Transcript: How Decagon Built Human-Level AI Support: Ashwin Sreenivas on customer obsession, early traction, enterprise complexity, and the AI concierge future

Can you talk a little bit about why customer service is, and what led to in this direction? This is actually very driven by our customers. Number one thing that you are trying to validate is, am I building something that people will pay money for? And what we've noticed is, my guest today is Ashwant Serias, he's the co-founder of Decagon AI, previously that one.5 billion dollars. Like Notion or Figma, you would never think about picking up the phone and calling Notion, right? You'd be like, oh, I want to like chat or email. Our customers complaining about voice agents or what has been the reality? It's also part of like customer trust. Most customers, they don't want to talk to a human, they want their problems solved. How they want their problem solved? They just want the problem solved. We didn't overthink, oh, what is the exact right two-year strategy or how are we going to build mots over three years? We said, okay, we'll figure all those things out when we get there. The only thing we're going to worry about now is how do we build something that someone will pay us real money for in four weeks? How do you acquire your first say five customers? I think customers think about value who weighs primarily. One is, Welcome to Startup Project. My guest today is Ashwant Sri Nias, he's the co-founder of Decagon AI, notable brands like Duolingo, Mercado Libre, Notion, Aura, and many others use Decagon's platform to automate critical support workflows. Ashra was previously co-founder of Fidya which was acquired by Scale AI. We've been covering a lot of companies where you know we're taking AI and solving a specific segment of problem and I think Decagon sort of fits in that category, where you're taking AI and solving a particular set of enterprise problem out there, in a better way than we did before. In this conversation, I'm hoping to explore, you know, how Decagon found product market fit. What is their secret sauce? What is the tangible business impact that they are delivering to their customers? So if this is the first time listening to Startup Project, don't forget to subscribe to us wherever you're listening to the podcast. Ashwant, welcome to the show. Yeah, thank you so much for having me. I'm excited to be here. So let's uh get right into it. Uh what is Decagon AI? What is the product do? And talk a little bit about the technology become behind Decagon. Yeah. So you can think of Decagon as an AI customer support agent. So, um for our customers, Decagon talks directly to their customers and has great conversations with them over chat, over phone calls, over SMS, over email, uh and our goal really is to build these AI concierges uh for these customers. Uh and you know, this this this idea of um um AI for customer support is not necessarily new, right? Like you've had chatbots for, you know, 10 years now probably. Uh but I think the thing that's really different this time is, if you look at the chatbots from, you know, as as late as three or four years ago, it wasn't a great experience, right? And the reason it wasn't a great experience is because you had these decision trees that everybody had to build and it was a pain to build them, it was a pain to maintain. And then from a customer perspective, if you have a question or a problem that is, you know, 1 degree off from the decision tree that was built out, it was completely useless, right? And that's when you got people saying agent, agent, agent. And I think the thing that's changed and and you know, a lot of the core of what we built is we've built a way to kind of train these AI agents like humans are trained. So you know, humans have standard operating procedures that they follow, and AI agents have agent operating procedures that they follow. So we're able to essentially build these AI agents that can have much more fluid, natural conversations like a human agent would. Oh. Talk about the products, you mentioned chat, phone calls, emails. Uh, do you have products for everything or if a company is coming to use or adopt and Decagon are they first starting with chat and then increasing to everything else or how does the customer journey look like? Yeah. Um, this is actually very driven by our customers, right? So for a lot of the more tech native brands, right? Think like uh a Notion or Figma, you would never think about picking up the phone and calling Notion, right? You'd be like, oh, I want to like chat or email. Whereas some of our other customers like Hertz, you don't really email Hertz, you're like, oh, I have I I I need a car, I'm going to call them up on the phone. So a lot of our kind of uh uh deployment model is guided by our customers and how their customers want to reach out to them. Um, but typically most customers start with, you know, the the the method by which most of their customers reach out to them, and then and then they kind of expand to all the other ones. So for a lot of people, very common to start with chat and then expand to email and voice, or start with voice and expand to chat and email and so on. So I wanted to double click on that point. You didn't mention the decision tree model like I think during 2015 or later during that Alexa peak error, everyone was building chatbots and uh I remember seeing the app ecosystem where you have to build you know apps on Alexa or even I think Microsoft had something like Cortana where you can build apps uh and conversational bots I think two or three years they were all the hype and quickly you know it got stagnated and that we realized that okay, all we are doing is basically the same options that we used to give on our customer calls where we press one for this press two for this. They're just doing that on the chatbot but you know, you're giving a different cities of decision trees uh and you're defining by the decision tree and anything else you're having basically an if else uh you know command line and at the end there's a catch all where you're actually driving it through a human to chat with it. Um, I think that was sort of like the um well we got quickly used to it it was slight slightly an improved version of what we had but we quickly sort of saw that oh this is the end of it and it was not um, you know, improved after that. Um, would companies, you know, there are obviously a lot of players in customer support uh and you know there are existing you know companies which have scaled during you know using this approach, other approach and a whole suit of tools. Uh do they have a specific edge on creating something like what Decagon is doing because of the existing data? Yeah. No, I actually think uh interestingly enough because uh uh I I I think these uh you know, customer service bots went through a few generations of tech. Uh you know, fortunately for us, the tech is different enough that, you know, you don't get too much of an advantage starting with the old tech. In fact, it's you you you start with a lot of tech debt that you then have to undo. Uh you know, let's say 10 years ago when you had to start with explicit decision trees, right? Where you have to program every single line. And then about 5 years ago you had the sort of Alexas of the world. And the it was a little bit of an improvement, but essentially all that it did was it allowed a user to express what they wanted, you know, they can come in and say, I want to return my order. And what the models were good at doing was detecting intent, right? Saying I'm going to get this natural language inquiry that came in and classify it into one of, you know, 50 things that I know how to do, right? But then beyond that then everything becomes decision trees below that, right? So you got a little bit of a jump, but then not much more. The thing is now with these models because you have so much flexibility, because you have the ability for these models like follow more complex instructions, follow, you know, multi-step workflows, you can actually build this rebuild this from the ground up. It's not just you say something, you classify it into some intent and then you get a decision tree below that. It is we want the whole thing to be much more interactive because it's a much better experience for the user. So we had to kind of rebuild it from the ground up to say, okay, number one, how does a human being learn? How does it how how do you instruct a human on how to solve issues? You have some standard operating procedures that you follow. and you say, hey, if a customer asks to return their order, first you want to go check this database to see, you know, what tier of customer they are. you know, maybe if they're a Platinum customer, you want to have a more generous return policy. if they're not, you want to have a slightly more strict return policy. You need to check the fraud database to make sure that this uh customer hasn't been flied for fraud before. You need to go through many of these steps and then actually work with the customer to get that return happening. So the core of what we've done is is is kind of build out uh uh AI agents that can follow instructions very well like a human does. So, uh I think this whole concept of AOP that you guys introduced is very fascinating. Uh so you mentioned SOPs, which humans really need and understand, and then you have AOPs, which is sort of you're telling the agent what it should do in different scenarios. Uh sort of like a protocol for the agent. Um, then who is converting the SOP into AOP? How easy it is to actually, you know, create this agent? Are you giving a generic agent which adapts to a SOP of a given customer or I as a customer, you know go ahead and you know some sort of um you know I have to build the agent. Yeah. No, so the core Decagon product is we've built an agent, we we've built one agent that is very good at following uh uh instructions and AOPs, right? And so, uh the reason we built this is for time to value, right? Like if you have to train an agent from scratch for every single customer, it's going to take a lot of time for that customer to get onboarded because you know you need to like fine tune a whole bunch of models from scratch. And then two, it's very difficult for that customer then iterate on their experiences, right? Because every time you need to change something you have to refind tune something. Whereas, if you build one agent like a human that's very good at following instructions, they can come to that customer and they're like, oh, here are the instructions I need to follow. You can be up and running immediately. Now in terms of how these AOPs are created, most customers tend to have some set of SOPs already. And AOPs are actually extremely close to these SOPs, right? The only thing that you need to change is you need to instruct it on how to use a company's like internal systems, right? So it's 90 99% English, and then there's a few keywords to let it say, oh, at this point you need to call you know this API end point to load the user's details. At this point you need to, you know, issue a refund and the way you issue a refund is using the stripe end point and things like that. So that's kind of the primary different from difference from SOPs, um, because you know you need to point it to specific internal system, but other than that it's almost entirely in English. So it's the the lift and shift is very, very straightforward. So if if you talk about like the, you know, the stack, the technology stack that you're building, uh you have the models, you know, which I'm assuming you're using external models and you're not training your own models. Um, correct me if I'm wrong. Um, and then you're top of that you're building this sort of like some version of fine-tuned uh model that is really good uh at uh answering or doing customer support. Is that the way to think about this what you're doing or what is the like the steps after, okay, there's a model, but then what is the difference between a model to what you're delivering to a customer and what kind of sau sausage uh sauce is you know you're adding into it? Yeah. So we spend a lot of time thinking about models. So we do use some external models, but we also train a lot of models in house. Uh and the reason we do this is because um, if you're using external models, most of what you can do is through prompt tuning, right? And we found that models are only so steerable just with prompt tuning. So we've spent a lot of time in house being like, okay, how can we take uh, you know, open source models and um, you know, fine tune them, use RL on top of them and use kind of all these techniques to steer them to do the things that we want. Um, so we we we use a mixture of both. Uh and the interesting part, the the the way you get these models to, you know, uh follow instructions well, is to kind of decompose this task a little bit, right? Saying, okay, I have uh uh a customer coming in, they've had this question. Now, I have all of these AOPs that I could potentially uh select from. So decision number one is, is any of these AOPs relevant? If a user is continuing the conversation, okay, are they continuing the same topic or should I switch to another AOP? So, at every step there's kind of a hundred micro decisions that you need to make. oh, did the user change their mind? Do I need to go back to an earlier step or are they continuing in which case I can proceed. Did they just completely change the topic in which case, you know, I should go to another AOP? Do I need to undo decisions that I made before? So a lot of what we do is kind of break down these micro decisions that uh that the AI agent needs to make and and and have models that are very, very good at these. I think one of the industry narrative has been that, you know, only the companies in a bit very large capital uh chest can actually do training. So you know the Open AIs and that topics of the world or the Microsofts and the Google of the world can train models uh and the estimate was in a couple of years small companies can actually you know train uh their own models. So uh are you seeing that cost drop down when when you mentioned, yeah you're also trying to train open source models. Uh are you seeing that um yeah cost being more accessible now? Like on if so like on what order are we basically looking at? Yeah. So we're not uh, you know, we're not pre-training our models from scratch, you know, we take open source models and then do things on top of those. Um and I think the thing that has changed dramatically is that the quality of the open source models has gotten so good that uh this is now viable to do um um you know pretty pretty pretty quickly. Which which models are you like um, you know other than for your use case? Yeah. We we use a whole mix of models for different things because we found that different base models perform differently for different tasks. So you know, the the the Google Gemma models that were just released are great at very specific things. Lama models are great very specific things. The Quinn models are great very specific things. So even for uh you know one customer service, you know, one message that comes in, it's not one message goes to one model and that's it. It's one message goes to you know a whole sequ of models, each of which are good at doing different things to finally generate that uh that that final response to the customer. the other thing that is also like has been uh debated is, you know, for anyone who's sort of like fine tuning or even doing more of post training things or like doing specialized things for their category or for their industry or you know something that is very you bring unique data and do more post training on top of existing models. The argument has always been like you know the bigger models will get those capabilities you know as we train bigger and better models whether it's GP5 or Gemini uh you know like GP5 is going to be slightly better thinking in you know some other models whether it's something else like what what has been the reality and what you're seeing because you guys really are looking at you know specific things that you want from all of these models for your use case. So, I actually would push back against that argument and there's two reasons why. Um, so number one, this idea of, yes, the bigger models are all going to have these capabilities. Um, I think they'll all have the capabilities, but I think the level of performance will change. I think it is uh if you have a well-defined task, you can have a model that's like a hundred times smaller, have a higher degree of performance if you just fine tune it on that one task, right? You're saying, you know, I don't want you to know how to, you know, code in Python and write poems in uh you know in 20 different languages. I just want you to get really good at this one task and then when measured at that one task, it will probably outperform models a hundred times its size. Right? So that's number one, which is you can actually get to better performance, right? Like maybe you know uh this will do better than uh GPT 5 today and when GPT 6 comes along it might be a little bit better, but as of today it's going to perform better, right? Based on what we what we have available. And number two, which is actually equally, if not slightly more important is latency, right? If I have a giant model, it's going to take a lot of time and by a lot of time I mean you know maybe 5 seconds to to kind of generate that response. Whereas if I have a really small model, that time is cut by a factor of 10. And there are certain uh you know in in certain situations like when you're chatting uh over text, 5 seconds might not matter. But if you're on a phone call, if you say something and it's silent for 5 seconds, that's a really bad experience. So we care about latency a lot as well and and and for that you you kind of have to go to the toward the smaller models. Someone said to me this like uh you know what a system prompt really is basically coming a larger model into a small model because you only want you get about certain parts of the model. So I thought that was an interesting phrasing. Um can you talk a little bit about you know why uh you and your co-founder they customer service as a segment when you you know decided to like uh you know start a company? Like why customer service and like what led to in this direction? Yeah. So um when we when we started this company, you know, this was around the time when you know G3.5 Turbo was up, GPT 4 was up. we were looking at the capabilities of these models and we're like, wow, this is now getting just about good enough that I can start doing things that people do, right? And as we looked at the enterprise, we were like, where where is there a lot of uh uh demand for like text-based task, very repetitive text based task? And and we were like, number one is customer support teams and number two is more generally like operations teams, right? Where you have large groups of people just doing repetitive work. Uh and as we talked to a lot of these like uh operations leader leaders, the number one demand from them was in customer support, because they were like, look, we are, you know, we're growing so quickly, because of that our customer support volume is scaling really quickly, which means we need to hire a lot more people to stop all this demand and we can't afford to do that. We are desperate, we really need something here. And initially, you know, it looked like it was a very crowded space from the outside. But as we went in and started talking to a lot of these customers, the thing was, it was very crowded for smaller companies, like smaller companies that were buying that had simple tasks that, you know, 90% of their incoming volume was, hey, I want to return my order or something like very simple like that. But for more complex enterprises, like there wasn't anything that had been built that could kind of really follow their, you know, very intricate support flows. And so that that kind of wedge in was what we took to say, okay, we're just going to build exclusively for companies that have these very, very complex workflows. Uh and number two, the other thing that was interesting was, you know, longer term, the way that we were thinking about this was, if you build an agent that can instruction follow very well, the thing that you enable businesses to do is start saying, okay, today I'm going to use this for customer support, but I can eventually start growing this into a customer concierge, right? And then and then what I mean by that is, let's say you are uh let's say you want to fly from San Francisco to New York next weekend, right? Um, what you would do is, you know, you would go to your favorite airline's website, you would type in I want to go, you know, San Francisco to New York, Friday through Sunday, um, you know, one ticket, you hit search, it gives you, you know, 30 different flights that you can pick from. You look through everything, you're like, okay, these two are the right time, but this one this one's too expensive for me, you pick one, you pick your seats, you fill out your information and you hit purchase, right? That's that's a lot of steps, that's annoying. However, a much better experience for you would be if you had your favorite airline's phone number that you could text and you say, I want to go to New York next weekend Friday through Sunday. And they had an agent on the other side, an AI agent that it knows who you are, it knows what your preferences are, it knows what times you typically like to fly, it knows your budgets, it looks through everything and says, hey, here are two options, like which one do you like? And you say, oh, the first one. And this AI agent also knows where you like to sit and then it just responds and says, oh, okay, by the way, I have a free upgrade available for you in premium economy in the window seat near the front. Is that okay? You say yes and it says, okay, booked. The big difference is, one, this is a much more seamless experience, and number two, there's this shift that happens, right? Where most websites and web apps today shift the burden of work onto the user, right? They are it basically gives you all the tools for you to book a flight and then you have to go do all the work to do that. Whereas now it shifts to this world where you express your intent to an AI agent that then does the work for you. And that was a really, really interesting shift for us and you know building these customer support agents, uh in our view is a first step to building these broader customer concierges for all companies where this can become one of the primary ways in which now customers interact with your with your business or service. I think this is almost um the extension of the prediction of almost like, you know, at some point we start to start to that Prime will predict what you will need in future, right? Uh so it's almost the same thing because you've seen like even in the example that you've given you know exactly you know what are the things other customers, you know, if you if you're an airline company, you know exactly what are the patterns that the other customers have taken and you learn from those patterns and you can almost suggest what are the two or three things and I can just you know it shows up as a notification and I can select between the three options. I going it's almost a little bit scary but also like yeah, you know, there's also a predictable nature of finding this patterns uh because you know there are repeatable customers already as you as a customer has done this before. So the airline company can basically tell you what are you going to do next and basically suggest your options based on that. In some form we're getting into that place. I I I I would think of it more as um instead of predicting what you want, it's more of learning what you like and remembering so that your experiences are more seamless, right? So instead of you know guessing I want a flight, it it knows that, hey, you know all of these 10 flights that are available, Ashwin typically does not like taking red eyes, so I'm going to remove these. He doesn't like flying at 6:00 in the morning, so I'm going to remove these. And these three are way too expensive, so I'm going to remove them. And okay, there's two options and Ashwin probably liked both these, so let me just recommend them to him and see which one he wants. So it's a lot of this manual work that otherwise I would have to do that gets kind of done on my behalf. Um so I I I think there's going to be a big shift um, you know in terms of the way that businesses kind of design websites, design web apps, to rather than saying, here's a bunch of information, you user have to now make all the decisions to, hey, you're going to like tell us what you care about one time and then our AI agents will remember and then work on your behalf to, you know, just tell us what you want and we'll do the work and kind of give you the service that you want. Can you talk a little bit about, you know, we talked about like how you got into like customer service as a space, but you know, how did you acquire your first say five customers? Yeah, yeah. What that journey look like? And how fast did you convert them like a free design partner to a yeah pain customer? Yeah. So, you know, early customer acquisition is always uh it's always very manual. There's unfortunately no, you know, silver bullet and you know fun hack that we did to to scale. It was just a lot of, you know, find everyone that you know uh in your networks, get introductions to good companies. Uh we did a lot of cold emailing and cold LinkedIn messaging. Um it's it's just um it's just brute force work to do it. Um but the other thing for us is, in the early days we never did free design pilots. Um we charged for our software from day one. Uh now, of course, this doesn't mean, you know, we charge them on day one of the contract. We typically said, okay, there'll be a, you know, maybe a four week pilot that we'll give you, you know, because we're such an early company. We're going to give you a four week free pilot. and at the end of four weeks we're going to decide up front, if you like it, this is what it's going to cost and this is, you know, how much you'll pay. So we never had an open-ended uh long-term period where we did things for free because if you if you recall in the early days uh the number one thing that you were trying to validate is, am I building something that people will pay money for, right? And you might not have built the product, uh sorry, you might not have built the product entirely, but if this is something that's truly valuable, you should be able to talk to your potential customer and say, hey, if I accomplish ABC, will you pay me this much in, you know, four weeks from now? And that should be something that if it is truly something valuable, they should say yes to. If it is truly a problem that is so painful, they should say yes, if you can solve this problem by doing ABC thing, I will pay you and I will pay you this much. So we were what what this does is it removes the cohort of people that don't really care that much about your problem, because if you ask them to pay you a reasonable amount of money, you know, a reasonably short period of time, the people that actually don't care about the problem will just quickly tell you they'll say, okay, actually, you know what? It's not that much of a problem for me. I don't need this. So that helped us kind of weed through um, you know, bad business models and and and bad initial ideas quickly. I wanted to talk a little bit about the business impact for your customers, you know, when you are using Decagon, right? you mentioned, you know, for the initial customers you promise, you know, we'll deliver A B and C. Um, and you know, if we were able to do that and we'll you know we'll propose some kind of pricing model. How how do customers now look at um, you know, what is the success metric that you know your customers look at or you look at whether this customer is being successful on uh Decagon or not? Yeah. Um, I think it's I I think customers think about value in two ways primarily. One is, what percentage of conversations are we be able are we able to handle ourselves successfully, right? Which is not I'm going to frustrate the user, but the user is satisfied and we've actually solved their problem, right? Because if we can solve a greater percentage of those things, fewer percentage of, you know, a fewer number of support tickets ultimately make their way to human agents. And now these people can focus their time on the more complicated problems, getting them support faster, things like that. But the second uh benefit actually that has been uh a little counterintuitive to us and was somewhat interesting was, a lot of these companies then expanded the amount of support that they offered, right? Which is before it's not that companies want to minimize the amount of support that they offer, right? Like they want to give as much support as they can economically, right? And so before if it cost me, you know, $10 for one customer interaction and all of a sudden that has become 80 cents, instead of uh I as a business, I'm not saying, okay, now I'm going to, you know, uh uh just save all that money, but I'm going to reinvest some of that in providing more support to my to my end customers. And what we've noticed is, customers uh their end customers actually want that increased level of support. So now rather than saying, okay, you know, support is only from the phone lines are only from 9:00 a.m. to 5:00 p.m. Monday through Friday, it becomes 24 hours a day. Instead of saying we only offer uh uh you know support to paid members of this tier, it's we offer support to everybody, right? So there's there is this kind of latent demand for increased support. And by making it much cheaper, businesses can now offer this more because at the end of the day, this leads to, you know, higher retention, like better customer happiness. And like every time a customer feels a slight paper cut, if you're there to like solve their problem quickly, that customer is likely going to retain longer, stay longer and be a happier user. I think uh support also is like a brand distinction, right? Like a lot of companies really take their support very seriously and provide extraordinary amount of support and you can almost read the success of Amazon to a little bit for this, right? and because because of their extraordinary investment and support. Uh you can quickly get access to Amazon support versus like you know try try getting support from a different e-commerce website. So that like support has been also like a very big brand differentiator, especially as you go to price point increases, uh support actually, you know, becomes much more uh almost like a uh thing that builds your brand and customer trust over long term. Um I I I agree. It's like especially to take that example, right? Like I think Amazon's a great example because I feel very comfortable buying things on Amazon because I know if there's a problem, Amazon will take care of me quickly without any hassle, right? Like if I have an issue with something I bought, I know I two clicks and they will refund me the item. Uh I think another great brand uh example this is American Express, right? Like American Express is known for incredible um customer service. So yes, completely agree that I think this will increasingly become a big part of like brand differentiator, which is if my customers have a problem, I will be there, I will fix the issue. And I think it makes people feel a lot better. Yeah, absolutely. Uh so the the other thing that you have support for is voice agents, right? And I think voice agent is a particularly interesting one um because you know the models are getting really good and good at uh identifying voice and creating voice uh in the last couple of generations, especially in the last three years. Uh um So talk a little bit about, you know, how has the traction been with voice agent? Uh are customers complaining about voice agents? Are uh customers praising voice agents? What has been the feedback like? And and do customers realize that they are talking to because now certain we're almost passed the tuning test of you know certain voice agents are really, really good in terms of imitating human behavior. Uh so like are there any specific techniques where you, you know, sort of pushing to the models saying that we can identify this as a non-human because that's also part of like customer trust, right? You also don't want to tell them then you you're speaking to a human while you're not, right? That's because some find it wanted. Yeah, no. So, you know, in general our all our voice agents say hi, I'm uh and a virtual agent here to help you or something like that. But I think the other interesting thing is, most customers that are calling in about a problem, they they don't want to talk to a human, they want their problem solved, right? Like they don't care how they want their problem solved, but they just want the problem solved. Um, and for us making it sound more human is not because we want to, you know, give the user the impression they're talking to a human or anything like that. It's to make the interaction feel more seamless, right? Like you want the responses to be fast because it's kind of annoying to say something and then wait 5 seconds to get a response back. Um, but at the end of the day the primary goal is just how can we solve the customer's problem? Because even if the if the customer, let's say the customer is very aware that I'm talking to an AI agent, but that AI agent solves their problem in 10 seconds, that's a good experience versus if I know I'm talking to a human and this human takes 45 minutes, that's a bad experience, right? So the the the main factor here is not whether it's human or AI, it's can they solve my problem for me? Number one and two, were there a lot of paper cuts, you know, did I have to say something and then wait 5 10 seconds every time. So if you can make it a seamless experience and the problem gets solved, I think at the end of the day customers are happy. And and we've actually seen this. Um, we have um, actually several customers now where the the NPS for their voice agents is as good or higher than human agents, because if the voice agent can if the AI agent can solve their problem, it solves it immediately. and if they can't, it hands it over to a human immediately, right? So either way you you end up having a reasonably good experience. I think there was this uh interesting case where Crowd Madded uh where they almost removed the entire customer service department but then they report it and they said they went too extreme right? Has there been like a drop in hiring in customer departments or support departments or what has been the trend in terms of like pure you know agents versus human um you know our agents replacing humans or are we adding it to it or are we getting I mean I think you're talking about this like increasing support but in general when you know your CEOs of the companies are looking at their you know how much they're spending on support, are they looking at you know staying flat or already using the account or increasing it? Like what has been your experience? Yeah, it it it really depends on the business and like different people do different things, right? Because if AI agents can handle a bigger chunk of customer uh inquiries, then you can do one of three things. One, you can say um look, I can now handle more incoming customer support. So I'm going to keep everything else constant and I'm just going to handle more support volume, right? I'm going to put it on every page, I'm going to give support to every kind of member, I'm going to do it 24 hours a day. So your top line amount of incoming