card is your main product and the main moneymaker. 00:28:49.680 --> 00:28:53.119 You'd much rather just have that, and the corporate card gets you into every business. 00:28:53.279 --> 00:28:56.559 And then once you get in the business, you sell them other things like spend management and so on. 00:28:56.720 --> 00:28:57.839 I think that's ideal. 00:28:58.079 --> 00:29:02.000 Marketing is very heterogeneous, and marketing has many buyers. 00:29:02.160 --> 00:29:06.880 And so by selling to marketing, we kind of don't get to have the one single entry point. 00:29:06.960 --> 00:29:13.519 Um, the person who's doing ads may not care what the IT person is doing or the email person is doing. 00:29:13.680 --> 00:29:21.920 And so we think this modular company makes sense when you have multi-persa and when you want to grow in like nonlinear ways. 00:29:22.559 --> 00:29:25.440 So uh for high touch, right? 00:29:25.599 --> 00:29:27.279 What could be the feeling for revenue? 00:29:27.440 --> 00:29:29.359 Could it be a billion dollars? 00:29:29.759 --> 00:29:33.759 Um, I would at the moment I think it can be multi-billion. 00:29:34.240 --> 00:29:37.519 Um, but we're always learning and iterating. 00:29:37.759 --> 00:29:43.200 So like we don't have any, we're not married to IPOing or getting acquired. 00:29:43.359 --> 00:29:53.920 We basically think like we should try to double revenue or more every single year and make sure that as we become a larger and larger company, we can sustain that pace. 00:29:54.079 --> 00:30:00.079 Um, going multi-product is one reason for that, because we want to make sure that even at like a billion of revenue, what's the next year? 00:30:00.240 --> 00:30:02.720 Is it 1.1 or is it 1.7? 00:30:02.880 --> 00:30:09.039 And we want to make sure that we have the right product mix so that if and when we get to that scale, we can still keep growing at the same rate. 00:30:09.200 --> 00:30:11.759 And so historically we've doubled year over year or more. 00:30:11.920 --> 00:30:14.559 I think last year was like 120% growth. 00:30:14.720 --> 00:30:21.519 Um, I want to be able to maintain that growth, which means investing in the right regions, the right products, and the right uh teams. 00:30:21.920 --> 00:30:32.319 So let's say we discussed about it, but you know, if you can put it in more details, if you are right about your lead on the market, what does the marketing look like in five years and what is the role of human? 00:30:32.720 --> 00:30:33.440 Uh, great question. 00:30:33.519 --> 00:30:33.839 Okay. 00:30:34.000 --> 00:30:40.880 So five years is unpredictable, but I think what will happen in five years is approximately this. 00:30:41.119 --> 00:30:46.480 Um, a very, very um creative and smart marketer will wake up in the morning. 00:30:46.640 --> 00:30:49.599 They'll have five ideas in their inbox created by AI. 00:30:49.759 --> 00:30:53.119 AI worked all night, thought about the five ideas for them to improve their metrics. 00:30:53.359 --> 00:30:58.400 Those five ideas could be in getting new customers, retaining customers, cross-selling, and so on. 00:30:58.640 --> 00:31:16.640 Um, the marketer gives AI feedback on each of these ideas, and then AI runs with those ideas to coordinate across different teams, get the collateral that they need, um, build the email or the ad, and then come back to the marketer with, okay, I've based on your brief, which I helped you write, I've come back with the collateral and everything that you need. 00:31:16.720 --> 00:31:23.440 Um, the legal approval, the brand approval, the collateral, um, the marketing, email campaign execution, and so on. 00:31:23.680 --> 00:31:26.400 And now let's work on this together and launch this campaign. 00:31:26.559 --> 00:31:30.960 Um, the more and more autonomous it gets, the less human discretion is necessary to say yes, no. 00:31:31.039 --> 00:31:34.799 But the human should always say, This is a good idea, or here's a better idea. 00:31:34.960 --> 00:31:37.359 So, very similar to cloud code plan mode. 00:31:37.440 --> 00:31:38.640 Um, the plan is always wrong. 00:31:38.720 --> 00:31:41.599 It's like somewhat good, but like always has something to improve on. 00:31:41.839 --> 00:31:49.119 Um, the marketer in this case should be even more um able to push and pull on what the plan is. 00:31:49.200 --> 00:31:58.000 So maybe the agent says, like, we're gonna run this tennis campaign, and the marketer says, hmm, give me some research on basketball, soccer, and squash. 00:31:58.160 --> 00:31:59.440 And then the AI comes back. 00:31:59.599 --> 00:32:02.559 So this is kind of the like kind of um push and pull. 00:32:02.640 --> 00:32:08.880 And then when we're ready to run the experiment or the campaign, um AI can run it end to end and come back in four weeks with the results. 00:32:09.119 --> 00:32:16.799 So something along these lines where every step has some human discretion: the legal step, the brand step, the creative step, and the marketing step. 00:32:16.880 --> 00:32:25.200 Um, but AI can like run with the task and seek these approvals and come back and constantly be running these multiple threads. 00:32:25.359 --> 00:32:31.200 So maybe one is a research thread for churn, maybe one is like this three new email campaigns and three new ads. 00:32:31.599 --> 00:32:40.000 Um, and it's very similar to like the stuff you see on Twitter for these engineers that are managing like 10 or 20 Devon agents or 10 or 20 cloud codes. 00:32:40.240 --> 00:32:41.839 Um, it should work very similarly. 00:32:42.880 --> 00:32:55.119 And I

think actually within one to two years, um the marketer will be saying that they can't do their job without AI or without high touch. 00:32:55.279 --> 00:33:02.880 So today, if you're an engineer, it's quite impossible to do your job without cloud code, codecs, or like one of the um coding models. 00:33:03.119 --> 00:33:06.000 Um, I spent like a couple weeks this year just coding full time. 00:33:06.160 --> 00:33:12.960 And I literally like got to the point where I just couldn't write, like I couldn't do my job if um cloud code was down for an hour. 00:33:13.200 --> 00:33:16.240 Um and they had a couple outages, like the whole team just stopped working. 00:33:16.480 --> 00:33:28.240 Marketers will start working this way, where without high touch and without like basically, like people will think of high touch as the harness for marketing teams, the same way cloud code is the harness for coding teams. 00:33:28.400 --> 00:33:30.559 Um without that harness, they can't do their job. 00:33:30.640 --> 00:33:47.359 And that harness should have all the context of their past marketing, their email campaigns, ad campaigns, have access to all the customer data, and know what the CMO's top-level goals are, and have the destination um access, whether it's MCP, CLI, or API, doesn't matter, but the destination access to actually execute those campaigns. 00:33:47.920 --> 00:33:54.559 But uh you will only like high touch is limited, I believe, by the customer data the company has. 00:33:54.799 --> 00:34:06.960 You don't have a data of prospects, let's say hypothetically uh dominoes or a uh uh wants to go into one into a new country. 00:34:07.440 --> 00:34:07.759 Yep. 00:34:08.000 --> 00:34:08.559 Right. 00:34:08.880 --> 00:34:14.159 Then then they then they are blinded because nobody has the data on the new country. 00:34:14.880 --> 00:34:17.519 Um I think honestly, post AI, we have everything. 00:34:17.840 --> 00:34:20.320 So we have built previous to AI, right? 00:34:20.400 --> 00:34:24.960 It was sitting on your data warehouse, whatever structured data you have, you have, everything else you don't. 00:34:25.440 --> 00:34:35.119 Now we've built agents that are scraping competitor insights, scraping competitor ads, um, figuring out what triggers matter based on weather or sporting events and so on. 00:34:35.280 --> 00:34:39.840 And we're starting to build a lot of synthetic data for our customers via these agents. 00:34:40.000 --> 00:34:41.119 And we host them all. 00:34:41.280 --> 00:34:42.800 It's all part of the platform. 00:34:42.960 --> 00:34:49.760 Um, and if someone were to drop in a lead list of, let's say, all customers we could go after in Japan, we have that lead list as well. 00:34:49.920 --> 00:34:53.599 And we can marry it to the first party data and have the agent recurse over that as well. 00:34:53.840 --> 00:35:02.079 So, like the world that we lived in before that was like, here's your SQL database, was very limiting, completely agreed. 00:35:02.320 --> 00:35:09.119 And that's why this company is so exciting now, because now like the agent can just operate on unstructured data quite freely. 00:35:09.679 --> 00:35:13.760 So but but the company would still need a qualified lead list to go after it, right? 00:35:13.920 --> 00:35:14.320 Without it. 00:35:14.639 --> 00:35:16.400 Kind of, but like marketers don't think this way. 00:35:16.559 --> 00:35:20.400 Marketers think of it as let me run 50 million of ads in Japan. 00:35:20.639 --> 00:35:23.920 By running the ads, I'll get my lead list in my database. 00:35:24.000 --> 00:35:26.079 And then I'll further market that person more and more. 00:35:26.400 --> 00:35:33.440 So, how would uh let's say if they want to spend$50 million in Japan, how would they approach it using high touch? 00:35:33.679 --> 00:35:34.239 So, this we do. 00:35:34.320 --> 00:35:47.519 We have like basically um we would have the agent look at all the ads that are working in other geographies, develop a brief or a plan for which ads we can run in Japan, and how to allocate that budget across the different content and campaigns. 00:35:47.599 --> 00:35:49.679 Um, and then they would work hand in hand with the marketing team. 00:35:49.760 --> 00:35:54.719 I think something like 50 million in Japan could take weeks for the team to develop together with AI. 00:35:54.880 --> 00:35:59.840 But they would come with a whole like plan of here's the creative, we'll try, and the experiments will run. 00:35:59.920 --> 00:36:04.239 And then the marketer will just incrementally run more and more ad experiments with AI and then recurse on those. 00:36:04.400 --> 00:36:06.159 So it's the same way an ad agency works. 00:36:06.320 --> 00:36:12.320 And it's the same way like high touch would launch in like, let's say, we launched in like um Australia and India recently. 00:36:12.480 --> 00:36:20.320 So we are running ads now in Australia and India, um, but we can do them localized in different languages with the exact same collateral. 00:36:20.559 --> 00:36:24.079 And um whatever doesn't work, the AI just will turn off. 00:36:24.159 --> 00:36:28.400 And whatever does work, it'll bring back to our marketing team and say, hey, this is working. 00:36:28.559 --> 00:36:31.039 I suggest you spend 5x more budget here. 00:36:31.360 --> 00:36:33.280 So that's kind of like how marketing teams work. 00:36:33.440 --> 00:36:38.480 They work in a very experimentation-driven way, and they just double down on what's working and they just kill what's not working. 00:36:38.960 --> 00:36:39.920 Correct. 00:36:40.239 --> 00:36:51.199 Um, and so this is like um, you know, like on Twitter, people are talking about loops and having like some loop where it's like the agent tries a thing, gets a reinforcement loop on if it's working or not, and then tries a thing. 00:36:51.519 --> 00:36:53.119 Ads is actually the best loop. 00:36:53.360 --> 00:36:57.199 So I think ads is a quite um configurable with AI.

00:36:57.519 --> 00:37:08.960 But how would you give the team an approximate ROI that spending$50 million of opening a new market, let's say Japan, would give you 100 million revenue in the first year? 00:37:09.760 --> 00:37:13.039 Um, it's this is like age-old problem in marketing, right? 00:37:13.119 --> 00:37:17.119 Um, marketing success does not necessarily equal revenue success. 00:37:17.280 --> 00:37:18.960 So that is ultimately up to someone else. 00:37:19.039 --> 00:37:25.679 I think what we would have to prove is that if you spend 50 million in ads, then the leads we bring in are worth$50 million. 00:37:26.320 --> 00:37:28.559 And then it's up to like, is it the right product mixed? 00:37:28.719 --> 00:37:30.400 Can we serve that market and so on? 00:37:30.559 --> 00:37:36.960 Um, but like usually ads teams are responsible for like cost per impression or cost per um conversion or lead. 00:37:37.119 --> 00:37:42.480 And then once we've converted that lead, um, it's really the economics of the business on whether it was worth it for that conversion. 00:37:42.800 --> 00:37:46.639 Yeah, but ultimately everybody in a company is responsible for revenue. 00:37:46.719 --> 00:37:46.800 Yeah. 00:37:46.960 --> 00:37:50.159 And this is how in AIH the company are getting structure. 00:37:50.320 --> 00:37:51.039 Are you a builder? 00:37:51.199 --> 00:37:52.719 Or let's say in any company, right? 00:37:52.880 --> 00:37:55.280 Are you creating products or are you bringing revenue? 00:37:55.440 --> 00:37:55.599 Right. 00:37:55.760 --> 00:37:59.760 So I think every department today in any enterprise had that kind of pressure. 00:37:59.840 --> 00:37:59.920 Yeah. 00:38:00.159 --> 00:38:01.840 Because you cannot hide behind anything. 00:38:02.000 --> 00:38:05.760 So a marketer cannot say that my job was to bring an MQL. 00:38:05.840 --> 00:38:05.920 Yeah. 00:38:06.159 --> 00:38:10.719 Now get the sales team to run the SQL or in a consumer company. 00:38:10.800 --> 00:38:15.440 And the marketers cannot say they I got the best leads worth 500 million. 00:38:15.599 --> 00:38:18.639 It's just that the team is not able to convert, so it's not my problem. 00:38:18.960 --> 00:38:23.280 So I mean, since our all of our use cases are mostly consumer, it's pretty solved in consumer. 00:38:23.360 --> 00:38:28.159 Like once you have the lead, um, you have the analytics of whether that person eventually bought the product. 00:38:28.400 --> 00:38:34.239 Um, and so my only comment on this is there are some businesses that can't prove revenue in the first year. 00:38:34.400 --> 00:38:37.679 They can prove like conversion of leads, but then revenue comes next year. 00:38:37.920 --> 00:38:43.199 There's other businesses like quick consumer businesses where the lead should convert in the first 30 days. 00:38:43.360 --> 00:38:46.719 And so I think those businesses can have a good reinforcement loop very quickly. 00:38:46.960 --> 00:38:51.199 And most of these companies that are

running like significant ad spend have that kind of reinforcement loop. 00:38:51.679 --> 00:39:14.079 So uh uh, you know, I have interviewed some of the guests in my podcast, which are like some of the leading AI researchers here in SF, and they have said that uh most problems today, even outside of you know the digital workforce, uh are can if they can be a coding problem, then they can be solved via AI and agents. 00:39:14.639 --> 00:39:17.039 What's what's your view on the marketing harness? 00:39:17.599 --> 00:39:19.519 Yeah, so I think um I guess two things. 00:39:19.679 --> 00:39:25.599 One, uh I completely agree that most problems are coding problems because coding is just an expression of logic. 00:39:26.000 --> 00:39:26.719 Maths, people. 00:39:27.280 --> 00:39:27.519 Yeah. 00:39:27.679 --> 00:39:30.079 And it's ultimately like, what is even coding, right? 00:39:30.159 --> 00:39:42.880 It's like um if if you think of it as like an expression of logic or an expression of thinking, then it's just a way for like the model to operate in like our plane versus like a different plane. 00:39:43.039 --> 00:39:47.760 Um so yeah, I think marketing problems will be solved by coding agents, if that makes sense. 00:39:47.920 --> 00:39:53.679 They won't be solved by coding agents without much, without a harness for them to use. 00:39:54.000 --> 00:40:13.039 But with the right harness, um content assembly, research, campaign ideation, all these things can be solved by coding agents, because the coding agent can like recurse over logic, over data, over campaigns, over image embeddings and video embeddings, and so on to figure out what to do next. 00:40:13.199 --> 00:40:21.360 Um, and then if I answer the question about like what is a marketing harness, I think like basically like you have a foundation model. 00:40:21.519 --> 00:40:25.280 Um, the foundation model needs a harness to operate on marketing activities. 00:40:25.519 --> 00:40:30.960 Marketing activities requires two pieces of context and then some execution layer, the, or three pieces of context. 00:40:31.119 --> 00:40:36.880 The context is your customer data, which comes from the data warehouse and some semantic layer, which HighCitch provides. 00:40:37.119 --> 00:40:42.000 Then it comes from all of your channel data, which is your ad channels and your email channels. 00:40:42.159 --> 00:40:48.639 Then it comes from your business context and your CMO saying what matters, your CEO saying what matters, and what are the business priorities. 00:40:48.880 --> 00:40:54.719 With these three pieces of context, the context window would be greatly exceeded. 00:40:54.960 --> 00:41:05.920 So the harness has to index this information and turn it into vector databases that, or like document stores, that the harness can then talk to and query in order to get the right information at the right time. 00:41:06.159 --> 00:41:08.480 This context engineering task is actually a lot of work. 00:41:08.559 --> 00:41:10.639 And it's heterogeneous between different enterprises. 00:41:10.800 --> 00:41:14.320 Some enterprises use this database, some people use this channel tool. 00:41:14.559 --> 00:41:21.119 Highly heterogeneous task, and that state management or context management has to be constantly refreshed and constantly indexed. 00:41:21.280 --> 00:41:23.199 So that's like the number one task. 00:41:23.519 --> 00:41:30.639 The second task is being able to actually execute that campaign, which means create the content and actually execute in the channel tools. 00:41:30.800 --> 00:41:34.960 Um I can go into detail for that, but that's quite thick as well, as we call it. 00:41:35.039 --> 00:41:36.639 So it's like a thick harness versus a thin harness. 00:41:36.880 --> 00:41:38.880 We'd love to know what's the thick harness, what's the thin analytics? 00:41:39.039 --> 00:41:42.800 Yeah, so I basically think of it as like a thin harness is like a thin wrapper. 00:41:42.960 --> 00:41:44.559 Anyone could build it in like five minutes. 00:41:44.639 --> 00:41:48.000 Um I could probably tell a cloud code agent to build it while we finish this conversation. 00:41:48.159 --> 00:41:55.920 A thick harness is something that it would require me to have like tens of engineers working for months to build this, and actually they still might not finish. 00:41:56.079 --> 00:42:02.639 We will be improving the thickness or like the capabilities of this context engineering probably for the next three years. 00:42:02.880 --> 00:42:10.239 Um, and the more enterprise customers that sign on, the more they'll challenge what context we provide them, and we'll build more and more on the context engineering piece. 00:42:10.639 --> 00:42:12.159 Why would an enterprise care? 00:42:12.320 --> 00:42:15.199 Like an most marketers are not tech non-technical users. 00:42:15.599 --> 00:42:17.119 Yeah, they don't care at all about the technology. 00:42:17.280 --> 00:42:20.719 They care about the AI being able to operate on their task. 00:42:20.960 --> 00:42:23.280 And without the context, the AI cannot operate on the task. 00:42:23.519 --> 00:42:23.679 Yeah. 00:42:23.840 --> 00:42:25.760 So why would they care about the thick harness? 00:42:26.239 --> 00:42:27.039 Uh they don't care. 00:42:27.199 --> 00:42:29.039 It's only high-tech cares, right? 00:42:29.199 --> 00:42:35.519 Because we need to make sure that we're working on a business that is long-term defensible and provides real value. 00:42:35.840 --> 00:42:42.480 If the IT team at this brand can simply create that harness in five minutes, there's no point in high-touch building it. 00:42:42.639 --> 00:42:43.840 They should just build it in-house. 00:42:44.079 --> 00:42:48.639 If it would take them a year to build this internally, then actually it's a great product for high-touch to build. 00:42:48.800 --> 00:42:57.840 So the thickness of the harness is much more about us building something that is durable and it is about us building

something that is differentiated from the foundation model. 00:42:58.159 --> 00:43:02.079 If Gemini and ChatGPT can do it, the marketer will just use Gemini and ChatGPT. 00:43:02.320 --> 00:43:09.440 But if we're the only um harness that can do this, then they'll come to us and we will earn the right to serve them that token. 00:43:10.079 --> 00:43:15.599 So for example, uh ad creation chat GPT can do it, Gemini can do it. 00:43:15.920 --> 00:43:16.719 I would disagree. 00:43:17.360 --> 00:43:17.760 Why? 00:43:18.079 --> 00:43:22.400 I don't think ChatGPT and Gemini can stay on brand, especially in enterprise. 00:43:23.599 --> 00:43:28.079 Um and additionally, not only can they not stay on brand, um, they don't know what performs. 00:43:29.360 --> 00:43:40.000 Once you do this context engineering task, I can look at all of your ads um for the entire history of your brand and compare that to the generations and tell you if this ad will perform or not perform. 00:43:40.559 --> 00:43:50.880 But for example, you know, uh if anthropic and chat GPT uh can give forward deployed engineers to uh an enterprise. 00:43:51.199 --> 00:43:52.159 Yeah, then this can be built. 00:43:52.400 --> 00:43:52.559 Yeah. 00:43:53.119 --> 00:43:58.000 So do you get these questions often that say, hey, what happens if anthropic and open AI? 00:43:58.719 --> 00:44:00.880 Um as soon as I Show them the context? 00:44:01.119 --> 00:44:01.440 No. 00:44:01.679 --> 00:44:15.199 So when I show them like here's all the places we're taking context from and indexing so the LLM can then read and query from this, immediately they're like sold that they need a higher harness for marketing versus waiting for the foundation model. 00:44:15.519 --> 00:44:24.480 So would you say that in the current market serving enterprise, uh high touch has uh has the best marketing harness or has the only marketing harness? 00:44:24.880 --> 00:44:25.280 100%. 00:44:26.079 --> 00:44:26.320 Yeah. 00:44:26.559 --> 00:44:27.519 Only invest for sure. 00:44:27.599 --> 00:44:27.679 Yeah. 00:44:27.920 --> 00:44:28.320 Only and best. 00:44:28.800 --> 00:44:31.519 And it's in its in in it's getting better rapidly. 00:44:31.840 --> 00:44:32.000 Okay. 00:44:32.159 --> 00:44:32.320 Yeah. 00:44:32.559 --> 00:44:34.559 Well, other companies aren't able to build it. 00:44:35.039 --> 00:44:36.800 I think they're not thinking about it this way. 00:44:36.960 --> 00:44:46.159 I mean, maybe they could build it if they tried, um, but they're not investing in context that is unstructured. 00:44:46.320 --> 00:44:50.800 They don't have the structured data context that we have built via CDP for the past five years. 00:44:50.880 --> 00:44:55.119 Um, they have not built the destination execution layers that we built for the past five years. 00:44:55.280 --> 00:45:06.159 So there's so many primitives that we get to inherit from our SaaS business for the last six years that are primitives that this harness uses, um, that are just easily available, that our engineers have worked five years on, right? 00:45:06.239 --> 00:45:07.760 So these primitives are easily available. 00:45:08.000 --> 00:45:18.960 Then the incremental 30 to 50 engineers that we've tasked with building the context, the execution, the content generation, and so on, they're not able to invest that kind of resource in building each of these things at once. 00:45:19.119 --> 00:45:25.519 They're focused on one agent to do one thing, such as JPEG generation, or one agent to do one thing, which is media buying on Facebook. 00:45:25.679 --> 00:45:26.800 But it's not the end-to-end. 00:45:27.440 --> 00:45:27.920 Got it. 00:45:28.320 --> 00:45:44.159 And I kind of find that like when agents can do the end-to-end, the power is not just like two X more, three X more, it's like 100x more, because it has context on your customers in Japan are interested in these things because the ads in Japan are performing better about these images. 00:45:44.320 --> 00:45:44.480 Yeah. 00:45:44.719 --> 00:45:51.599 So I should tell the whole rest of the agentic and human marketing team that this is working, and then they'll execute on this. 00:45:52.159 --> 00:45:52.960 Got it. 00:45:53.920 --> 00:46:03.440 So for a company to implement high touch across all its geographies, how much time and to see its results across all its geographies? 00:46:03.679 --> 00:46:06.800 US, Japan, Asia, how much time does it take? 00:46:07.199 --> 00:46:09.440 Um, it will probably take a month per geography. 00:46:09.519 --> 00:46:12.639 Um you can do in parallel or you can do serially, so it's up to you. 00:46:12.800 --> 00:46:14.960 And we do give you a forward deployed team. 00:46:15.199 --> 00:46:27.599 So especially for creative assembly, um, we have like forward deployed designers actually that will sit with your designers and onboard them onto this platform, and forward deployed marketers that will sit with your marketers and onboard them as well. 00:46:27.840 --> 00:46:32.079 So, but I would say like month is enough for any single geography. 00:46:32.400 --> 00:46:36.159 And and what what is the stack that you usually replace during this? 00:46:36.559 --> 00:46:37.679 We don't replace staff. 00:46:37.920 --> 00:46:38.559 No, stack. 00:46:39.039 --> 00:46:39.679 Oh, stack. 00:46:39.760 --> 00:46:40.320 Okay, see. 00:46:40.480 --> 00:46:43.679 Um I don't know. 00:46:43.840 --> 00:46:47.599 So like we can offer we can replace the CDP. 00:46:47.760 --> 00:46:51.679 We can replace many parts of Adobe and like journey optimization and so on. 00:46:52.000 --> 00:46:57.039 But in this like marketing harness world, it's really not about replacing a stack. 00:46:57.280 --> 00:47:04.480 It's really about taking your stack and suddenly like superpowering that with significantly more personalization, more campaigns, more content, more data. 00:47:05.599 --> 00:47:11.840 So we're really thinking like, it's like if you ask like me, like which part of the

stack does cloud code replace, it doesn't. 00:47:11.920 --> 00:47:13.679 It's just enabling significant more. 00:47:13.760 --> 00:47:18.320 And now there's like 10x more PRs on GitHub, 10x more bugs in big GitHub, and like so on. 00:47:18.480 --> 00:47:21.199 But but cloud code has completely replaced the IDE. 00:47:22.320 --> 00:47:22.880 Yeah. 00:47:23.119 --> 00:47:26.480 But like I don't know how much you even pay for ID anyways, it doesn't matter. 00:47:27.199 --> 00:47:30.639 So, you know, we are towards the conclusion of the podcast, right? 00:47:30.719 --> 00:47:32.800 You know, kudos of the journey of building high tech. 00:47:33.199 --> 00:47:35.280 Like it seems like you're just getting started now. 00:47:36.000 --> 00:47:36.239 Yeah. 00:47:36.719 --> 00:47:37.840 It feels that way too. 00:47:38.159 --> 00:47:47.519 Uh and uh if you have to summarize, what are some of the decisions that you know you as a founder's or team made, right? 00:47:47.679 --> 00:47:50.480 That made high-tech today where it is. 00:47:51.199 --> 00:47:55.280 Um we went enterprise really early, which everyone said not to do. 00:47:55.440 --> 00:47:58.320 We went multi-product very early, which everyone said not to do. 00:47:58.480 --> 00:48:01.440 Uh, we hired salespeople very early, which everyone said not to do. 00:48:01.599 --> 00:48:06.480 Uh, we made all these controversial decisions that were wrong for most startups, but correct for high touch. 00:48:06.639 --> 00:48:10.480 And so every single year we ask ourselves: what is best for high touch? 00:48:10.719 --> 00:48:19.760 And like, how do we think first principles about the needs of high touch customers and high touch employees rather than like the standard operating playbook of a startup? 00:48:19.920 --> 00:48:21.199 And I think those things have really helped us. 00:48:21.519 --> 00:48:24.880 And why do you think they work for high touch and they don't work for some of the other startups? 00:48:25.199 --> 00:48:26.880 Each individual thing has its own reason. 00:48:27.119 --> 00:48:34.239 So most companies, if they go multi-product too soon, they will lose focus, nothing will work, and each thing will be mediocre. 00:48:34.400 --> 00:48:41.199 Um, for us, we knew that like A, the products were so linked together, and B, that the reverse ETL product would become commodity. 00:48:41.360 --> 00:48:46.320 So we needed to build the next level of the platform to actually like really increase revenue growth in our own metrics. 00:48:46.559 --> 00:48:50.159 So it's just these decisions that are controversial. 00:48:50.320 --> 00:48:53.199 Like you could say that like Docker should have made some of these decisions. 00:48:53.360 --> 00:49:00.239 They should have focused on enterprise, they should have focused on like what comes after the initial product and so on. 00:49:00.480 --> 00:49:03.519 So some of them have parallels, but like many of them don't. 00:49:03.679 --> 00:49:08.800 Slack did the right thing by

focusing on one product forever and not caring about enterprise versus SMB. 00:49:09.599 --> 00:49:16.800 And and uh why enterprise was the right choice for you, though it would have, I assume, taken many months to close your first pain enterprise customer. 00:49:17.039 --> 00:49:17.599 Yeah. 00:49:17.840 --> 00:49:22.400 Um, because we had product market fit in enterprise, and we did not have product market fit in SMB. 00:49:22.559 --> 00:49:24.639 How do you define product market fit for yourself? 00:49:24.960 --> 00:49:38.559 Um the ability for customers, well, basically us able to provide way more value than alternatives in enterprise versus SMB. 00:49:38.719 --> 00:49:43.440 In SMB, A, we can't provide value because we're not even fit as an architecture. 00:49:44.079 --> 00:49:50.159 B, like even if we can be fit as an architecture, the value is like minimally better than alternatives. 00:49:50.320 --> 00:49:55.280 In enterprise, the value is significantly better than alternatives, and the alternatives don't even work. 00:49:55.440 --> 00:49:58.719 So it's just great product market fit versus like mediocre. 00:49:59.360 --> 00:50:10.000 So in your journey, you know, as a CEO from 0 to 1, 1 to 10, 10 to 100, and 100 to X, what are the things as frameworks that help you scale? 00:50:10.880 --> 00:50:13.119 Mostly just, I don't know. 00:50:13.199 --> 00:50:15.599 I don't really have any frameworks that helped us scale. 00:50:15.760 --> 00:50:19.679 Um in the early days, we really believed in marketing driven growth. 00:50:19.840 --> 00:50:22.800 So we thought like hiring sales is too slow. 00:50:22.960 --> 00:50:26.800 Um we want to hire salespeople to close deals, but we want the pipeline to come inbound. 00:50:26.960 --> 00:50:32.159 And so we really focused on like what can we build and like do marketing around so the pipeline could come inbound. 00:50:32.239 --> 00:50:35.039 And we didn't settle for like triple, triple, double, double. 00:50:35.119 --> 00:50:38.559 Like we went from two to 10 and then 10 to 20 in revenue. 00:50:38.800 --> 00:50:42.880 And like we just really focused on like how do we get as many logos as possible? 00:50:42.960 --> 00:50:46.800 Don't worry about charging them crazy amounts, um, just focus on land. 00:50:47.199 --> 00:50:51.119 Um, then we switched over to focus more on revenue and like providing value. 00:50:51.360 --> 00:50:55.280 Now we're focused on actually with AI, there's so much more impact that we can make. 00:50:55.360 --> 00:51:04.079 So let's actually build this modular multi-product company and do the end-to-end stack for marketing teams, all with AI, rather than like some pieces of the stack. 00:51:04.320 --> 00:51:14.480 And so each decision and each like um each decision is made stage appropriate based on like what opportunity you see that year and how do you plan for the next opportunity. 00:51:14.639 --> 00:51:20.960 But every year, like the reason multi-product, if it doesn't work this year, it doesn't matter that much, but it needs to produce next year. 00:51:21.199 --> 00:51:23.679 So that's why we start doing it like six months ago. 00:51:23.840 --> 00:51:24.000 Right. 00:51:24.079 --> 00:51:31.519 So like I think like a lot of the founders that don't scale, they're always firefighting the needs of today versus building the company of tomorrow. 00:51:31.679 --> 00:51:39.280 And so I really try to make sure that we're building the company of tomorrow versus um night and day living in the business of today. 00:51:39.760 --> 00:51:45.840 So then uh how do you prioritize meeting the the quarterly targets and everything? 00:51:46.079 --> 00:51:47.679 Yeah, that's what the exec team is for. 00:51:47.920 --> 00:51:57.840 The exec team's job is to meet the quarterly targets, and then our job is to make sure that we are working or they are also working on the targets for next year. 00:51:58.719 --> 00:52:01.519 So then the product reports to you or your co-founder? 00:52:01.840 --> 00:52:06.880 Yeah, so um we kind of share all the responsibilities, but it's me and two other co-founders. 00:52:07.039 --> 00:52:14.960 Um, and then the reporting structure is also kind of flexible, but all of product and marketing reports to him, and then like revenue and finance reports to me. 00:52:15.280 --> 00:52:15.760 Got it. 00:52:17.519 --> 00:52:20.639 And and where do you spend 80% of your time on? 00:52:22.079 --> 00:52:23.599 Um, it really varies. 00:52:23.760 --> 00:52:27.679 I think of like every month or every almost like two weeks as a sprint. 00:52:27.840 --> 00:52:30.960 So like last two weeks I might have spent on headcount planning. 00:52:31.039 --> 00:52:35.199 Um, the previous two weeks I would have spent on how do I make the ads platform scale quickly? 00:52:35.360 --> 00:52:38.079 And are we like ready to scale it? 00:52:38.239 --> 00:52:41.360 Um the previous two weeks I might have spent on something completely different. 00:52:41.599 --> 00:52:45.760 But I just don't think like I think it's like much more like the Jensen type operating model. 00:52:45.840 --> 00:52:48.559 Like do one thing deeply and then get out, and then do one thing deeply and get out. 00:52:48.719 --> 00:52:48.960 Got it. 00:52:49.119 --> 00:52:50.239 More so than like every day. 00:52:50.639 --> 00:52:51.039 Got it. 00:52:51.119 --> 00:52:54.559 So, so some uh, you know, the last few questions are rapid fire. 00:52:54.719 --> 00:52:55.119 Cool. 00:52:55.360 --> 00:52:59.840 So so one AI tool that has 10x your productivity. 00:53:00.400 --> 00:53:02.159 Um definitely just cloud. 00:53:02.639 --> 00:53:07.679 I'm not very good at using AI outside of like Cloud and oh actually OpenCloud, OpenCloud, Cloud, and Glean. 00:53:07.920 --> 00:53:08.239 Yeah. 00:53:08.719 --> 00:53:13.199 Uh one uh one interesting GTM hack that has worked for you. 00:53:14.960 --> 00:53:16.559 I don't have any hacks. 00:53:17.519 --> 00:53:22.239 Back in the day, we used to do like a

bunch of like scripting on like LinkedIn, but that's like a long time ago. 00:53:23.039 --> 00:53:28.719 What has helped you build uh the largest pipeline at every stage of the company? 00:53:29.199 --> 00:53:38.159 Um just being like in the details with the salespeople and like the sales team, I mean, being like really in the details of like which customers are they talking to, what are those customers saying? 00:53:38.320 --> 00:53:44.079 If you're in the customer conversation, then you can like do the right marketing, do the right pitch, build the right product, and so on. 00:53:45.679 --> 00:53:52.559 Uh one founder, operator, or investor that has changed the trajectory of the company. 00:53:52.960 --> 00:53:54.239 Oh, interesting. 00:53:55.119 --> 00:54:02.800 Um well, maybe one example is um there's this YC partner named JJ, JJ Fliegelman. 00:54:03.119 --> 00:54:08.480 And he like in the very early days, he kind of gave us the idea for reverse ETL. 00:54:08.719 --> 00:54:14.079 Like we were working on um something adjacent, and he was like, guys, like, this is never gonna work. 00:54:14.239 --> 00:54:17.679 Just sync the data to Salesforce, and here's how you could do it. 00:54:17.920 --> 00:54:19.519 And we were like, why would you pay money for that? 00:54:19.599 --> 00:54:20.800 He's like, I don't want to build it. 00:54:20.960 --> 00:54:22.960 And I have like great engineers, I'm never gonna build it. 00:54:23.039 --> 00:54:25.199 He was the founder of company at the time, now he's a YC partner. 00:54:25.519 --> 00:54:28.320 And he was like, just build it, I'll just pay you 24K. 00:54:28.719 --> 00:54:33.119 And like, this is just gonna sell like hotcakes because everyone needs this. 00:54:33.440 --> 00:54:37.360 Gotta one book or podcast that has helped you a lot. 00:54:38.480 --> 00:54:42.800 Um, is really lame, but the qualified sales leader, it's by John McMahon. 00:54:42.880 --> 00:54:45.280 He's like the classical like Bible for sales. 00:54:45.519 --> 00:54:53.599 I think it's just like still really accurate on the things that like a founder CEO should know in order to operate a sales team. 00:54:53.840 --> 00:54:57.280 And and what message in the book resonated with you the most? 00:54:58.159 --> 00:54:59.679 I can't paint any one thing. 00:54:59.840 --> 00:55:08.480 Like, I think like the best thing I have, I don't know if it's from the book or just in general, is like you, as the CEO, need to do the pipeline generation with the team. 00:55:08.719 --> 00:55:13.039 Everyone thinks I'll hire SDR team or marketing team or like a CRO and they'll just do it all. 00:55:13.280 --> 00:55:17.199 If you're not talking to the customer and sourcing the leads with the team, then you're not gonna figure it out. 00:55:17.440 --> 00:55:20.079 So you sit on the pipeline generation every day? 00:55:20.400 --> 00:55:23.840 I was sitting on it every day for the last like four years. 00:55:24.079 --> 00:55:25.360 I stopped doing it this year. 00:55:25.519 --> 00:55:25.840 Okay. 00:55:26.159 --> 00:55:34.320 But like before that, uh yeah, and I would hand select every patch for every rep in the company. 00:55:34.719 --> 00:55:36.079 Every pot for every rep? 00:55:36.480 --> 00:55:41.039 Every um account select like the um territory for each rep, I would just hand select. 00:55:41.199 --> 00:55:44.960 So which um customers or prospects, sorry, can this rep reach out to? 00:55:45.199 --> 00:55:47.119 I would just choose with the rep together. 00:55:48.000 --> 00:55:48.639 Got it. 00:55:48.880 --> 00:55:49.679 Well, thank you so much. 00:55:49.760 --> 00:55:51.360 I loved doing this podcast with you. 00:55:51.599 --> 00:55:52.079 Thank you as well. 00:55:52.320 --> 00:55:56.079 I clearly see like uh high touch going to a ten billion dollar company. 00:55:56.159 --> 00:55:56.480 Very easy. 00:55:56.719 --> 00:55:57.119 I hope so. 00:55:57.280 --> 00:55:57.519 Yeah. 00:55:58.079 --> 00:55:59.679 It's been such a pleasure hosting the podcast. 00:56:00.159 --> 00:56:00.559 Thank you.