The Hidden Cost of AI Agents No One Talks About

I met a customer in New York City a couple of weeks ago. He says, "My boss tells me he needs agents and I just keep building make workflow automations and I tell them they're agents and he's very happy." So by empowering that AI agent that has a role and understanding what it should do and then giving it the right tools and very precisely focused on giving it scowples, not sledgehammers. We learn more about the capabilities of generative AI. We learned AI can actually reason and that's an interesting thought unto itself and if AI can reason it means it's not just part of the workflow but it's actually driving the workflow and ironically the more you throw at it the less successful it is. They have people who for whatever reasons in the past have shown themselves not capable of successfully taking care of a pet. So they do not adoptless, it's name a person, maybe an address, phone number, the business process is making sure that a person that comes in is not on that do not adoptless. Fuzzy matching is like a hard software problem to solve normally like is this is Tom Thomas and Thomas Tom like are these the same person? It's a baseline expectation when you come to work now. So imagine building hundreds of automation workflows across your company, sales, marketing, customer service and you're scared to change anything because you don't know what you might break. That's the reality for a lot of companies today with AI and automation and today we're going to talk to make about how to solve this exact problem. Welcome humans to the latest episode of the neuron podcast. Cory Knowles editor of the neuron joined us always by our daily writer Grant Harvey. How are you Grant? Doing good. Cory doing great actually. Really excited about this one. Nice. Nice. Well, today we're going to talk with Darren Patterson VP of market strategy at make the automation platform used by about 250,000 organizations worldwide. Make just like two awesome new products AI agents that can make autonomous decisions in your workflows and make grid the world's first real time visual map of your entire automation landscape. And Darren has a fascinating take that goes against what everyone's saying right now. While the industry is screaming AI agents everywhere, he's arguing that companies need to get their automation house in order first or they're setting themselves up for failure. Darren, welcome to the neuron. Great to have you. Cory and Grant, it's great to be with you. I'm actually a big fan of your newsletter and read it every day before I get started. Oh, that's awesome. Thank you. Really appreciate it. So Darren, I guess our first question is something Cory and I talk a bit about is we usually tell people that there's a lot of companies that are marketing AI agents, let's say, but they're actually just you know, agentic or automation workflows. Both are cool, but for an AI to be an agent, it needs to reason and act on a user's behalf. So which one is make building? All of the above. So we're makes in a actually I describe it as either we were present or quite lucky, but we've been doing automation for some 10 years now. And of course, the world of automation has changed significantly in the last few years as you guys would certainly know. And so, you know, and that fateful day in November, I think is 2022 when chat GBT became known to the masses. That was a moment in time that we saw people begin to experiment with how do I actually incorporate AI into my business processes. And so make was actually supremely suited to be able to take advantage of that trend and enable you to experiment very quickly with incorporating it into your business process. But of course, as we learn more about the capabilities of generative AI, we learned AI can actually reason and that's an interesting thought unto itself. And if AI can reason, it means it's not just part of the workflow, but it's actually driving the workflow. So we actually commonly describe automation among a spectrum. And the spectrum has on one end, automated workflows that are highly deterministic. And on the other end, you have a genetic workflows that are highly non deterministic and you're giving a lot of flexibility to an AI agent to make decisions. Of course, in the middle, there is even workflows that have steps that are AI to generate content or analyze content to whatever the case may be. So we have a firm belief that all those approaches are absolutely valid for businesses both small and large to be able to really achieve significant value from what the automation spectrum and of course, agente parts of it, but non agente parts of it as well. I agree. Yeah. I think there's use cases for both, like everywhere. Yeah. Yeah. And increasingly, we see people need a skill set to figure out when to apply the right one. Right? So it's not a black and white sort of piece of so understanding what's important for that particular business capability. It is a man. You mentioned the automation industry was just so perfectly placed and teed up for all of that to happen. It's like, it's almost as interesting as, you know, Nvidia and chips. It's this whole idea that like, here's this tool. It's already there. You know, we're almost on the scale of Nvidia. I think it's very close. Just a few trillion dollars, right? You'll get there. You'll get there. So I guess what does a make agent do that standard automation workflows can't and when does the difference actually matter for business? Yeah. So agents operate most effectively environments where rules are hard to define or they may change over time. And so a make agent deployed into a business process works similarly to some other agents that people might have interacted with in the sense that I give this agent a high level role in life, so to speak, it has a goal, it has a set of instructions and kind of criteria that I wanted to operate with. And then of course, what makes an agent an agent, it's ability to make decisions about how to carry out that goal. So make is again, ideally situated because of course, if I want my agent to be able to effectively do things for me or access information for me in order to make the right decisions on that overall goal, I needed to connect to my tech stack. So of course, that agents empowered with lots of tools, make itself has over 3000 built-in applications or connectors to all the different types of applications that run a modern day business, everything from your air tables and Monday.coms to your net suite implementation or your Oracle implementation. So by empowering that AI agent that has a role and some content and understanding what it should do and then giving it the right tools and very precisely focused on giving it scalples, not sledgehammers, we'll talk a little bit about that if we get a chance. Then I can make sure that the agent can identify what tools to use, a win and let it go on its way. And what's most important about using an agent versus a workflow automation is if those rules change over time, if I have for example refund policies that determine what customers I'm going to refund in an online e-commerce store, then those rules are nuanced, they change all the time and agents respond really well to that as opposed to having to redesign a very complex workflow and think through all the potential edge cases. And so that's where agents really shine, they also shine in terms of their resilience so to speak. Unfortunately in the world we live in even now, especially maybe now with a company that rhymes with proud flair, it goes down from every once in a while, the systems go down. And so system goes down and in traditional automation, it stops, but an agent says, well what's going on? And it retries and it thinks through edge cases and it can think through and reason so to speak in order to make things happen effectively. So this is kind of where we see the differences mostly occur between that kind of traditional automation approach and where an agent might make sense. That's awesome. How do you, what's the trick to really recognizing that difference as far as light is what I need just an automation here? Is there, do I need to go with an agent? Should this be a human? Yeah, I don't know if it should be a human. I subscribe to the world where we're all going to end up on the beach enjoying life and what that looks like. We'll start back on that. All right, we're all agreed. Perfect. I mean, certainly there are still times that that human intervention is required or desired based upon the type of business process that you're interacting with. But actually, I really appreciate it. Open AI actually produced a fantastic paper on on exactly what are agents good at versus whether they not get at and it's really looking at the types of automations then that are resistant to being able to to be automated because of the fact that they're highly dependent upon context that changes all the time or different qualitative types of inputs that that would be struggled with in a normal automation. So then a fantastic paper actually refer to it all the time on kind of the five reasons you would want to use agents in your in your particular process. And I lean on that a lot. We'll drop a link to that in the in the description below this video too. Beautiful. That's excellent. I think what's interesting is to make sure that people don't just start with the solution and they start with the problem they're trying to solve. And so all too often today, especially people are like, I need an agent for this and I need an agent for this and I need an agent for this. He says, my boss tells me he needs agents and I just keep building, make workflow automations and I tell them they're these are agents and he's very happy. Very happy. You just want a non-human to do it. I think that's what he said. That's exactly it. 50.800 agents this week. Yeah. Yeah. Branding. So would you mind walking us through or if you're set up honestly even maybe you kind of show us what it looks like to to build and make? Yeah, I'll never miss an opportunity to show what it looks like and make. And it's a little bit of what what makes it unique. It makes an incredibly visual platform. We believe that that this type of technology should be accessible to people and taking complexity and simplifying it and frankly doing it in a playful way as well if I dare say so. Yeah. I'm gonna appreciate that. Make it like Billy Legos. Exactly. I have heard a lot of people respond exactly that they have in very serious business context they love using make and what I can do. Yeah. You can really nerd out quick when you first walk into doing this. Once you once you've accomplished something for the first time I can see the the addictive element there. You're exactly right. I get to meet with customers all of the world that have that reach that aha moment and it becomes I don't know not just like a task to do something but a passion literally and both how they built it and the value they're getting out of it for sure. Yeah. All right. Well, what are we looking at, man? Well, what are the most challenging and interesting parts about working with the platform like make is man the variety of use cases that we end up tackling in the world of business is ridiculous. So I always want to give a little bit of context to the types of business problems. I know that we probably have listeners that are working in all sorts of different roles and different teams and different sized companies but if you could imagine for a moment that you're a busy account manager in a New York City advertising firm, right? So this is a high stress kind of a job and environment. Don Draper. Say it again. Oh, don Draper. You're Don Draper. And you want to focus yes exactly you want to focus on the pitch. You don't want to focus on all this other management sort of stuff. So it's actually not uncommon in a situation like this for even today for a New York ad agency to get briefs like this from their customers. So we can imagine that I'm a customer called APEX athletics. I want to launch a marketing campaign and literally this qualitative information is written down in this way. This is my first clue high qualitative content. This is my first clue that this is a prime area for automating a core business process. So it used to be as an account manager you take this information you would analyze it and ultimately create a project plan for this campaign to be able to launch this effectively for your customer. That means coordinating with your creative team to get all the content right and all those sorts of things to ensure success. And a lot of people that when you think about okay well I really am a busy account executive so maybe I need a chatbot but I believe firmly that work should happen when I'm not working. And so really the first kind of principle in this process is to recognize that the trigger for this is not just a chatbot. We can certainly interact with chatbots or slackbots whatever the case is. But the trigger is it for the client basically says hey this document's ready ready for planning and I'm ready to go ahead and process this. So in make we see a highly visual indication of what this process is that's not only useful for when I build it first but also and we have customers that are running literally hundreds of businesses processes and understanding and able to zoom out and understand how these things work together is absolutely critical for their success. And visually in make we can see sort of an outline of this process. We can see some of the concepts we've talked about so far. We can see the first step represented by here is looking for documents that are in already for planning state. We can see that we're pulling these tells about that document. And very simply we can see that the again kind of just get all the automation stuff we can see that if the name of the document includes the word campaign we're going to go down one route. And if the name doesn't include campaign we're going to go down a different route and treat it in a different sort of way. I love that it's like a really user friendly approach to basically conditional logic. That's exactly what it is. That's exactly. And of course there are people that are right brain to love brain so of course you can design this exactly the way you want. There is make art all over the world where not only does your process work functionally but it's in the shape of a unicorn so it's really pretty good. I have never seen that. That is amazing. There's a rabbit hole I need to go down. That's exactly what that is. We're going to look to frame your art and put it up. I'll look forward to it. There we go. So you'll notice of course we got those key concepts of hey I've got a workflow and understand how it works together. But then you'll notice that we're leveraging AI as steps in that workflow. Again not necessarily yet yet using AI to determine the steps. But here we're using some built-in tools and make as a wide variety of customers we have some that are that could tell you everything there is to know about open AI's response model and how it works with temperature settings and output tokens. And then we have lots of customers that are just getting started to incorporate an AI. So here we have an example of how we're trying to lower that bar where we're kind of pre-packaging some of the things that AI does really well which is extracting key information in this case. We're pulling out that that that's text content from that document we just saw on the screen earlier we're pulling out the client name and the campaign name and you'll notice an example good example here's the total campaign budget. So we all know in a Google doc somebody might have put a period instead of a comma all sorts of things but AI is incredibly good to extract that information effectively. And of course we've done the hard prompt so to speak to you don't have to think through what does this prompt look like in order to extract this information. Yeah so I just want to I just want to point out a couple things because I think like everyone can understand the visual workflow here but then once you get to these like little field boxes and want to put in where that's where people get intimidated and perhaps you know stop stop working on it but it seems like here what you're doing is you have the model in that first so you're using GbD 5.2 and then underneath that you have identified the types of information that you want to grab is that correct. That's exactly right. So here every one of these steps I have set of fields and of course can include information from a previous field or I can just write out the name of the information here so I can include the content from a previous field and again when I hover over that you'll notice a little pulsing on the left side. Yeah, that's nice. Exactly. And I assumed that unlike working with you know traditional software approaches like the formatting here is not even that crucial you know the name of the client it's going to know what that means because you're you're dealing with a frontier model. I mean it knows okay we'll find the name instead of like find the line that is capital C client underscore right you know like it might have been in the past or or absolutely was. Yeah that's exactly the case. I think there are some people will be sad to hear this but there's a lot of dead rejects sites. You know no longer need to think about what rejects means and if you're listening and you don't know what it means don't worry. Don't worry about it. You don't need to. You're lucky. Exactly. That's awesome. Yeah you'll notice also into your point Grant and people kind of start out simple but you know as you think about even more complex things and in larger businesses you think about things like transforming data and we have a whole host of capabilities that enable you to very simply transform data like different date formats things like that or even math things like that and ultimately you can think of them in the same way you think of like formulas in Excel or Google Sheets and things like that that help you manipulate data from one format to another just all done in line and very simple. Right. What happens after this? So it's extracting all the all those fields are extracting information then it goes back to chat GPT. Yep and so at this point we and again we're using kind of just a couple different models but here at this point I really wanted to kind of leverage the power to kind of breaking up small tasks for for different steps within AI as opposed to if I just created an agent that said hey you're a marketing expert do marketing things I'm probably not going to get that great result. Don't swallow an elephant right like you want to try and like break it up into smaller tasks as well as they'd be successful. Exactly. So here I've said hey I'm going to I'm going to leverage open AI very specifically I'm going to give it a text prompt here and I've just said hey you're just you know just like the prompts that you guys produce in the daily on a regular basis we're given great instructions about what makes a great prompt I've just simply getting some instructions it said hey given this campaign brief I want you to give me a simple list of tasks and so So that's what this step is going to do. It's going to digest that information and run it. And then my last stop here is where I'm actually investing in, okay, I'm going to hand over the rest to the marketing campaign agent. So my agent represents lists as another step in my workflow. But of course, it's a special kind of step and it's unique in many different ways. And it's unique because it has a set of configuration unto itself. So that system prompt that helps me describe what its role in life is. And as we talked about before, an agent's really powerful because it has a set of tools and capabilities to do things. Here I can see that the tools I've given it some simple, like, send an account, a Slack message to the account team. And some that can be complex and nuanced based upon my business, again, makes ideally suited because it already has connections to all these different things. You can see that as in this case, I'm creating a specific Canva folder structure. So not just telling AI, hey, you can create folders in Canva, but it's actually a scenario that I've created that says in our business, we always create a draft folder for each customer and we create an impregnance folder for each customer. And so creating that right folder structure every single time and mixing these concepts of determinism and on determinism. That's cool. Okay. Can we, can we, number one, zoom in on this if that's possible? And number two, if we could just kind of break out. So I see there you have a, let's just go from the top. So you've got the tools listed. Then what's the next category? And what's the next category after that? Yeah, fantastic. So we have both the types of tools that we have, ability is a single module, essentially one of these modules you see, or an entire scenario, which executes, you know, kind of everything. So we can see that example that I just gave where the tool I'm giving to the AI agent is it can start this process and it's always going to run these four steps and then be done with that process. And this is sort of the interesting part because you're playing with the situation wherein you may or may not trust the AI agent to get it right every time. If I have to break out caps lock, then then probably like in a place where I should fit a different approach. Yeah, that's fair. Yeah, never do this. Never ever ever. Right. So great. So you've got those tools like easily add and of course I can continue to add additional tools as well to say, hey, and the example I use all the time that's kind of interesting here, both of these approaches make perfect sense. But at the end of this process, I could add on another step, which is a Slack message and say, and then at the end, just send a Slack message, all determine the format and exact message. Or I can give a tool to the agent and say, hey, when you're done with everything, send a Slack message. And these are both legitimate approaches and one of them I have more control over and one a little bit less control. Whether that's give me an update or send me a form message that says project A is done. Yeah, that's exactly right. Okay. Cool. And at the bottom there, it seems like you have MCP, which is your connector. Is that just a connector to Canva or what is that? So these are additional options. So of course you can use modules or scenarios. A third type of tool, so to speak, is MCP, which increasingly a suspect many of your listeners are familiar with. But this is the model context protocol. It's a standard way of an AI client talking to a tool most often, a software application. We could definitely talk about MCP for a while, make operates in a unique place where this is an example where we're using MCP as a client and just to kind of show off that you can connect to any of these MCP servers. Yeah, it makes interesting because it's also a MCP server itself. Right. Yeah. That's a bit. In the sense that you're connecting to different other platforms like the API. So you are sort of like the ultimate MCP server in a way because you connect everything. That's exactly it. Yeah. And you actually get a little bit more control over that again. So how much agency do I give AI versus the kind of precision that that making it able to kind of manage an orchestrator at a higher level? I had not considered that essentially you're an MCP. Yeah. It makes so much sense. All of the integrations. Yeah. Yeah. And it's good because you know, oftentimes I hear the criticism of MCP is that it sends making blowout the context window, right? Really bad. Because you're connecting all these different functions that the AI can use at any given times. And then if you include instructions with it, that's very complicated. This make is a lot better MCP in some ways. I mean, you can tell me all the ways because it's very deterministic and very token efficient in my contrast. I didn't know that. Right. You nailed it. You're like, you couldn't have said it better. Something we spend a lot of time thinking about and kind of positioning how do you know, MCP's new to a lot of people. So they're learning the best way to use it. And then how do we think about and understand that you make ads on top of that? Those those context windows and ironically, this is a great irony. Agents and AI need good context to be successful. And ironically, the more you throw at it, the less successful it is. So this is about precision. Look at an example of that, actually, that exact concept as well as we wrap this up, but this particular demo. Yeah. Yeah. And I had two more questions about this at the end. I don't know if you can address them. Yeah, please. The two questions are number one, the trigger is great. I'd love to know a little bit more about how that works because when I'm building my own automations, oftentimes coming up with the perfect trigger for something is a lot of times my sticking point ironically. Yeah, that's interesting. Yeah. I actually spent a lot with some people earlier this morning and they were having those exact discussions. So very often people think when they think of AI, they think the trigger is, I say, I send a message to it and then it talks. And we've kind of been trained that way right from the chat interface is like, okay, I need to do something. Now I ask for it. Yeah. I always wanted to be something really simple. Like I built an automation in here two and a half years ago where what I wanted essentially was to when I found an article that was important when I was just reading the news on my phone, I wanted to be able to hit the share button and have it go somewhere that triggered it. And I wound up running it through. I can't remember if it was feedle or anywhere or one of those and it goes through there. And then it grabs it, sends it to a Google sheet and then the AI would grab it and the AI would run it through the process. I'd set up there and deposit the answers back and I wanted to be able to do that from my phone at a red light. If so, if so be it. If I chose and I wanted to be able to hit share and be done and just know it'll be there in the morning. I love it. And it was really effective. That's awesome. Yeah, I mean, there are three ways I typically think about kicking off work. And so one of them, which is not the example you described, Corey, but one of them is like on a regular basis. That regular basis might be every one minute or it might be every 10 minutes or every day. I mean, I do something. So that's a pretty straightforward way of kicking off work on a routine schedule. So depending upon the software applications you're working with, working with obviously Google Docs here, it requires that. So every five minutes, this is kind of checking. And of course, I can kind of fine tune that as much as I want. But it's checking Google Docs to say, hey, are there any files that I have not yet processed? That's kind of a key phrase that actually kind of people might not always understand the nuance, but that I have it yet processed that's in the ready for planning folder. And I'm basically constantly pulling this to see if it's where to go. So that's one way to kick off work. The second other way is slightly technical, but it's a web hook, but it's an instance. It basically means that other application has the ability to send information my way. And I might have wired this up to my e-commerce front end. And it's immediately going to send information to this scenario and kick off that agent and interact with it. And the last external action is essentially the trigger. In the e-commerce scenario, it would be, you know, someone makes a purchase, for example. It makes a purchase. It signs up for something or precisely. Yeah. Okay. That's exactly it. And then the last one in the super simple way to describe to do what you were working on Corey is we actually have a concept of mail hooks. Basically, I can have a set up a magic email address that's like, hey, to do lists later, I just give it that contact in my phone. And then I would just send it to do lists later every time it receives an email. This is kind of old school technology. We don't have the facts option, but we do have the email option. So if it kicks off, uh, based upon a new email, so what would you go address? That's all. I wouldn't be surprised with robotics if facts comes back in some way. I don't know how, but I could just see like, I recently heard that how you communicate with Dolly Parton. I know. You've fax Dolly Parton or that you receive faxes from her. Oh, I don't remember where I heard that. I remember being just floored that wow, there is someone still using it. Well, we got we got some investment advice. I think they're from Grant. That was excellent. And people got that worth out of this. podcast for sure. That's right. It's the next big thing. Yeah. Back's is the next big big. My other question and perhaps you can address this because you're going to show us the scenario of dealing with MCPs and the precision is let's say I'm intimidated by all of this. Could I ask an AI agent to set one of these up for me, whether that's track TBT or maybe perhaps something inside the mic? Yeah. So you read our mind. So we're as we intro the show today, we talked a little bit about some of the things that we've released and we're really excited about over the last year. We recently had our customer conference in Munich, Germany and got to spend some time with just fantastic, amazing people from all of the world. And we unveiled one of our most exciting things that is in cooking as we speak. I'm not ready to show it off today, but we're launched. Maya by by make we're working on some select kind of interviews with some some key customers to make sure that we nail this experience. And essentially, that's exactly as you describe it is a in product experience that allows me to chat with AI in order to build out these automations. And the one thing that I really want to emphasize about that that I'm particularly excited or sorry, I don't have the most amazing image on the screen, but I'm Korean Grant. I suspect this is true, but I don't know. Have you guys vibe coded before like we can vibe. We give it our best definitely. Definitely definitely. Yeah, perfect. Some people get offended if I use the word vibe coding. So I hope that's. I am nowhere near close enough to a legitimate engineer to be offended by. So you're all good. I resemble that remark as well. The beauty of it like for me is sort of it's the difference between a like a one shot. And that's what you'll see a lot in products like this in other places. You'll see like I give it one prompt and let's let's just hope and pray it comes up with the fantastic scenario that's amazing. That like that's not the real world that we live in. I've been doing this a while. I've never seen two businesses even of the same shaper size that operate the same way. So much like universal creativity on how people operate a business. And so Maya by make is actually a highly interactive and is designed to very similar to that vibe coding experience where it's not just one shot. It's actually a continual interaction where it's going to ask you clarifying questions and build alongside you. So I'm particularly excited about that release coming soon. I said I will say you know today people obviously they'll go to cloud or they'll go to a chat GPT they'll even create a workflow diagram on a napkin and upload it to a cloud and say create a mix scenario for me and have pretty pretty solid solid success. Yeah, that's not the experience that we totally aim for. And so we're working heavily on bringing that to market. Well, you know something I feel like is really valuable here is like. Like I would say Grant and I really probably started doing this in like an in a din kind of tool that's a little more technical little less user friendly. And as time has gone on, you know, we're getting more and more to a time where it's just going to be expected that employees are capable of building deploying and maintaining their own agents. And that's just going to be a fact of life. And what I like about this is it's not it's not intimidating. It's not scary. I love the idea of being able to build these with natural language. So I'm really excited to see my one is ready. And and I think that that real that kind of approach really can enable a bunch of people who right now might be scared to death of the idea of building automations. Or being automated or being automated. Because I think if you can build these types of automations that makes you valuable because then you can scale how much you can do as an individual. Yeah, you're not going to automate yourself out of a job. You're going to automate yourself into one. Into more work. No, now I mean, I often theorize that look, what I would describe is orchestrating AI orchestrating agents. It's the new middle management. It's no longer the case that hey, you're doing great. I'm going to promote you and you can hire some people, but you're doing great. I'm going to promote you and you can 10X your output by creating a managing an army of agents. Yeah, I think that's 100% true. Whether or not people like that. The remains of you see that I mean, certainly knowing the skills and knowing the tools will empower you in whatever role you're in. Yeah. So I agree 100%. I kept joking. Sam Altman told me I would have a fleet, but we would all have a fleet by the end of this year. But the truth is I have six. I'm religiously, but definitely by the end of next year, I can't imagine a scenario where just about everyone isn't dealing with a lot of these. Absolutely. It's a baseline expectation for when you come to work now. Yeah. This is a great demo, by the way. I think this is going to be incredibly valuable for people who have never used this before. Yeah. I think it's good. I keep thinking of the people on our team who really love the colorful workflows in our project management tools that we use and stuff. I always struggle with all of the colors for whatever reason and often use a like grayscale thing on my screened out. So many people think better and are less intimidated with that approach. And I think, hey, what you've got here's really cool. That's a big compliment. I can tell you it was very intentional. I get to I'm lucky enough to spend some time with our co founders. Actually, here's a here's an interesting insight. Our co founders based in Prague and the Czech Republic. And part of the inspiration for make was literally coming from the Metro maps that you come from. If you think about like a business process and you think about taking complexity of how to get from one part of the city to another, the most effectively, like literally what you're seeing care has a little bit of inspiration from those those Metro mats that you I see. I totally see it. Yeah, absolutely. I totally see it. I'm thinking of like, yeah, if you're in London and you're on the tube, you see these little lit up screens all over. Yeah, absolutely. And the lines are different colors like the purple line, the pink line. Yeah, I mean, there's so much like subtle happening here like like the way that that's highlighting up is very intentionally directional. Yeah, you feel and that's exactly the word I use all the time. You feel powerful. You feel connected. And it seems strange for software to be a motive, but it is it is agree. Well, you talk about building an AI stack with four layers that's data models interfaces and speaking of orchestration orchestration. Where do you find that more companies what most companies get stuck when they try to build this type of stack? Is it the data layer model layer? Yeah. Yeah, so the one thing that I find and when we think about that kind of very often people in the no code world, they refer to that the trifecta of no code. They think about data. So and very commonly tools like air table that and even Monday they distort it in very flexible sorts of ways. Then they think about interfaces and building on interfaces on top of that. And then they think about how do you automate and integrate between all those things. And what are the places where so people come up with insane ideas on how to create incredible value for their organizations, both internally and externally by combining the power of these different stacks and incredibly flexible data source, incredibly flexible interfaces. And of course, automating everything in between. And the piece that we actually saw and we started to see this several years ago is when you start to the great irony of. You've got a fantastic business. You start to automate and and no code every part of your business. It becomes difficult to rain in and understand what's happening. And connects to all the various pieces. That was really the inspiration behind behind a concept called make grid. And we're actually going to ignore the while these errors are intentional. So I should highlight that we actually might actually let's get rid of those for a moment. So make grid and represents how this entire stack connects to everything. So very what we found with make we were looking at a very specific process like one process, but your business is way more than just one process. And so often you were looking at the tree but missing the forests in the broader scheme of things. And so what make grid is a unique one of a kind automatically generated view of your entire automation and landscape. Not only does it figure out what how processes are connected to each other, but it actually identifies how key assets in your company are connected to different processes. The direction of data flow to. Yeah, I love that. It's a beautiful thing. I noticed they're not all the same. That's what caught my eyes. That's exactly that. Not to put you on the spot, but is this real time because that would be really cool. This is so this is showing the direction of data. We've actually continually pushing on more and more layers that will put you to that kind of let you look at different. And so there's a volume of pieces over time. So there's only there's not a lot of volume there, but if I look at actually I could see. the volume here. This is a very highly active workload. Fun fact, the process we're looking at right now, that's the conference registration system. Let's see. Ah, this is amazing. So we have a business process. So I get to travel to Prague regularly. We're super lucky at Prague. We serve lunch to everybody there on Wednesdays and it's fantastic catered lunch. Like it's good stuff. But if people don't come, then we waste food and if we don't make enough, then that's a problem as well. So completely automated process every week on Slack, you get a notification and you respond to the emoji, like a meat sign or a vegetarian sign or like what you want. And then you get a QR code that's a wholly automated process. And we're looking at that process. Like that process has lots of different steps. It has lots of different pieces. I can see the Slack channels that we're using to manage that and connect all those pieces together. But what's important here is I see these processes how they all work together. But I also see dependencies. And so in this case, air table and this particular base in air table turns out to be really important for at least three of these business processes. And if I change something about that air table and in fact, I can even drive a little bit deeper here to understand all the different connections. These are the different attributes in that particular table and air table that they're dependent upon. Like has a lot of insight into not just how your automations are working, but actually your whole tech stack and how it's all connected together. And that works for no code things like air table, but it also works for enterprise things like work day and net suite. So you can understand that these are critical assets that connect multiple processes together. That's awesome. Right. Now is this connected to the idea of the precision that we had mentioned earlier, like in how this works? This particular piece is really about zooming out. I think precision as zooming in a little bit where you're getting that's fair. That's fair. Precisely with the scalpel. Yeah. Yeah. Yeah. And this is great because I don't know of any other tool that shows you at this higher level that you can visualize all your automations. I could imagine a CEO coming in and scrutinizing this. What are we doing? Like how can we do it more efficient? Like you look at the data flows. How can we get rid of expensive subscriptions that we don't need anymore? Because one of these other services replaced it. Yeah, you're absolutely right about that. If you're connecting core software through low code automation, it makes it less reliant on that core software and you're able to not be vendor locked in for sure. Yeah. Yeah. Because you can say, oh, we can swap this vendor right here and replace, you know, all the automation still work if we switch it up. Yeah. Amazing. Yeah. This is a one up kind of experience so I can really kind of go deep on any one of these pieces understand how the relationships work and it's incredibly visual and all updated in real time. Is there an easy way to know like here's everywhere I need to go make a change or at least the dependencies I guess. Look up all of my automations connected to click up for example. Yeah. Like that is a little bit of what this is getting at. So I can see here and so many people use a spreadsheet as a database. Yeah. Whether we agree that's a good idea or not, that happens all the time. It does. So this is that that spreadsheet is an asset that it's become an important asset for my business. And I can see exactly that spreadsheet. Of course, if I click on open spreadsheet, I can go directly to that real spreadsheet. I can see very clearly it's Google it's spreadsheet and I can identify all the different links that are connected to that particular spreadsheet. But we also have literally all the attributes that are in place. So literally there's a word, there's a column column F in that spreadsheet called classifications. And it's being referenced by these three scenarios. And so I do know that it's really if somebody adds one because it may not be relevant for these automations. But I do know if somebody needs to change this or modify it or delete it or if we decide that we're moving from Google to Microsoft. It's going to be critical that I know where all this is connected and how to read the line up. That makes sense. Yeah. That's that's exactly what I was looking for. I was just curious. My theory is that over time as we have more and more and more of these, we're also going to have more and more and more ways to break them if we're not careful. And I think a system like this is a really important part of building out agents having that overhead sky down view of your automation situation for like a better phrase. That's exactly where we sit and we're you know, again, either lucky or precious. I don't know. But if you think about what's necessary for AI agents to be successful in real production situations, you're going to need to know the relationship between AI agents and what scenarios, what tools they access and be able to drill into that to understand exactly what they did at any point in time. And that's very much our focus. Awesome. Here's a here's a feature idea for you if you haven't done this already. So if you have, I'll be really impressed. Nice. How do you export all of this information to give it to an agent so that it understands how your company works? Is it possible to do it? Yeah, I'll definitely categorize that in the bold ideas from Grant. It's interesting. Max machines. Agent, Agent maps. Yeah. I got to write these down. I don't have. What are these days you're going to make literally some of the dollars? Yeah. I'm just I mean, there's there's an interesting relationship here. I don't know if you guys are aware, but so make is actually so we're very lucky because make is is managed as its own distinct brand, its own distinct customers, but we actually have a parent company called Solonus, which is a data mining company. It really helps companies map out and understand their actual business process based upon data that they collect and of course knowingly so. Yeah. And yeah. So there's an interesting tie in a relationship there to help you understand what your actual business process is like you think you have an account receivable process, but then there's the reality of the data and how it gets processed and it's different. Could could we go look at the the workflow again? I have a question. Yeah, you got to suppose I I know we're talking low code, no code, but suppose I have something that is just complex and it's a pain in the neck and and my only solution is I need to shove in a code node somewhere. Is there such a thing? Yeah. Why would you want to do that? No, I'm just kidding. I don't. I promise, but if in the event that was my best choice. So some of your listeners will perk right up. They'll be excited to see exactly this, but and of course I don't code, but Gemini does. So that's right. I'm actually incredibly proficient with including a make code module when you need it. And of course, again, code, one of the things that's great about arbitrary code, it can do anything. Yeah. But of course that means doing it in a secure way. And so we've invested heavily on enabling you at any point in your scenario to be able to execute of course, you know, the standard JavaScript or Python code and to be able to dynamically send data to that as well. Yeah. Oh, that's awesome. That's awesome. I just it was just a thing that crossed my mind. I was like, you know, the goal is not to, but you know, those are goals are what they are. And sometimes they don't happen. So if you need it to, but you know, if you were dealing in snippets and stuff, the average person could absolutely go in and find their way through that using a chat GPT or a Gemini cloud, whatever. And and stumble through some small like, I can't connect this to this. I need something there. And probably it'll tell you, you need to use this code. I'm going to write you. That's exactly it. Yep. Okay. So we have a couple lightning round questions for you. Is there anything else you want to show us before we get to those? And I can do this all day. Maybe very briefly, we touched on MCP. So let's like very briefly do kind about a demo of what that looks like. That'd be great. Perfect. Yeah. Tell me. So this is totally made up. I can't spell Neuron either. I don't even know how. Oh, it's fine. It'll know. The AI smart. They're good at typos. Right. But I'm going to say, tell me about my customer. I'm in pretender my customer of the same tree. But we all have imagination. So I have cloud up here. We'll see at the demo gods are have been properly satiated today. But I'm within anthropic and you'll notice that this is, I've already set up and kind of configured make as my MCP server. So it's looking at various tools that it can utilize in order to effectively support my query, my request. Yeah. So from a technical perspective, I've actually set up a scenario that's designed in that way. And what's important to recognize is my scenario could be, you know, grabbing information from Salesforce and fresh desk and the news for that for all that matter. So it's everything. And it's going to call that one tool. I can actually see the inputs that it provided. I can see actually, you know, if I really want to geek out, I can see actually what it did do. But now I can talk about my customer in sort of a natural way. Like I don't know any salespeople that want to log into their CRM. I can imagine that there's a world in which that doesn't exist anymore. Yeah. So, it's a happy world. It's a happy world where everything is nice. So of course pulling out information, that's not too difficult and it's probably okay for it to be a sledgehammer. But when I get into more distinct business processes, I have a new customer to my CRM and I'm going to say Starbucks. I just landed them. It's happy days. Oh, congrats. So, this is joke. Oh, I'm sorry. I was in on that. I'll try to be less literal. I was like, hell yeah, go Darren. Commission checks on the way. So I've done this as well. So you might say to yourself, why not just have like the hub spots here, the MCP server or whatever the case is? So here you can see, I actually, I said, add to Starbucks to my CRM, it says, okay, the website's that, is that good? And I'll say, yeah, that's good. So at this point, it's got enough information that it feels confident. It's going to go execute that tool. I'm going to show you what that tool looks like inside of make and why there's power for the make MCP server that's different from just what I would call the standard one. So, by the way, I've got to know a great awesome. So this is that very simple scenario. It's probably oversimplified, but in make, when I create an MCP tool, basically a tool that's available to any AI, I get to define a few things about it. Actually, I get to define the description of it. So like, this is the tool, this is how it's going to be described to the AI. So the AI kind of knows how to use it. So I have control over now. I get to decide what inputs this tool expects. So the name, it expects the name of the business, and you'll notice the business domain. And I just wrote a very simple phrase here. I said, always confirm the business domain by the user. So you can kind of think of that as a business rule. So rather than just giving AI, here's all the possible fields and HubSpot and go for it. I've only said there's two. Maybe there's more in reality. I'm going to add a few more, but I've only said there's two here. And I've added in that this one's required. I always need the web address. And then of course, I can go in here and I can add it to the CRM. So basically, I've got a wrapper around the MCP, or if you think about it that way. I have total control over the inputs. I have total control of the outputs. You'll notice the output I provided was the HubSpot URL, exactly. So I get that back. What you're likely to get back from HubSpot's actual MCP and HubSpot's awesome partner, but you're likely to get back an ID. And then you're like, OK, cool. But here I have complete control to get precision. And then of course, I can see every time this scenario is executed. So we haven't talked about it a lot, but I can geek out on this all day long. Every time these workflows get executed, knowing what actually happened, what was the inputs, what was the outputs is incredibly valuable, especially as you scale your business to be able to have that insight. >>Exactly. >>Yeah, that's really cool. >>That's awesome. I've never thought of this use case. So this is amazing. >>Awesome. >>Yeah. So we said in the beginning, make has 250,000 organizations using it, which is massive. What patterns are you seeing in how the most successful teams structure their automation and AI strategy? If I'm starting with make, what should I be doing? >>Yeah. So the patterns that we're seeing and we're looking across all of our customers and understanding what resonates and who's most successful. One of the most distinct patterns that we see emerge is that we see different types of business functions where AI and automations most likely to be deployed in and most likely to be successful in. So we're seeing incredible uptake in business functions like marketing. Maybe this is not a surprise. But the reason for that is in the world of AI, the things that matter a lot are high content and high context situations. >>Yeah. >>And you really can't think of more high content and context situations than within marketing. So we see a significant push from there. We see the same push and things like customer service as well where again context is high and content is high. But one of the interesting insights I've had most recently is when I narrow it down and I look at some of our largest customers. So these are customers with many employees, many different departments and divisions that are thinking holistically about what I would call AI transformation strategy. I've been blown away. I didn't expect this at all. But in the data, I see significant uptake and also qualitative and quantitative within finance departments. I would have thought they were the last to really start to adopt these technologies. But I've talked to amazing customer bars that's a food delivery service and they're literally using AI agents to pre-create journal ledger entries. It's been a long time in my creations. I've created one of those. It's interesting to me. I think a lot of finance people I assume would be stuck. I'm going to create it in the old fashioned way. But they're happily and ready to take that information from the business and leverage AI to to create those. It's been a fascinating trend. Also, I have a question that might even be a little corny. But I'm going to ask it anyways. What do you think is the single most impressive automation in make? Wow. It's a neck-cording. It is very corny. It's very corny. I was just curious. What's the, if there's one that you're like any time I want to really impress somebody and rock their world, this is the coolest thing you could do. I'm going to lean into your corny. I have a soft heart. So I'm going to lean in on that way. So I get a chance to work with all sorts of different types of organizations in the, in the, the Pacific Northwest, to work with an organization that runs a pet adoption center. So this is a pet adoption center. It's not necessarily the big enterprise use case you think of, but it's like how make it's applied in the real world. And there's like one guy who's volunteering his time to help them set up. Because if you think about it, if you work at a pet adoption center, you don't like literally they have Salesforce to track people and track the adoption process, et cetera. They don't want to spend a lot of time in Salesforce. And so literally simplifying that process, I was, I was blown away that one of the like oldest school problems that make an AI does really well at is actually matching. So what I would call fuzzy matches. So it within a pet adoption is actually a do not adopt list. So they have people who for whatever reasons in the past have shown themselves not capable of successfully taking care of a pet. So they do not adopt list. It's name a person, maybe an address phone number. And so if you think about that business process, so to speak, the business process is making sure that a person that comes in is not on that do not adopt list. Fuzzy matching is like a hard software problem to solve normally like is this is Tom Thomas and Thomas Tom, like are these the same person? Yeah, that's solved. And AI does a really good job at it. And so I'm going on leaning in on your fuzzy warm billing kind of those are the types of use cases that get me really excited. I can talk about big ROI ones for big companies all day long, but I love seeing stuff like that actually deported the real world. That's really cool. So where do you think like make grid and everything that you're building goes next? What becomes possible when everyone in a company can understand their automation landscape? Yeah. Yeah. This is what I spend most of my time thinking about. So grid is a significant investment into not just fancy things that you saw on the screen before, but it's a significant investment into becoming not just understanding how that set up, but becoming proactive at the macro level. So identifying where opportunities or challenges exist in your automation landscape and then responding to those. And maybe it gets at sort of the very big picture of make and how we think about it today, which is so automation at its core. Yes, it saves you time. That's absolutely true. But what makes design to do is to support agile businesses, basically businesses that can respond to a very fast technology environment or competitive environment. And if we can help businesses do that and empower people that are closest to the business functions to be very responsive and innovative in the broader market, that's where all the value comes from. And that's what we're trying to serve with things like make grid and even all the little things we do in the scenario designer to create that visibility so that you can not just automate it. Forget about it. But it's actually continue to innovate and iterate on it in order to outlast your competition or do better things. That's really cool. Darren, thanks so much, man. I really appreciate you coming today, joining us and showing us how to build agents in make. It's a really cool tool. Where can people go to learn more, especially if they want to learn beyond what they saw here today? Yeah, fantastic. One of the best kept secrets is, certainly, our hands-on training that's at academy.make.com. If you're just getting started, this amazing set of resources that are walking through and at whatever pace you seek and lots of amazing resources. If you're more of like an executive decision maker and thinking, man, my company needs to really transform and change the way we are. operate and you want to know where to get started. You're not going to drive in and necessarily be operating the tool. I recommend checking out playbook.make.com. This is how you think about driving AI transformation in your company and where to get started. Great resources. Absolutely. We will have links to all of that in the description below today's video. I want to thank everyone for watching. If you haven't yet, please take just a moment to like and subscribe so we can keep bringing you these interviews and telling you about the cool tools and the interesting people in the space where we all live and work today. But that's it for today's episode. Farewell for now, humans. We'll see you next time. [BLANK_AUDIO]