n8n CEO on The AI Product Rule That Grew n8n to $100M ARR | Jan Oberhauser

Carlos González de Villaumbrosia 0:00 I just went to your website, and you have this big button that shows your GitHub repo with almost 200,000 stars, and it's right there next to Sign Up.

Jan Oberhauser 0:09 Obviously, we prefer if people use our hosted solution and actually pay us because obviously it allows us to reinvest the money. But at the same time, like, if people use us for free, that's totally fine as well. Like, that's just two different ways of using n8n.

Carlos González de Villaumbrosia 0:19 Jan Oberhauser, CEO of n8n.

Jan Oberhauser 0:21 What made it special is, like, very early on, we kind of really focused on the community very heavily. Created an, um, an amazing platform, but it also really ensured those people had a great experience in there, like they're supporting each other.

Carlos González de Villaumbrosia 0:31 And how is it possible for people to train their own agents as they execute some of these workflows?

Jan Oberhauser 0:35 That is very powerful. You can build very complex systems there. You can just say, "Here, something happens there," then actually send it to do something with that output, send it to another agent. Then they can use it again as an input, like people can build really, really complex, uh, use cases.

Carlos González de Villaumbrosia 0:48 And how are you seeing the companies that are more advanced kind of take that AI option metric a little further down the line?

Jan Oberhauser 0:53 AI cannot do anything it's not supposed to be doing.

Carlos González de Villaumbrosia 0:59 Hey, this is Carlos, CEO at Product School and your host on The Product Podcast. My guest today is Jan Oberhauser, the founder of n8n, the open-source automation platform that's crossed $100 million in ARR at a $5.2 billion valuation with nearly 200,000 GitHub stars. In this episode, here's what we cover: The fair-code license bet. Why he rejected traditional open source and still won. Why sprinkling AI on top kills products and what to do instead. How n8n beat Zapier and Make by giving the code away. Who should actually build automations? Hint: Not the product team. The metric he uses to know if an AI feature actually works. How one company routes 75% of support through an n8n agent with happier customers. Let's get into it. Welcome to The Product Podcast, Jan.

Jan Oberhauser 1:43 Uh, thank you for having me. Great to be here.

Carlos González de Villaumbrosia 1:46 I have to start with this. Why did you call your company n8n, man?

Jan Oberhauser 1:51 It caused a lot of confusion out there, and the simple reason is I'm a solo founder, so then, like, I started and then I started to program, and then knew it, before I launch it, I have to kind of... I need a name for the product, but I also knew it's not gonna make or break the company. So at first said, okay, like I find a name, I kind of time box like an hour, and then I just, uh, started. I wrote down like all of the names I, I liked, and all of the ones I liked they were already taken. As you know, all of the good domains are always gone. So one name was left, which was Nodemation. The idea was there's kind of, it's a node-based system and automation combined. But the name was quite long. It didn't really look nice or exciting, so I thought, "Hey, maybe I can make it more exciting." So I thought I'd do the same like Kubernetes, K8S, N, then eight letters N. It's, it's the same thing actually, and then n8n stuck. Um, n8n.io was available. After the seed round, I could buy n8n.com for 5K, uh, as well. We don't use it right now, but at least we, we own it to make sure it doesn't get more expensive. [chuckles] Uh, and now everybody has to say that strange tongue breaker, and there are so many different ways of saying n8n, which I didn't even imagine, like, from n, like Nation, Natan, Nathan. We have N, acht, N in German. Like, there's like a million different ways. It's quite exciting to see, um, [chuckles] what people come up with.

Carlos González de Villaumbrosia 3:01 Hey, you made it work. Like, people recognize the name. It's obviously not, not very common, but still, you know, like the product works. People will find a way to remember it. You started this, this journey over seven years ago, right? Way before LLMs became mainstream. So I'm very curious to... But just recently I start hearing about you maybe two, three years ago. So I'm very curious to know kind of how did you see this opportunity of LLMs and decide to pivot into that.

Jan Oberhauser 3:23 As you said, we started pre-AI, and honestly, I didn't imagine at all we're gonna ever do anything with AI at all. Um, but then obviously ChatGPT came along. We saw like the large language models getting kind of strong and strong and better, and so we just were wondering like, like how do we ensure that, like n8n isn't just part of the, the, like, there's not just a future for n8n, but we actually participate, um, in this amazing AI wave, which we kind of realized will be the future. So we kind of took a step back, and we just looked at what are we doing, what are our competitors doing? And all of them, uh, including ourselves, honestly, we just sprinkled some AI on top. Then, then we thought, okay, like to really kind of make use of it, really make sure we, we are part of the value chain, that actually people are not just using AI with n8n, but actually people building AI product applications, agents with n8n. Um, and that's when we then realized that we actually, n8n fits in perfectly. Like, what do these kind of large language models actually need? They need a lot of data from different sources. Like, no matter if it's, you're from Google Drive, is it from your local files, is it from Salesforce or whatever, it does something with the data, it transforms it, it makes a decision, and then it outputs this data again, um, again into a system or writes it to a file again and so on. And we just realized we had all of those pieces already in place. But the in-between piece, like this agents, that has had to be done like externally or like just very limited via HTTP request nodes. We just realized by really kind of adding these kind of capabilities like in front and center of n8n, we really kind of are able to kind of create something really amazing and powerful and just to not just, again, of allow people to build simply automations but really build really powerful agents as well.

Carlos González de Villaumbrosia 4:51 Obviously, that vision makes sense in retrospect, right? Because we're all using agents and understand now the concept of agentic workflows. But back in the day, that wasn't obvious at all, and, um, there was also a lot of competition. There were so many other platforms that maybe were more well known, at least in, in the US. And you guys are... You, you, you announced you crossed over $40 million in annual recurring revenue. You had massive funding rounds at, uh, latest one at $2.5 billion valuation with participation from Nvidia. So there is definitely something unique there. So I'm curious to know if, what was that kind of unique thing or secret sauce that helped you be on the map with this incredible traction?

Jan Oberhauser 5:30 Sure. Maybe first to, to clarify, we're actually over €100 million ARR a-already, and we are v-valued at five, $5.2 billion, uh, right now with, uh, kind of our announcing our partnership with SAP. So we're growing still nice. We don't ex- uh, announce, uh, the, the, the exact revenue numbers, but just people, we, we crossed 100 million ARR a while ago. And sorry, and, and the question was like what, what enabled it, this explosive global growth?

Carlos González de Villaumbrosia 5:52 Yeah. By the way, I love to be wrong about those things. Like, oh my God, now it's over $100 million ARR. So there are so many- Tools out there that were trying to do something similar. In fact, you weren't the first mover, right? There's so many other platforms like Zapier and Make and others that were-- They already had, had more market share. But you kind of came out of nowhere in the last two, three years and achieved this incredible traction. So I'm curious to know, in early days, what helped to get that initial tailwind?

Jan Oberhauser 6:18 I think, like, from the very beginning, n8n, n8n was all about, like power and flexibility, like I think that's what we kind of missed with a lot of the existing platforms. Just kind of think a lot of the existing platforms, you kind of got something built very fast, but if you actually wanted to get a bit and bring it in production, or you actually kind of wanted to scale it, you kind of, this, this platform really max out. That's why we kind of made the decision to kind of really focus on power and flexibility and kind of really, um, ensure people don't outgrow our solution. Um, and I think what, what made especially is like this focus, but also our focus on a community. Very early on, kind of really focused on the community very heavily and kind of created an, uh, an amazing platform, but it also really ensured those people had a great experience in there. Like they support each other, we supported them. We provided, um, a great, great support. We, we cr- we had, uh, online events. Now we, we do offline events. But it is just making sure, like putting those builders, which is really at the center of n8n, n8n is about empowering people. And even our kind of, one of our core company values is we are builders. Like, like we are builders, our community are builders, our users are builders. I think like having like this, this kind of, this alignment be-between, um, everything kind of really, uh, ensures that you kind of attract the right people, um, and it really kind of provided value. And I think especially then in combination, I think we grew very nicely the first years, but when it really took off is when we then focused on AI, as you mentioned before, where I think there were a lot of other platforms out there, but I think none of the other platforms kind of really had like this kind of very deep, very, um, powerful A- um, AI capabilities like we had them. Like all of them had like some AI integration, so you could kind of call out to, I don't know, an OpenAI or anything like that. But like, what we really, uh, did from the very beginning is kind of really ensuring that it's not just building, like calling out to, to, to an API somewhere, but you actually build like proper agents. Like agents, anything from memory. You can have like multiple models. You can have, um, different tools. You can configure them very easily. You can add scripts. Um, and kind of this kind of power and flexibility that was nowhere else available, and I think that's what people really craved. In combination, obviously, of other advantages like, um, for example, you can self-host us. Like people really appreciate that, that they know where their data actually lies, knowing they can run it behind a firewall. They can scale it. Um, they, they like, we have a big focus on, on reliability, like reliability, scalability and, and auditability. I think where you can just very easily understand what's actually going on there, even in more complex flows.

Carlos González de Villaumbrosia 8:39 I want to unpack some of the concepts you mentioned, because you're right. I mean, obviously you, you built a visual layer that makes it easier for people, you call it AI builders, to, to create this type of agentic workflows. So, and, and I think that expanded the market because this is not just a technical audience. Technically, anyone could be a builder. But I think you also mentioned something that I think is quite unique compared to other platforms, which is you allow the possibility to self-host your own product. What was the rationale behind making that decision?

Jan Oberhauser 9:06 I think the, it was very core incentive for the simple reason, like when I started n8n, n8n, I wanted to ensure multiple things. First, I really wanted to make sure that we have the, the possibility to become like the default tool out there. And if you want to become the default tool, you have to obviously make sure a lot of people get access to it and get access to it for free. Um, and so I think that's why I always, especially like a lot of, you know, I think like, I think that's, that's really a requirement to really kind of really get to that. The other thing is like, I saw a lot of open source tools out there, and I saw that they were able to build these amazing communities that really kind of made the, the success and this kind of, um, usage even, uh, uh, possible in the first place. So I think it was very clear I want to make, um, it available for free. And again, open source was one of the, one of the directions to actually do that. I went down a slightly different path very early on, where I saw that a lot of the open source companies actually decided to kind of change the license and kind of change it to something more restrictive. And I saw that, um, the community obviously got very angry about it. Less because of the license it shows, mainly because the, actually the company suddenly changed the rules and said, "Hey, first I, I can do all of this," and suddenly take, took like some rights away, and that really pissed people off a lot. And I thought, hey, I, you cannot afford it. Like building up a, a successful community is very hard. It's also like certainly very easy to lose it. So I thought, hey, I'm gonna be very upfront about it from the very beginning. So I said, "Hey, I like make it available for free for everybody because I'm a good person. I actually want to build a business. I want to kind of pay my own salary at some point. I want to hire a lot of people that kind of provide value." And I wanted to ensure that this money actually goes back in the company because I knew we kind of reinvest it into our support, into our docs, in our product, in, in doing quite great events. Only like this we can really kind of keep on growing and kind of make, uh, build a successful company. So from the very beginning, I chose the slit different license, which is now we call like a fair-code license. So we say you can do more or less the same thing as you can do with open source. The only limitation you have is that you cannot commercialize the code. So you can use it commercially, so you can use it in your company at any kind of scale you want, but you cannot build, for example, a hosted version of n8n and then charge other people for that. Which we thought again is, is a quite fair ask. That's why it's called a fair-code license. And I think that's, um, we did a-- So from the b-beginning, we thought about, hey, how, how can we enable the scale? How can we ensure we build a very great community, can really build up the trust? Because I think trust is really at the center of a community. What, what's really important is like people can say a lot of things, and they can also change their mind tomorrow, but only if you kind of are very consistent and from the very beginning people say, see that the things you say, you actually do, and you, you could, not just saying one thing today and another thing tomorrow, you can really build up this very strong community. So it was very important to kind of really show that from the beginning. And I think over the last seven, eight years, people can definitely see that.

Carlos González de Villaumbrosia 11:51 I mean, I just went to your website, and you have this big button that shows your GitHub repo with almost 200,000 stars, and it's right there next to Sign Up. So it's a really, I would say, unique approach to, first of all, pursue this open source path with a self-hosting option and still finding a business model that enables you to run a business while also making your product widely accessible to, to people. I've seen that approach in other, in, in other companies, uh, but not in, in your own specific space. Like we hosted the SVP of product from, which is this database comp- uh, well, uh, Fivetran just acquired dbt, and, and we, we had them here, and then MongoDB was the other example. But, uh, I think you are kind of the first mover in that automation. And by the way, how do you define your industry? Because I've seen terms floating around, automation, orchestration, now loop engineering.

Jan Oberhauser 12:41 Yeah. Uh, like we say, we are more like the orchestration layer. So I think orchestration is probably how I would, like an orchestration platform is how I would, uh, define it. That's, um, how I would call it. But also back to the previous point, I, I think like a lot of open source products start as, as, as open source and they kind of are very proud about it and at some point they start to hide it. The source code is still there, but they kind of try to hide it more and more and kind of get people, try to get people down the paid path. And we very deliberately say, "Hey, like we honestly, like obviously we prefer if people use our hosted solution and actually pay us because obviously it allows us to reinvest the money." But at the same time, like if people use us for free, that's totally fine as well. Like, that's just two different ways of using n8n. And the, the free one, like the free one is a very important, it's actually super important for us, like how we got adoption of all of these large organizations that we get used from anywhere from, I don't know, Meta, Nvidia, Dell, Accenture, in Germany, Vodafone, like Vodafone, Deutsche Telekom, Mercedes, like all of those companies and we got adoption the same way as by people having a problem starting with our free version and bringing that into this enterprise organizations. I think that is what, what people underestimate. They think very short term, think, "Hey, how can I ensure that this person pays me tomorrow for my paid hosted offering?" Versus thinking long term as making, ensuring that these people have a very great experience to get a lot of value out of the product. And when they get a lot of value out of the product, at some point they're gonna become a, a paying customer and even if not, like most ones will. So I think like again, back to the adoption piece.

Carlos González de Villaumbrosia 14:04 Even if someone chooses the free option, I imagine you start connecting to different LLMs and I know that you allow the possibilities to switch different LLMs.

Jan Oberhauser 14:12 Mm-hmm.

Carlos González de Villaumbrosia 14:13 So I guess there are open source LLMs-

Jan Oberhauser 14:15 Mm-hmm

Carlos González de Villaumbrosia 14:15 ... non-open source LLMs. So how do you go about enabling those, the, that optionality and still allowing people to, you know, pay when they have reached a certain level of usage?

Jan Oberhauser 14:24 Like in the end, what, what make us different with n8n and to, to most other tools is like people bring their own key normally. Like we, we have some functionality inside that you can get some, some test credits and we are also working on a feature that you can kind of, kind of use these LLMs via n8n and kind of to, to also pay with them. But un- up to this point in time, people really bring their own key. So they say, "Hey, I want to use an Anthropic model, I want to use an OpenAI model, or I want to use something else via, I don't know, Hugging Face." And they bring their own key and connect it. So again, depending on the use case, depending on, on how much money they have available, they kind of just decide by themself which, which one they want to choose. But I think the very important thing and the, the power of n8n is also, again, you can switch later on. Like you can switch like it's prob- like theoretically it's a switch of, of a second just with deleting one node and putting in another one. But obviously, but at the same time you still have to make some adjustments there where we also kind of have the f- functionality that people can create evaluations inside of n8n just to kind of test what does this change do and, and how do I have to adjust, for example, my prompts to make that possible. But again, it's, it's quite simple and much simpler than other platforms to actually change your model provider there from one to another, especially if it's, if you change, for example, I don't know, use Open- an OpenAI model by OpenAI or then hosted by Microsoft tomorrow. That's literally a switch of a few seconds.

Carlos González de Villaumbrosia 15:36 Yeah. I mean there's all this chatter now, right, with the Chinese, uh, open source models are, uh, performing apparently better than some of the US-based models. So I think clearly having the optionality for the user is an advantage. It allows you to not be locked in with any of the providers and ultimately you can see, uh, in the news a lot of the time when, um, or Anthropic or OpenAI releases a new, a new model, there's all these headlines, "This company just killed 10 other startups," right? So I'm curious to know from your perspective, what is your market positioning and how you're thinking about, uh, the modes, but also like the relationship that you, you have with these different LLM providers?

Jan Oberhauser 16:12 I think why people choose, like there's definitely a lot of killing going on and I think like they think killing apparently gets a lot of, um, clicks. That's why I think that that happens literally daily that something kills, kills another thing. So I would probably not take it too seriously. Most of the killed startups are, are doing still very well, obviously not all of them. So I think partly that there's s- something true there. But I think like what, what makes n8n special is like our focus on kind of really audibility, like flexibility, um, like the self-host ability and, and the security piece and kind of really giving people really control. Like you can normally, you can see more or less like you can say the LLM is kind of the engine and, and we are kind of the car and also kind of the, the rules and the kind of the traffic, the, the, the whole, the roads and, and the whole system. Like we kind of really connect everything and kind of we are the kind of connective tissue between all of them to, to kind of make use of this kind of pile of different models and agents and business systems and bring it into kind of a actually a process you can actually use. I think that is, um, it's like we, we, we all need each other. We, we need those, those, those model providers. Again, not a good thing is you can replace them, but again, we have a lot of models, we have a lot of different assistants you connect to and we kind of bring it all together and allow you not just to kind of build a nice demo out of it, but we, we have a, like a very big focus on actually building like reliable systems that are secure and that you can actually understand what's happening. So if you build an automation with n8n, you can know exactly how data flows that way. It's using those tools, it's not gonna do anything else. You see when something goes wrong what, what exactly went wrong and I think that is again why especially large organizations choose us for business critical use cases because they can already get real ROI. They actually can deploy us in production, not just for a nice demo out there.

Carlos González de Villaumbrosia 17:51 I think we've, we've talked about the product enough. Now it's time to show it and I love that you were the, the one who volunteered to share the screen and, and demo what you got. So Jan, I would love to see your product in action.

Jan Oberhauser 18:02 Sure. In the, in the past people built workflows themself manually, but obviously with, with things like Claude Code and Claude Code Work, people realized, hey, actually building can be much simpler and we also obviously realize it as well. We obviously want to give the great, same great experience to people inside of n8n and also just allow literally anybody to build. Like I think a lot of feedback we always received is that n8n was a kind of a more complex system focused on more technical users and the goal was always to kind of lower the bar and thanks to AI we're now finally able to kind of really empower everybody to build. So what you can see here is our new, uh, AI assistant is we just released it, um, one and a half weeks ago. But what you can do more or less here, you kind of de- describe what you want to do. Like we give you some examples. You can say, hey, I have a marketing use case, um, SEO and then you can just Kind of define what, what it should build out for you. I already prepared something because it takes, um, definitely a few minutes to actually build something. So like here we have... Actually, I just make it a little bit smaller. You can see better and kind of scroll up. Here, I, I kind of just ask it to build me like a, a personal workflow agent, um, a kind of personal agent that kind of work with my emails, with my calendar. And it starts with a chat trigger, so I can actually talk with it. I told it what my email address is, and then I told it like all the things it should be able to do here. Um, and then I also made sure that, again, here you can see With Human in the Loop. That means, like, certain actions I want to ensure that it's not just a AI just does its stuff by itself. Like, I don't want it to send an email in my name without me actually confirming that. So I just told it, "Hey, please ensure that certain steps can actually not be done unless I approve them first." Um, so it then started building. It kind of, um, said it understood what it, uh, what's going on, and then it asked back a first question. It kind of wanted to know like, um, actually, where was the first question? Uh, here. It kind of asked like which model it should use because it can obviously, there's a lot of different models out there. Uh, where we then simply said, "Hey, um, we actually want..." I, I show it's on here. This is the, the first question. Davide said, for example, um, "Hey, please, um, use Claude through O-OpenRouter." Then it asked some other questions as well, and you can kind of see the answers down here as well, how it should work. Then it kind of kept, kept on building. Like, normally, you don't have to call about, care about most things here. But again, for the more technical users actually want to understand what's actually happening there, they can actually open up here and kind of see exactly what's the kind of the AI system it is doing and how it kind of starts to build that workflow. Then it was, was building a while, um, and at some point, it, it said it's done. And then I also said, "Hey, um, actually, the, the way you currently build it, um, I want to get, get it changed slightly." Then it kind of went in there and, and kind of made changes to the workflow. And then in the end, it was done. And you can now see here, you see like one of the artifacts which you can now see on the side is that the workflow that has been built. So you can go in here, and let's also make it a bit larger that you can see it better. Um, and that's the workflow that, that the system built by itself. You can see like it's, it's quite easily understandable. Like here, you, for example, you have like if you start with a chat node, uh, here the whole conversation starts. You input the data. It flows into this a-agent. This agent, by default, uses, um, kind of Claude Sonnet five. But we are also have it configured that as they are not, you know, Antropic is not always that reliable. [chuckles] We also have a fallback model. They can say, "Hey, if this model is currently not available, fall back to a GPT, uh, 5.4 via OpenRouter." And then you see here all of the different tools we had defined previously. You can set of list recent emails, um, or you can have like, um, it can get emails for you. It can even create draft and so on. And here we go. We see the whole human in the, in the loop steps. There you can say, "Hey, you can send an email, but it can only do that if I get approval before." We kind of can see the configuration of each of those nodes by going in there. So we can see here, like here we say, be-- for approval, it kind of shows this message, send this email. Please provide it before it goes out. It gives you like an overview of who it's gonna send it to, the subject, and the body. And so before it actually does that, it kind of does always the confirmation. We can see the same thing here as well for kind of sending a reply or for like creating calendar events. Mm-hmm.

Carlos González de Villaumbrosia 22:03 In this, in this case, so I'm seeing there are two different workflows, right? Ones that could be fully automated and ones that require a human approval before-

Jan Oberhauser 22:11 Mm-hmm

Carlos González de Villaumbrosia 22:11 ... they are executed.

Jan Oberhauser 22:12 Exactly. Like the, and like, and it, and can totally exist in the background. You can say, "Hey, every time a new lead gets created, um, in, in whatever system, create something in Salesforce for me," and it just does it totally automatically. You can also have like a, a, an internal chat system like this one I, I'm showing of here where you can just say, "Hey, I create a personal, personal, personal assistant for myself," which in this case kind of works through a chat. Like you can test it in here, or you can also use it, for example, via Slack or, you know, via, via Telegram. Um, or you can have even have stuff, something exposed to, to your users. We have, for example, a, a big, uh, online travel agency which kind of automated a lot of the support actually via an AI assistant built with n8n, where literally like seventy-five percent of the requests of the customers go through that agent where people can change, uh, hotel rooms, change bookings, can ask questions, and so on. Um, and again, the nice thing about that example is even that the people that use the AI assistant, like this, uh, AI customer represe-representative, people actually seve- like actually happier with, with that than with a human, um, um, assistant there, which is obviously amazing, which shows AI is not just about kind of lowering the cost and so on. It's actually providing also real use, uh, real value for the users. They can definitely-

Carlos González de Villaumbrosia 23:27 Yeah

Jan Oberhauser 23:27 ... again, anything from fully automated to literally everything needs approval. You can build anything.

Carlos González de Villaumbrosia 23:31 Got it. And just to clarify also on the, this visual graph was created automatically based on the configuration that you did before, right?

Jan Oberhauser 23:39 Exactly. That was built, uh, automatically. I mean, actually, we, we did actually, uh, uh, normally you have to kind of sometimes, um, select, um, for example, like which credentials to use. To give you an example, here we, for example, um, a Claude Sonnet. We can say here, "Use the n8n built-in credits," um, or you can say, "Hey, use my own, um, API key." Um, and you can do this, can, can do it with literally any tool here. You can say, "Hey, here, I, it's connected to those credentials of, of my Gmail account," and so on. So you can install, you can build it automatically, and then you can say, "Hey, I can actually now go through and, and make sure I can understand what actually has been built." Because I think the, the problem with a lot of if you write code, code, like you have normally ten thousand lines of code that you hope does the right thing, but actually kind of reading and understanding them is almost impossible. Because it takes you hon-honestly more time to partly understand it than actually to kind of create in the first place. And here it's just very simple. You can literally say here, here I have a chat node. I, I get a message. Here, for example, you can say, "Show me my last five emails." Then the data flows through here, and then you can see, ah, here now I have this AI assistant as AI agent. And you can see here now it says assistant message. That tells me, "Hey, I'm, I'm an assistant. I have those tools. I have like certain rules in place." Here I get the chat, uh, the chat message And then I call, I have access to those different tools, and you can very easily add more tools. Like there's, you have, you can even add code tools. You can have like, you can do actions in f- in, in different apps. Um, like you can see a lot of them. Anything, um, you can, you probably, your, your heart is asking for and, and you can set up very easily.

Carlos González de Villaumbrosia 25:04 Can I ask you a question about those, those tools? So I saw on your website you have over 500, uh, integrations with, with tools. And so how do you go about that? Is it something you have to set up one-to-one with each of those tools, or is it a more of automatic way for you to create integration?

Jan Oberhauser 25:19 Um, in the end, like we, there are d- different kind of i- integrations, um, in multiple levels. There are like certain built-in integrations that, that we created ourselves, we maintaining ourselves. This is for example, like all the most important ones. Like for example, j- anything with, with Google, for example, or Salesforce. We created them and we maintain them. Then there's also like custom int- uh, like community integrations. The community build them, and then we check them and make sure they're actually secure, and then people can also access them on their cloud solution as well or can ins- manually install them. There are actually a lot more than those 500 out there. And then even if there is no integration built by us and no integration built by the community, you can still kind of connect to the system manually via an HTTP request node. You can just say, make a GET or POST request to a certain endpoint. You can also ask the AI assistant to build it for you as well, to kind of configure the nodes for you. But in the end, you should be able to connect to literally any kind of, um, API-based system out there.

Carlos González de Villaumbrosia 26:12 Yeah. I remember the first time I set up my OpenClang agent, it was quite an ordeal. And I, and I'm relatively technical, but still there was no visual layer. So here I can imagine you can choose to drag and drop and, and set it up, and I guess you also don't need to buy a MacBook or a Mac Mini, right?

Jan Oberhauser 26:29 No, no. It's, it's running totally in the cloud. You can also just like test very easily and literally see what's g- see what's going on. If I say, say now for example, let's say I can say, "Send an email to james@," I don't know, "email.com saying hello." And then now I can literally see here how it's actually going through the data. The, the message got received, and now the agent starts building. It kind of adds data to the memory. It calls out to the model. Um, then it kind of tries to figure out what to do next. And then you can see that it's gonna call one of those tools that actually gonna answer. In this case, it said, "Hey, um, send out this email. Please review, review it before I send it out." So actually see it went here, and because we said, "Hey, there's actually a human in the loop," it actually stops now, and it kind of keeps on waiting till it actually says it should send it out. So I could now say, "Send it" or, uh, "Don't send it," "Send" or "Cancel," and you can see either it would go on or actually stops here. So like, again, you can, again, and if something would have gone wrong, you would see like a node is red here, and you can just very easily debug. Um, and again, now you can also just go in here and see, hey, what actually happened. You can see the logs. I got the input. I got the model. I, I sent this kind of input to it. Then it kind of, uh, called a tool with that kind of data. It kind of, um, kind of asked for approval. So you can literally, uh, go through the whole, uh, to, to everything that happened very easily and kind of debug and understand what if it actually does the right thing you want it to do. And you can do that for literally any execution that happens.

Carlos González de Villaumbrosia 27:52 So as you know, loops are hot now, right? This whole concept around now you can train your own agents or you can set a goal and they will self-improve. How do you think about that, and how is it possible for people to train their own agents as they execute some of these workflows?

Jan Oberhauser 28:06 Like in the end, what, what happens there is, is a kind of loop. Like it, you call the agent, it, it calls a tool, it goes back to the agent, uh, like to the model. It calls another tool. It can respond to the user and then can, can go back here again. You can also literally can, like, you can literally loop. Uh, like literally can just tell it to kind of talk to itself if, if I want to by just, um, adding this connection here. Um, and it then is very powerful. You can build very complex systems there. I can just say here, something happens there. Then actually send, do something with that output, send it to another agent. Then I can use it again as an input. Like people can build really, really complex, uh, use cases. And also like maybe actually let me share this one here. This our website. You can see here templates. And you can see there's like over 10,000 templates that our community created. So you can say, "Hey, I want to use something like a sales," i- in sales, for example. You can see here, um, nice things people built. Or you can say, "Hey, I want to have something with," I don't know, um, "Google Sheets." And you can see like a, a nice example how you kind of auto create TikTok videos or anything else out there. And then you can kind of click there, and you kind of can see here literally the actual, um, automation that actually has been built. And you kind of already explore it here, like the same way you can like, like just explore it. You can see here what's actually happening, how did it get set up. And then you can literally just say, "I want to use it." You can copy it, and then you can just literally copy it, um, in here. So let me just, um, start a new one, just create a n- new workflow. Now just literally just copy-paste, and you can see the whole thing is in there. And now I can kind of configure it with my own settings. For example, now here, in this case, it's, it has an OpenAI model, so I could say I can use my own credentials or use the n8n ones. So you can... Oh, sorry, I don't f- I just forgot. I forgot to share the right tab.

Carlos González de Villaumbrosia 29:49 No, I, I get it. I think this is, this tackles a very important problem we see in the product, which is the starting with a blank slate sometimes is hard.

Jan Oberhauser 29:56 Exactly.

Carlos González de Villaumbrosia 29:56 So having some sort of recipe that allows you to validate what you need and still have the ability to, to edit it, right? Based on the tool stack that you have or the specific use case that you need.

Jan Oberhauser 30:06 Exactly, yeah. And, um, maybe also to f-finish the, the previous example, I think like, um, I said, I said before, like we had our assistant here and, and it, again, now it build it out. Like and I have now this workflow here, but I can actually as well like, can actually now say, "Hey, that, it's great what it built here, but actually I want to, can extend it." So I can, for example, say now, "Do you have the existing workflow ready? But now add, for example, this like calendar tool where I want to book my one-on-ones with people. I want to have consistent format. They have like over thirty minutes," and so on. And now I can just set it to the agent. And again, now I can see like how it actually works. Now it goes away. It, it does some thinking, and then it kind of will start to extend the existing workflow. So you can say, hey, you can start very simple. And you kind of not, you don't have to kind of build and like go totally crazy and kind of come up with the whole solution from the beginning, but actually you can also do it in a iteratively. You can say, "Hey, start with the most basic version, and then I want to add this, this other things, um, in time as well." Now I can see now it's, it's working here, and now it kind of, uh, does some reasoning there. It kind of goes back and thinks about what it should actually be doing. Uh, it will probably gonna take like a few minutes. It's a good addition. I need a Google Contacts tool name. Um, again, it figures out kind of the different nodes it needs. It looks of if credentials already exist, for example, and it says, "Hey, I need Google Context credentials." It says, "Hey, I already have one existing one, so I can actually connect, um, to, to, to Google already." Um, then it will, um, hopefully be done in, in, in less than a minute. And then you can see that it extended to the existing workflow as well. And again, that's, uh, the amazing thing is that let, uh, what, what sh- want to show here is that again, literally everybody can do it. Like, because you have to know what you actually want to build, and if you know what you want to build, you just describe what you want to, uh, have done, and it goes out there and actually does it for you. And again, and then you can test again, literally test it by yourself, um, and, and, and ensure it, um, it, it actually works the right way, and then it can, um, very easily, um, deploy it in production and kind of use it on a daily basis.

Carlos González de Villaumbrosia 32:06 I want to ask you a question about the main bottleneck that I see these days with, uh, this type of assistant. So people are able to now create their own, they call it second brains or AI chief of staff or assistant, uh-

Jan Oberhauser 32:18 Mm-hmm

Carlos González de Villaumbrosia 32:18 ... that seems to work really well for their own productivity.

Jan Oberhauser 32:22 Mm-hmm.

Carlos González de Villaumbrosia 32:23 But the multiplayer bottleneck is real, right? Like, how do you now create a shared environment where people can benefit from the work that others have done, and they can really collaborate instead of creating their own versions of their own assistant?

Jan Oberhauser 32:35 And then, like we just, I think, not sure if you released it yet, but it should be closely with, uh, shortly before releasing, is where instead of kind of just creating your own, like connecting your own credentials, you can actually say, "Hey, assign these credentials dynamically." So in this case, if I interact with it, it would, um, use my credentials and would access my Gmail account and my calendar. If you would actually chat with it, you can a- give it access to your calendar, and it can do it for you. Think that is this very important piece where you can adjust, again, build every, you have, I don't know, 10,000 people in an organization, where you have 10,000 people building, um, something separately. You can actually have literally anybody just having one of them deployed, and then the whole organization gets access to the same thing as well. Again, generating workflow. I said, uh, that's fine that they can do that. And I think that's, that's this very important piece, um, as well.

Carlos González de Villaumbrosia 33:24 Totally. 'Cause ultimately, you become an super agent manager, right? Like, you have all of these different agents doing different things. Sometimes those agents are yours, sometimes it's your teammates. And so what ends up being the, the control panel, kind of a, the surface for the product leader to really understand what's going on and be able to make any type of modifications?

Jan Oberhauser 33:44 Yeah, I, I agree. That is super important. Also, by the way, here you can also already saw maybe very fast. It's not just building the workflows, actually kind of already kind of tested it for you as well, so they actually can be sure that when you build something, um, it, it kind of, um, it is already, it's in a, in a working state.

Carlos González de Villaumbrosia 33:59 Really cool. And the other thing that I, that I noticed it's, it's becoming challenging for especially larger organizations, is the deployment into production, right? Like this, this type of situ- the demos work very well when you have full control over the tools you're trying to connect to. But as more people join the party, then you also have more controls and checks, right? So if we're talking about large enterprises that already use n8n, like how were you able to get over all of those procurement and guardrails, uh, problems to make sure that they trust your solution?

Jan Oberhauser 34:30 First, like we talked about before, self-hostability. I think that's just an important piece already that people actually know where, where it actually is running. Um, and they can also very often, they kind of use their own local models, for example, to actually deploy n8n. Um, and then other thing, the other piece is like, literally, like if you build something with, with, with AI, like it can go rogue very easily. Again, for example, I don't know, it deletes data it shouldn't delete, or it sends out emails it doesn't wa- it shouldn't, uh, it shouldn't send. But again, because, uh, we make it very simple, like literally this agent can do only those things. It cannot go rogue. We have exactly specified all the things it can be doing. Again, also the things it can only be doing once actually a human approved. So it means like be- before those things can go in production, somebody else can very easily review it. Again, and they don't review, again, 10,000 lines of code. They actually review a workflow they can e- easily understand and go through and kind of know exactly what it's gonna be doing. I think that just makes it very easy for these enterprise organizations to not just, again, have a nice demo, but actually kind of release it in production much faster with other systems.

Carlos González de Villaumbrosia 35:31 And I think this is a very common challenge for a lot of, um, companies that maybe started more on a SMB segment or co- direct to consumer, people who are like really curious and are building something, and then they're bringing it to their enterprises. So in addition to just proving that you have the right security checks, like what does go-to-market motion look like? Because PLG can only take you so far.

Jan Oberhauser 35:56 First, like th- we got into the most organizations, like via bottom-up PLG, uh, growth. In- interestingly, and it's not just bottom up, we got a, in a, into a lot of organizations as well, literally top-down as well, where people on the board, where C-level, um, I don't know, the CTO or the CEO ever saw content around n8n and just was wondering like, why don't we do the same thing in our organization as well? And they kind of, um, bring it into organizations and then very often find that they're actually already using n8n, just in a more limited way, and then it just, just makes it much, much easier, um, to actually, um, do that. Um, at the same time, obviously like at some point you want to accelerate things. Um, another nice way to accelerating things is definitely partners, and I think that is one thing that also works very well for us. Like we have a lot of, we talked about the community before, like a lot of the community consists of people that have agencies, has built up agencies around n8n, and it really kind of not just, again, they go out there and say, "Hey, I could, I can actually help you to build this workflow. I can actually maintain it for you. Um, I help you to kind of even figure out what can, can be automated." They're kind of this, this accelerator. We also work with, um, for example, just a month ago, we announced a partnership with Accenture as well, and we talked with other system integrators as well, and there are other partnerships getting announced very soon, um, on, on top there. So like we're looking for this kind of accelerators that kind of really help us to kind of reach more of our user base and, and reach more of the opportunity, um, easier because we can obviously not have a one-to-one relationship with everybody. So we, we have to ensure to actually kind of use all the channels that are actually available to us.

Carlos González de Villaumbrosia 37:24 So as you partner with these, uh, companies, consulting firms and others as your implementation partners or FDEs, call it however you want, it seems like one of the challenges that these, uh, companies have, and by the way, we do that for a living as well, is like you get the, you do the initial sprints with a client. They, they see the magic. They love it. But at some point, you want them to be more self-sufficient. They also want to be more self-sufficient, right? So I'm, I'm thinking about the actual adoption curve, like maintenance. Like what is the approaches that you are seeing now for these type of large enterprises? Is it like having their own center of excellence or chief AI officer that is o- o- overseeing all AI implementations? Do you see more functional leaders trying to handle their own automations? Like what is your, your vision on that?

Jan Oberhauser 38:06 We, we definitely see All of it theoretically, like all of it is out there. Um, but what, like, just the, the thing we are doing internally as well, and which we, um, suggest for our partners as well, is not just not having it only centralized. Like, I think, like, you can definitely get... If you have a centralized team that builds automations or AI assistance or agents, uh, in whatever way for other people, like, you, you can reach, like, I think you get a, to a lot of value, but you don't really empower people. You build for them, and again, it's, it's, at some point it becomes unmaintainable because you suddenly, like, this, this, this department literally has to kind of maintain hundreds of thousands of, of automation that doesn't work. And it's also the opposite of why I created n8n. Like, the reason why I created n8n is, was because I saw that the people that had the problems were dependent on other people like myself, who actually helped them to kind of build things, to build automations or, or build agents. And n8n, the idea with n8n was that the people that have the problems can solve it themselves. Because again, that's why we bu- we try to make it simpler and simpler and simpler. So, like, this, again, uh, we have an, an AI and automation department internally as well, but what they're supposed to be doing is, like, still help, uh, like, the idea is about empowering people inside of the organization. Again, the people that have a problem should be empowered to build it themselves. They're still there for you. Say, "Hey, I can't, I don't know how to do this one thing." Then they help you out there, or they help you kind of try to figure out first what is actually possible to build. It kind of, trying to kind of, the idea is to kind of help the people to fish, because in the beginning, they have literally no idea what's actually possible. And then kind of getting, starting with something small and kind of empowering them more and more, that's how we really see, um, how we can really build it up successfully. And then obviously document what they did, kind of, uh, and, and show you have the kind of the right, the, the right tools in place. But it, we kind of see them as kind of creating the guardrails, kind of. They, they, they own n8n, they own, uh, everything around it, their own education, but the building should be owned by the people that actually have the problems because they, like, nobody else can do it better than them. They actually know the whole process by heart. And explaining it to somebody else is just super inefficient, it's never gonna be the same thing. If literally any problem you have, again, any simple change, you have to go back to another person again that has the same context and it's, it's just super inefficient.

Carlos González de Villaumbrosia 40:12 Yeah, and, and, uh, exactly, and that is why this thing is so complicated to measure ROI, right? Because first you, of course, need to understand what's possible, and once you taste that magic, it's hard to go back. So adoption seems to be a good way to get started. But at some point, people will start questioning, okay, but all of this productivity goes where, right? But at the same time, you can't fully give a, a one fits, uh, a one-size-fits-all answer to all the companies. And here's how you measure ROI of your AI implementation. So I'm curious to know what is your take on how you're seeing the companies that are more advanced kind of take that AI option metric a little further down the line.

Jan Oberhauser 40:50 I think there's, it's, I think it's definitely also something we struggle with as well. That's just thing, because there's so many, like I said, so many different ways to actually measure, like, the, the impact. Um, and again, like, we, we also started, uh, initially thought, hey, that the simplest thing, but you can also do very easily with n8n, and it's literally just track, like, how much time it actually saved. But I think that's actually not, not, not, not the best measure. We kind of realized it's also, like, very easily, you kind of, you can just cheat very easily, you know, kind of in certain automations. Like, even though they run every minute, it's not just, um... And, and it would take you 10 minutes to run it every time manually doesn't mean, like, it, it saves you crazy amounts of time there. So, like, w- what we actually now are looking, I think one thing is, is the ROI piece, but what we are kind of in, in, in, we, we actually, i- internally, we care less about the ROI kind of measuring it the right way because we just know there is, and, like, it, it, it has a lot of impact. What we actually want to drive forward and want to ensure is that actually change is happening. That actually every department actually thinks about, like, what can they be doing to kind of drive change forward? How can they make sure that they're actually not, that they think differently, how they kind of have more impact? And that's why each department has to define, like, their AI vision and kind of really kind of define the certain things they want to achieve every quarter. And then we as a company ensure that each, that each department kind of reaches those goals, and we as a whole company actually reach, um, at least 90% of those, uh, AI targets that have been set. I think, again, y- you can measure anything. I think, again, also, I think it's not one thing, one size fits all. Maybe for some of them, it makes sense to kind of track, again, of how much time is saved or how much money is saved, whatever. I think you have to decide it yourself. But I think honestly, the most important thing is right now, certain things you can take as a given. You know it provides the impact. I think you just have to kind of, I think you, you sort of want to make sure, like... I think it becomes then problematic when you have, like, on one side an unknown outcome, and on the other side you have a lot of, uh, high, very high spend. Because then it's, it's not very clear anymore. And I think that's again why people appreciate n8n very much, is because the cost of running an AI system built with n8n is much lower than in other systems, because we're not just saying, "Hey, AI is a solution for everything." We say, "AI is one piece of it." Again, what you want, you need AI, but at the same time, you still want to have deterministic logic. In this case, for example, the step approved those emails or, I don't know, do X or Y depending on a certain number. That's deterministic logic. Like, it doesn't need an AI to actually do that for you. Because AIs are expensive, as we just figured out. They're not very reliable, and they're very slow. Like, why would you do it with an AI? And then you still have to have the human in the loop, again, which again, en- ensures that you cannot just use it for a nice demo, but you actually can use it for business critical use cases because you know the AI cannot do anything it's not supposed to be doing. And I think that is again why the, the cost of, uh, running an AI system is actually in-built and, and is actually quite low. And then now n8n, the, the ROI, um, is very clear. Again, in the most simple form, start with just, uh, start to track time saved, and then you can actually go further and kind of literally kind of track your own things, like... I mean, n8n, you can even do it dynamically. You can just say, "Hey, I have this, this arbitrary metric," um, and you can, for each execution, save that number. Like, it would maybe get saved as time saved, but you can literally kind of use it for anything you want, just to make sure it has the kind of the impact what actually your organization is aiming for. And again, one last thing I want to say there is the example I gave previously with this kind of AI assistant for this online travel agency. I think people, like, think this ROI thinking is one thing, it's great. You ob- you want to have ROI, but very often, like, the advantage is not just money saved or anything like that. Very often it's also a better customer experience as well. Again, people prefer the AI assistant. Why? Because never gets tired. It can speak any language. It's, works 24/7. It has the perfect context. Like, that's great. And, yeah, and I think that is this other thing, is like certain things, it's not just normally one value. It's certain things you can drive forward. You can say, "Hey, maybe you don't want to... It's not about saving money, it's about having better customer experience, or it's about saving money." I don't know, like, whatever your, your, your company is driving for, what you should measure that one thing.

Carlos González de Villaumbrosia 44:46 I like that because sometimes with the introduction of new, new tools, it seems like we're forced to now coming up with 10 new metrics to measure the specific impact of that new tool into our business, while in reality it's all intertangled. That it's just hard to isolate a specific variable. So using The same business metrics that your business uses to measure productivity, outcomes, performance, ultimately, I think is the easiest and most reliable way to see if, okay, how you're getting there is actually helping or not. But, um, that's at least also how I try to measure the own impact of our own AI implementations in the business. Like, okay, are we seeing the NPS of the customer go up? Are we seeing revenue grow up faster? Then great. It seems like what we are doing in the house seems to work or not. But it's like if you go to a restaurant and order a chicken, if you like the taste of a chicken, who cares where that olive oil is coming from, if it's AI or non-AI.

Jan Oberhauser 45:41 Exactly, and I think, like, the thing is, like, you can even say very often it's maybe the right thing to actually spend more money because of AI, because, like, what you care about is, for example, speed. It's like if, uh, the difference if I have a human that does it and it takes him, for example, 15 seconds, I can do it with an AI which does it maybe in, in three seconds. Like, maybe that's worth paying five times that, but it really depends on your own business, um, and what, what is important for you.

Carlos González de Villaumbrosia 46:03 Jan, it's been a pleasure to have you on the podcast and see you demo, to you also explain the product, go-to-market strategy. Um, thank you so much for your time.

Jan Oberhauser 46:13 Thank you for having me. It was great. [outro music]