LIVE: Top 10 Things We Learned from the New SaaStr AI with Jason Lemkin + Assistant UI

Blue Sky, whatever streaming platform you're watching. Welcome back to Saastr Workshop Wednesday. We're going to have some fun today. We are, we have a great guest and a great theme. A lot of you have noticed that we have launched our own AI, our own Saastr AI, and it has been trained on a decade of our content and hundreds and hundreds of CEOs and speakers. And it's only been up for a month, but we've already had over 6,000 sessions on it, which is pretty crazy. And I wanted to do two things. I wanted to share my learnings because honestly, it is—

I'm—

to be honest, it has opened my eyes to how to think about what AI can do in the real world, not just in SF pitches and even for coding and programming, but what it will really do for all of our apps in the real world. It's opened my eyes. So I wanted to share those 10 learnings. And then I was so excited about what this did. I was lucky enough to recently invest in a company called AssistantUI, and we have the CEO and founder here. Simon Farshid, who is— has the leading platform to add a ChatGPT-like interface to your own app, right? So I'm gonna have Simon here. I'm gonna introduce him in a second, but I'm gonna have him chime in on all my learnings, um, and dig deep a little deeper on the application technical side. But he's got— I'll show you in a minute— he's got, uh, 4,000 GitHub stars and a lot of great early customers on this and a lot of traction. So he's seen how these— this framework works and how you could add it to your own app. So Simon, thanks for joining us.

Nice for having me.

Yes. And I'll introduce you a little bit more of the app in a second, but, and we'll share our learnings, but you guys probably saw this and I want to frame this slide. I know I talk to a lot of B2B founders and it's really only recently they're coming off the sidelines in AI. A lot of folks are talking about hallucinations and I want to show you what we've learned in a minute. They're talking about security concerns. They're saying that there's, that our customers don't need cute cartoons. But I think the skepticism is fading away. And I think everyone's realizing whether they're in a space like support or coding, which have already been profoundly changed, profound, you know, almost half of support is done by AI in many cases. And a lot of leaders, as we've talked about, are now completing about half of their programming in, using AI tools like Cursor and Windsurf and others. Last week we showed, you know, one of my fastest growing investments, RevenueCat. They've doubled their pull requests using AI tools, right? So that's already happened. Others we've been skeptical. I don't— I think folks aren't skeptical anymore. HubSpot has launched its own tools. Everyone has. But to summarize, this is kind of crazy from Sam Altman. ChatGPT was just 26 months ago, 1 million users in 5 days, now 1 million in an hour. I'm not sure how many orders of magnitude that is. Simon, can you help me calculate orders of magnitude? It's a lot of orders of magnitude difference, isn't it?

It's quite a lot of orders of access, 3 or 4 or something. Yeah.

So wherever you are, I think by learning from the Saastr UI, wherever you are, there's something. So I, when I put this together, it was 5,200 messages in a month. Um, we precede, if you look on the right, we precede it with how do I hire a great VP of sales? Should I come to Saastr Annual? How do I get funded? But people ask all different things. But if you think about it for a minute and we'll dig into it, we have almost 6,000 folks in less than 30 days. I could never have this many, Amelia could never have this many conversations. Simon couldn't, right? So if nothing else, if you think about the scalability here of AI, and we talk a lot about replacing humans and efficiency, but in many cases it's about doing things there's no human to do. We don't have anyone at SaaStr that can have 6,000 founder conversations. Like we don't have any, but so, but this is a lot of data and a lot of learnings. and then we're going to get to it in a minute. But Assistant UI, just to frame the conversation, Simon, tell me what you've learned, what's changing, what's going on here, what have you learned from your customers in deploying this type of interface in their own app? Give us kind of a 60-second summary of all the learnings so far.

Totally. So Assistant UI helps developers add AI chat interfaces in their apps. These interfaces look like ChatGPT. And the developer can go and customize and add their own custom elements to the chat as well. When I started this, I wasn't sure if there's gonna be any market for this, but it turns out that the chat interface is really useful for agentic applications. And that's what's been fueling the ChatGPT growth as well, where it's like, initially everyone was like, okay, it's nice and fancy, but how big of a use case is this chat gonna be? But it turns out that— you can do everything from image generation to controlling, uh, AI agents just from a chat interface. And it turns out to be a pretty decent, uh, user interface for that as well. So the market ended up being much, much bigger than I thought for a product like this. Initially it was just gonna be GPT wrappers, but, uh, ended up people building, uh, a complete AI agents that are controlled entirely by a chat interface.

Cool.

It's been a really fun journey. There's all sorts of long-tail use cases that people use this product for because it's open source. And then I've been doubling down in a few niches, mostly around pulling agents and adding agent UI in applications, as well as a little bit in commerce and support use cases.

Cool. All right. Well, it's pretty cool. There is an open source version that you can, you can use today, right? I mean, you can use it all today, but if your team, if you're thinking about adding, like, I think I wasn't even sure about this 30 days ago. I think every single person right now should have some sort of ChatGPT-like interface, some sort of Copilot, some sort of app. I don't care whether you're selling to salons or dog walkers or whether you're selling to the techies. It doesn't matter. Everyone, this is the way we don't want to input data anymore. We don't want to hit refresh. We don't want to figure out the, the 11,000th UX that we don't understand. Like, you've gotta have this at least as an alternative. And again, literally the fact that 6,000 folks went to just SaaS websites and did this in one month shows you this is— so take advantage of Assistant UI, check it out, have your team check it out and challenge yourself. Can you just launch something this month? This is the— I hate this term, cracked. I hate this term cracked, Simon, but this is the area of the cracked engineer. Like, Don't wait to launch your AI for a year. Like, use some— look at Assistant AI and other frameworks that are out there. You don't have to do it from scratch. You don't, you don't have to build— you certainly don't have to build a model from scratch, right? You don't even have to build the chat framework from scratch, right? If your engineers care, they can get this done in a week. You can get something up in a week. So do it, because people— you're going to look pretty old school if you don't have anything like this very, very quickly. Um, so just on how this works, our app, we use this app called Delphi. Ai. Lenny uses it. Brian Halligan, founder of HubSpot, uses it a lot. A ton of folks with a lot of, a lot of reach used Delphi. You can check it out. And it's pretty interesting. And so here's a quick look at the backend that we've done. It's trained on— you can see it here. I know it's probably small on folks' screen, but you can see it's trained on 17.6 million words. I think that's a decent amount of content. It's 10,000 posts, 2,100 YouTubes, including like, you know, CEOs of wherever, you know, Asana, HubSpot, Box, whatever. All of their SaaStr sessions, it is trained on 800 pods over the years, almost everybody, and 100,000 tweets. Like, it's a lot of content. And I want to— I want to get into the 10 points, but I think my meta-learning here It's one of the points is how important training is, because it's pretty good if you talk to something in these 7.6 million words. And I'll show an example in a minute. If you— my learning is, and I don't know why exactly, I sort of know why. If you just randomly ask Claude or ChatGPT about this type of content, the answers are pretty mediocre. But if you do train your, your app on 17.6 million, it's pretty epic. And this took Listen, I'm not an expert in how long ingestion takes. Maybe you have some thoughts, Simon. It took hours to set this up. It's a lot of content to ingest, in some cases to scrape. I don't know exactly how it works in the backend, but once it works, it's pretty powerful. This is how it works. Simon, anything you want to add on that ingestion or that's helpful on this process?

I think it's a very good thing to point out that all of these AI chat interfaces needs quite a little bit of data to work. So for Delphi's use case, uh, you're in this privileged position where you already have, uh, 17 million, uh, words about yourself, uh, on, uh, readily available to ingest.

Yeah.

Um, if you want to add some sort of AI assistant, uh, about your app, there's a likelihood that you don't have 17 million words about, uh, your product. The way you collect this, uh, is to just launch something really quick and collect data early on. Because the highest quality data they can collect is the user questions and the user conversations people are having about your product. Those are like the perfect set to figure out what questions your chatbot really needs to answer.

Yeah, the questions, it's a good one.

The questions occur, right?

90% of the questions are the same variants of the same 20 questions, right?

Yeah.

Yeah.

Yeah.

But unlike— what's fascinating here, unlike a wiki, which can just list the same 20 questions, if you train it right, it's— and you ask variants of those questions, it's really good at giving you the right answer, variant answer. That's right. It's pretty good at tailoring it. We'll have folks that come into our AI and say, listen, I'm at $3.6 million ARR, growing 34%. My customers are barbershops and tailors. My deal size is $5K. Here's the issue I'm having. And it can tailor the response pretty damn good to that in a way, you know, a knowledge base never could. Never could.

Yeah, exactly. And one more thing to point out is that like 90% of people, everyone's questions are going to be the same, but everyone asks 10% different questions.

Yeah.

And the magical experience is when you can answer those 10% as well.

Yeah.

So that's where having more data helps a lot as well.

Yeah. And it is interesting. I saw this in, I just wrote up Fred writing on Freshworks at $800 million in revenue, which is one of the leading support and service public companies. And they talked about their AI and they said something related to this that I've also seen a gorgeous ROM on the board which has 20,000 SMBs doing AI support. Uh, do it. This is work, right? This was a day, right? Brian Halligan from HubSpot trained his AI. He didn't have 17 million words, so it took him a week of training, right? We'll talk about that. Freshworks said, you know, and Gorgeous sees this too, the bigger the customer, the more successful the deployment is today because they have people to do the training, right? A lot of SMBs have nobody. They just expect it to work. AI, a lot of it like this still took a day, right? This did not— as great as it's cool as this is, it does not work 100% out of the box. Like I literally, I have a different slide. I literally tried a similar tool that said it would do everything. Like I didn't even have to set up the RSS feeds or the links. It didn't work.

Like we're not at that point. That's the hard part in DevOps. Yeah. Yeah.

So it may take you 30 days to fully nail this ingestion and training process, right? So you do have, you do have, we have to commit to it, but, but if you do, it's magical. So this first one, guys, this is going, Simon kind of made a version of this point, but this is really powerful for your apps or whatever you do. Folks will talk to our AI about things they won't talk to a human about, right? Here I just did a search on everyone thinking of quitting. Quitting. Simon, you're a repeat founder. You ever had a moment where you thought about quitting? I have a brief moment. Never.

Couple, couple of times. Every now and then. Every now and then.

Did you have anyone to— honestly, did you have anyone to talk about it?

Sometimes, yes. But a lot of times it's also been journaling. Yeah.

Journaling.

Yeah, totally. These AI chatbots are a tool for thought in a lot of ways.

But we've had a lot of conversations over the years on this. I've written about it. Ben Chestnut from Mailchimp has talked about it. These ants, this is pretty good therapy. This is pretty good startup therapy right here. Right? Amelia, have you ever thought about quitting? So you should say no. Have you ever thought about quitting? Once in a while for a brief moment, once or twice over the years? No, I'm not quitting, but I think there's times where it's, you know, I mean these, I think I've seen the backend of our AI bot, right?

So I know it a little bit more, but like, I think it's, it's more cathartic that they can talk, like they can't talk to actual Jason M. Lemkin.

Like, you know, our team can talk to actual Jason M. Lemkin, like, cathartic that they can talk to a version of you that's pretty good. And based, I think your AI does pretty much tell people to keep going, right? Like to go long and it takes a while, but in those moments of weakness, it's nice that they can talk to digital Jason. It is nice. I think at the app for sure. And I think at the application level, my, my learning, my, my, which I didn't know before is if you do this right, if you make the interface really good, um, people will talk to you about things they otherwise wouldn't. Right? I've written a lot on SaaS about how we just hate talking to customer success now. They show up, they never want to fix our problems, and they just want to do upsell. We have no one to talk to at our largest vendors. We have no one to talk to. If your, if your AI interface can really talk to you and work through these issues, that's very powerful. That is an example. We would love, we have two vendors coming up for SaaS Annual. Everyone come May 13th, 15th. The SFA, we have two vendors with problems. We can't get anyone to get our help. I would love to talk to an AI that is so good. Not only can it help solve my problems, but I would want to talk to it, right? So just look at this. These are conversations that, that would, that would never happen.

Um, another version of this is like, so if you add the sort of chatbot to your B2B SaaS app, there's going to be support requests and issues and problems and customer insights that you would never, that nobody would ever tell you directly, regardless of like how many customer conversations you're going to have.

Yes.

So yeah. Yeah.

I think as we do this, as we all get better, it is exactly true. This will be in many ways like Gong on steroids, just like from Gong, which was so early to ML. We would learn what our sales team was saying. The right interface here should tell you things about your customers and products. Otherwise you could never learn. That's a good point. You could never like go out and meet them in person. This is a SaaS thing. You've got, you'll learn things in person you'll never learn, right? But you will learn if you nail this interface, you will learn things from your customers you would never learn from a human support that takes an hour to get back to you or that doesn't want to solve your problem. You will learn things that you will never learn, right?

That's right.

This one's pretty crazy. So, um, uh, our— obviously the Saastr AI is trained on everything I've said or talked about startups for 12 years. So, um, this is from Owner, one of the fastest growing companies at Saastr Fund. We had it— we had all the board and investors provide feedback on their investment deck and the AI won. The AI won. It beat me, the real me. It beat Redpoint. It beat others. It had the best feedback. And, um, it did because it basically— for what it's worth, it basically said what I would have thought, but it didn't forget. It remembered 12 years of my thinking, and it had the best feedback on TAM, team positioning, competition, uh, markets. It was, it was really, really good, um, and much better than I would get out of a generic No matter what it was, I get out of a generic Claude prompt or input, but it won. It won this, right? So if you have— folks have already done this. If you have a deck for investors, upload it and ask the SaaS AI what it thinks. Is this— is this what you'd expect, Simon? Is it crazy? What do you think about this?

I totally agree. I think there's another version of this, which is like you're probably not going to work as hard as you could 24/7. And having an AI, it never loses focus. So it may perform better than you do 95% of the time, where if you really, really, really cared, you could probably do a better job than the AI. But most of the time in the average work you do, you don't do this.

You don't. Yeah, it's a good point. It's, it's pretty, it's pretty, it's pretty disruptive. I didn't expect this. This is better than us. This is better than all these fancy VCs. Not entirely, but overall it was— this was just from the other day from our AI. Okay, this one I think is worth spending a few minutes on because this was a learning for me, right? Everyone, especially folks that are not particularly technical or do not consider themselves prompt engineers, remembers the first time they fired up ChatGPT and we just all made fun of hallucinations, right? In the very beginning, we— anything we— it was just— we would just get crazy answers, right? And some of that has faded away, but so many founders I talk to are still skeptical about AI in their application because of hallucinations. And obviously the issues are still there. They are going away. Here are my meta learnings, and then Simon, tell me what you see. There is some— there are— I would say hallucinations are roughly— I don't have the exact data— less than 5%. I would say 80% of the time our AI is really good. It's really good. And I'll show you some examples. 15% of the time it's meh, right? It feels a little generic, right? For whatever reason, maybe there aren't enough of those 18 million words to synthesize. I don't know why 15 are meh, but it feels like something anybody could write, right? Not— that's where it's unspecial and anyone can do it. And 5% of the time it says something wrong, and it's not that common, but I pulled two examples. And it may be how, how it's interact— how it's interacting with the data you ingest and the APIs that it uses from the models. The one at the bottom left was someone asked for Saastr Annual, where, where they can get a discount. And it added a strange answer that it would be at the Hilton Union Square, rebranded Saastr Tower. That was 2018. So mostly it's done a good job in support. I'll show you an example in a slide, but this was pretty bad telling people to go to the venue from 2018. I give that an F, right? And then this really weird thing. People ask sometimes creepy things about me, right? But they ask about personal details. This entire story is manufactured. This whole story about me that I'm about work-life balance. I have 3 kids. I don't do emails when I'm on vacation. When I'm on vacation, I, I like to sit at the pool. I'm, I'm, I'm on social media all the time. Okay. And then this crazy story here, which is somebody else's story. Um, this aunt's play and work and teaching a 3-year-old to ride by visiting my grandmother and fast food. None of this is true. And I hate skiing. At the bottom. It says I'm passionate about skiing. I mean, listen, when I was in college, I pretended to like skiing for the social reasons, but I don't want to spend like a whole week skiing. I got to build stuff. Simon, we don't have patience for skiing, do we?

We don't have patience.

Right. So, so these ones are crazy. But I guess, I guess this personal stuff is, is it's making up and this email address is wrong. But I would say these are the only two exceptions I can find. These were the worst ones I can find. I'll show you some others that were like A tier. But what's, what's the current learning on hallucinations that you see, Simon, from the field?

Um, I think you can classify hallucinations in those that actually cause any harm and those that don't. Yes. And the user interface also really matters in terms of like how susceptible to hallucinations you are. So in a chat interface, you always have a human in the loop. Every AI response is read by a human. There are no automated actions being taken. So the worst thing that can happen is someone misunderstanding something. And like the worst, worst case is someone going to the wrong venue of SaaS. That's embarrassing.

Yeah.

But you can probably classify what like the overall shape of the risk is that like what's the worst thing that could happen and make sure that for those responses, the AI is doing a good job. And then when you have a chat interface, if the AI hallucinates random facts about you, it's probably not that big of a deal. Depends on how often it happens, but like, as long as the harm isn't that bad, like the cost of hallucinations in a lot of times in chat applications isn't that high because the end user ends up understanding at some point that like, wait a second, this doesn't add up. Yeah.

Let me ask you a different question. So, so this, these were, it's, It has exceeded my expectations. The hallucinations are rare and they're usually random things are off point, right? That they're off point like personal life or when people ask random things that, that probably aren't part of the 20 million words that we've ingested. But at a lay level, I think it's less than 5%. But do you think we'll get it down to less than 1% in the next 12 months? Do you think just using the tools and frameworks we have, it will go down almost another step function or magnitude? Or do you think it will— we'll have to just live with these types of issues?

I think it will go down. Like, if you look historically, the hallucinations have gone down tremendously.

Well, that's what I mean. Absolutely. Last 9 months, down, down dramatically. Right.

And like most hallucinations happen around the AI dealing with your own custom data. So like ChatGPT used to hallucinate and I can't think of the last time where ChatGPT in its generic form hallucinated. To me because it has so much training data. Um, and then like the only hallucinations I see is around RAG where the AI agents, instead of like taking the data that you give it, uses a generic answer. And so like that's a class of hallucinations we're currently seeing. But all the labs have— that might be a little bit more familiar on the family stuff, right? So I, I totally think so. I totally think that, um, this number is going to go down and the problem will fix itself.

Yes.

And there is going to be some point in time where we can finally say, okay, we can actually trust the answers of this AI to the same level that I would trust the real Jason to give me an answer. Yeah. And that will prevent, prevent us from having to double-check every answer. And it's going to be a massive step change in how we use these AI assistants as well.

That's a good insight. And maybe that's how I can wrap this. I picked these two because they're wrong, but they're the— but it's been, you know, I can't— I'm not privy to all of them, but if it's 2 out of 6,000, it's not bad, right? Less than 5%. I actually think my shocking learning is not only are the hallucinations uncommon, right? In most cases minor, right? But I think for this AI that's trained on SaaS information, it's already at your point. It's right so often that I don't believe we can do 6,000 a month without any human review. I think it's— I didn't know it would be right this often. Right. That's my epiphany. If it gets just 50% better, I think it would— these then would seem like anomalies, right? They're close to anomalies. Right. Okay. Here's the next one, which I admit I did not get until I was on the other side. This is a weird paradigm. Okay. But I'm combining— Brian Halligan, chairman of HubSpot, has done one as well, and he added video. I've decided for the moment not to add video. I think I will add video. We could talk about why. It's pretty interesting. I don't know what API it uses. Simon, maybe you know, but it can see you when you're on video. It can see the room. It knows what sweater you're wearing, who you are. If there's— if Simon and I were both in the room, Digital Brian would talk to both of us. It's pretty cool, right? I know this. You know this. But what I found, this paradigm is when there were 3 of us. What I mean is, if it's me and Amelia in the office at Saastr and we bring up digital Jason, right? Or me and Simon bring up digital Brian, and Brian can listen to us in full duplex and participate in the conversation, then we really have this, this IRL copilot, this assister in the real world. And it's really— so like, I can't remember everything that I've learned in 12 years. I don't know if you can. I can't remember all the 1,000 speakers we've had and guest posts and writers, but you know who can? Digital Jason. When the 3 of us get together, it's crazy, this 2 IRL plus digital. So this is something that I know most people aren't building in their apps today as a core paradigm, but I think this is crazy how effective this is because if the digital, if I'm 100% right in my mind and Simon's in the room too, and then Digital Jason's 95% right about 12 years, Like it's, it's, it's a superpower. This is the Mac for humans. This is. And so what I'm wondering is I'm actually shocked that we don't already have this for sales calls. For real. I'm shocked that we don't already have this for SDRs because I think by the end of the year, every single sales call should not have a note-taker. Well, there's, there's probably tons in this. Every single sales call should have digital Simon, digital Amelia, digital Jason on the sales call Clearly digital, we're not playing any games. But when I forget how the feature works, when I don't know how the assistant UI integration really works with Deep Seek 3.6.4b, when I can't answer that question, Digital Jason should honestly answer it. I think this will utterly change how we do sales unless I'm missing something. So I'm shocked in fact that we don't already have this because I think we're almost, almost there. So any learnings from this? Is this a, this paradigm, Simon?

Um, we talked about therapy and, uh, there's probably a reason why most therapy sessions are over voice and not over text. Mm-hmm. That's just because the friction is so much lower. Uh, and so there's an effect of that where like, make it 10 times easier to say something. You again unlock conversations that would not happen otherwise.

Yes.

And so you're basically exploring a different idea space through the law of friction. I can't, on another note, I can't, I've lost count of like the number of sales calls I've had where the person I was talking to didn't know an answer to a small technical question. Yeah. Something that even seemed like obvious. Usually happens with larger companies that are no longer doing founder-led sales. But it's like the most typical problem that you get when you try to build a sales, a reputable sales motion is like you need to train your sales employees to be able to answer questions. It's so frustrating as a buyer to— It's terrible. Yeah, it feels like you're not really sure if the person really knows what company they're working for. So having a sort of like Copilot in place, which can immediately answer all sorts of questions about the product. would completely solve this problem.

Yeah. I think every human interaction in business that is on the internet at least should have a digital human, a digital person with it that knows all the answers. Every single one. Every single one. I don't, I, in fact, in fact, I'm shocked now that I've done this with our own AI and I've seen how good it can be if it's trained. I'm shocked this hasn't already happened, right? There's so many, we're, we've totally changed so many parts of software already with AI. I am shocked that every sales tool doesn't already have this functionality and that it's not pretty good. Because Brian and I, Brian Halligan, me and somebody else together, we're going to close the deal, right? There's crummy, like, sales solution engineer tools that are like wikis on steroids. There's nothing wrong with them. But anyhow, if guys are— folks are thinking, and you have any, like, human, like, internet component, I, I think your app will be better with this. They will get nothing wrong. Everyone should bring Digital Simon to the meeting. This was— this is my aha moment, right? Okay. This next one, you talked about how therapy folks prefer voice, right? This is just a little bit of data, but I thought this was interesting. A lot of folks on social media talk about how the future is all voice with AI. I'm— I've, you know, I've— I was the first investor in Talkdesk, which is worth $10 billion. That was redoing voice for the contact center. I've done chat. I like— I get it. I've been passionate about voice. I'm passionate, but I think we, we might also be in a VC overhype frenzy where there's a different medium for a different use case, right? So this is just what we learned for SaaStr over a couple of weeks. But if you look at it, web was 2,000, voice was 500, and then embed, which is, which is, which is really just chat, but on our home page is 2,500. So that's only 20% voice, 80% typing on a keyboard. It's something to bear in mind if you think Again, customers are very different, right? I've invested in, you know, one of my fastest growing investments is Owner for restaurants. No one in Owner is like, is on the web all day. Like they're in the restaurants. You better either get them through text or voice because, but I think this is telling how people want to interface with AI today. What have you learned, Simon? I know you're adding more paradigms to your framework, but what have you learned on this voice versus chat versus et cetera.

Um, because I'm a developer tool, developers can add more paradigms than I even anticipated. So I have a voice mode and a text mode. And what I've also seen implemented is this sort of like unidirectional voice mode where the input is via voice, but the output of the AI agent is via text and, uh, interactive elements. And that also seems somewhat promising. I do wonder whether the friction around people using voice is mostly around us not being familiar with it. Um, because when I talk to humans, most of the time I use voice. I don't text them, especially if they're in the same room.

Yeah.

But with a computer, we seem to prefer to type instead of using our voice to talk to them. And I do wonder whether it's just a habit thing, uh, or whether there's, uh, any actual, um, benefit to text. Text lets you think harder about what you say and be more explicit and think more deeply in some ways. So again, I think both paradigms make sense. Yeah. But I'm not sure if like the, the ratio that we see right now is going to last, is, is the optimal one, or whether we are going to shift to a world where we mostly talk to our computers.

Yeah, I don't know either. I think, um, I just think this data is interesting for this type of use case, learning, right? Um, that it's— that voice matters, right? This is a, this is a, a multi— multimedia, multi-paradigm, right? Multimodal, right? Voice matters. Take it seriously. But right now, today, despite all this talk on the internet, for, for this use case, chat is— chat's 80/20. So you got to be multimodal. But, um, I just I think it's a reminder to not just like get— believe in the hype, but get data, get data. Maybe this will change in 12 months or 24 months. I don't know. I'm just not jumping on the bandwagon that because of a bunch of VCs say it, it's true. I just think in the real world, we may find that people like to chat and text. Right. And in the real world, we are perhaps sadly talking— we're meeting less, we're talking less, we're picking up the phone less often. Right? You and I are much more likely to DM than we are to pick up the phone. I just, just don't go all in on voice and be multimodal because you need it all, I think, is the learning. Okay, this next one. I mean, if you're deep in AI like Simon, this won't surprise you, or these next ones. But to me it is. One pretty smart person made fun of our AI. It says, can't you do it just on ChatGPT? Because ChatGPT has ingested all SaaS content. ChatGPT uses YouTube, whether it steals or not, it's a different question. It certainly can crawl all of our blog. It's crawling every blog multiple times a day. If you look at the data, it certainly ingests Twitter data. I believe, I don't believe it's blocked in any way. So 99% of Saastr content is in ChatGPT and Claude, in Deepseek. It's there somewhere. And so I was skeptical of our AI. And I got to tell you, it's like 50 times better than than ChatGPT or Claude, right? And so I had Claude, you know, write up, and I don't know why, but how do I hire a great VP of sales? This is one of the classic sasterpieces of content. I mean, it's just, and from Jason M. Lemkin. Now again, I could iterate this. I could evolve my prompt. I could be nerdy, but this isn't very good. This is so generic. It's correct. The answer is good. Like, it's fine. The answers that our AI gives are 50 times better than this. 50 times better than this. Um, and, um, so I don't know that I have all the answers to why, but I think my meta-learning is no, ChatGPT can't just do it. No matter what anybody says, not, not just do it. It's not the same. It's not the same relationships. There's more hallucinations. Um, and it's just not the same connection. It doesn't feel like talking to a digital Jason or digital Sasser. It doesn't feel like it at all. So that's my meta learning is you can't telling folks to just do it all in ChatGPT. I don't know today. It's the, it's, it's, it's, it's not the same. It can be much, I guess it can be much better, right? If you match the interface and the training to it, even if all the information is on the internet. So any learnings on that, Simon?

OpenAI ChatGPT has the GPTs. Which you can like, where you can like add custom data to them as well and change the system prompt. The hard part about this is like the data ingestion. That's like, I think where Delphi really shines and that's quite missing in ChatGPTs, GPTs. And there's another note here, which is like people like you, Jason M. Lemkin, probably are interested in building a relationship with their users. And there's no way for you to do that over ChatGPT. So you're way less likely to invest the time and energy to actually import all your data in a ChatGPT GPT to give it to everyone. Whereas in the, on Delphi, you own the customer relationship in a way. So if you at some point in the future want to use your audience and the relationship they've built with your bot in a way to, like, better understand you, the, the, the, your users, and also directly reach out to them, you can. But on ChatGPT, you really can't. There's no data you get. And for sure, for sure. One of the friction points that I've seen are a lot of times people use my assistant UI as a way to get GPT, but then embed it. ChatGPT is GPT, but embedded in their own website. And the main benefit they see from this is that it's now, they now own the customer relationship.

Yeah, at a business level or application level, using Assistant UI to embed ChatGPT on your own site, owning the data, owning the output, that's, that's, you should do that instantly, right? That's much, much better, right? I guess my, my, the bigger point I was trying to make that shocked me, it was not just that you should not, you should not send your customers to ChatGPT. At a minimum, you should, you should use something like Assistant UI to have it yourself so you have access to all the learnings and data, like 100%, 10,000%. There's a different paradigm which we wondered, we wondered this like in the early days of Jasper for marketing emails for AI. It's like, why, why, why do I, like, I wondered, cuz I didn't know what our AI would look like in practice. I wondered, is it gonna be any better than ChatGPT or can I just go to ChatGPT and access SaaSdoor content? Here's the part that shocked me, because the SaaStr content is ingested by OpenAI and Claude. It is there. All of it's there, isn't it? Much more S— all of it's there. And so I was wondering how much better can our AI be than what's there? It's like 50 times better. And I still don't totally understand all the whys, right? I get we could run our own GPT and do essentially what's being done here, which is uploaded into through the API. But forget about that. Just out of the box, I thought with the fresh perspective today, Claude and ChatGPT would be closer to this AI since it's the— since it has access to all of it. But it isn't. It's not remotely as good as this AI, right? This is an incentive to really roll your own and go deep because it's not as good. If you could learn all this from ChatGPT, you don't need, you don't need our AI, right? So I'm not sure I understand why completely, but it's true. Like, it's much better than the same question into Claude or ChatGPT, even though they have the data.

They have too much data.

Well, they certainly have too much data, right? That, that— but, but, but, and listen, there's, there's probably an element here. I don't know how the models are being used, and so does Delphi, because Delphi is taking this, right? And they're putting it through the API, and they are, uh, they're, they're combining this with data from, from the models, right? So I don't, I don't think it's just they have too much, but maybe it is, maybe it is that simple, right? Maybe it is that simple. But what I can tell you is it's much better. So at a minimum, like, roll your own with, with, with, with Assistant UI, but even better, go further with the data because it can be 10 times, 50 times better. Even if your information is already public, it can be— I didn't expect, I didn't expect that the gap, it would be so much better than asking the same questions about SaaS or content there. Okay, point 7. Um, this was super interesting to me. Um, there's a lot of talk about AI SDRs, about outbound, right? Some controversy. I've evolved my thinking. I was a skeptic a little while ago on AI SDRs because all I want, you know, getting 10 million more slightly better automated cadences, I don't think is high ROI. Time has gone on. Now I really see them more as tools for humans. Going to Simon's point, I think they're very powerful. So I'm an AI SDR fan, but this is slightly different. This is a BDR. This is how well out of the box with, with not being a, with not being a verticalized tool, being a horizontal tool, can Sastris AI answer a customer question? And this is just a learning for me. I gave it a B+. Like, this is not a sales tool. Okay. This is not trained for it. But here's a question on the right. How can I exhibit at Sastris Annual 2025? Didn't have to wait a day to get back to you, to you, like it can happen with our tiny sales team sometimes. Didn't have to be during normal business hours. Gave a pretty good question. Here's the sponsorship offerings from Diamond to Silver. How does the expo hall work? Who, if you want to talk to a human, what's the right email address? This is correct. It's nearly full. This is true. What are the dynamics like? 10,000 in-person dynamics, 80% from 2 million. This is all correct. And going to Simon's point, if we hired some random BDR tomorrow to qualify it, they would get most of this wrong. At least their first few weeks. They wouldn't read the prospectus, they wouldn't read the website, they give the wrong answers. These are all correct. So are all— have all the sales type questions been perfect? No, but I'm going to give it a B+ out of the box. Out of the box, it really— it was my— it was just my learning. I did not expect out of the box it would be this good. It's literally just reading our prospectus, right, and the websites and a few documents. It's not— this is not the greatest accomplishment in the history of AI. This is not redoing Private Ryan in real time, uh, but this was— I'm— this was an aha moment for it. Like, this is already B+ BDR. It's already a B+ BDR. So if nothing else, add something like this to your product, because if it can be a B+ BDR for us, it could be one for you too, right? Any— any— are you surprised, Simon, or is this what you'd expect?

Uh, It's what I expect, and I'm, and I think a very interesting point is that people will end up using a chat for more use cases than you initially intended.

Yeah, that's, that's part of it, right?

Well, almost always converges to your AI chat needs to handle everything from support to product questions to sales questions. And it's really hard to make your users not ask the wrong type of questions in one AI chat and to direct them between different AI chats. What I've seen almost everyone do is build everything in one AI chat and teach it to do everything from sales to support to answering product questions and everything in between.

Yeah, you're right. We're, you know, at— we're seeing this in, in the extreme in e-commerce where support, sales, and marketing are already one interface, like already, right? Going to you. But another way to flip around is just if you have a good AI, customers are going to expect to do everything— sales, marketing, support, research, right? Everything. Um, and it is, uh, it, it's, it, it, it raises the bar, right? But it's also very, it's very powerful, right? But it's something you quickly see. This— you're right, this is the point. Folks quickly, if it's any good, if it's actually good, folks will use your AI very quickly for things you didn't expect it to use, right? And they'll use it for everything.

That's right.

They'll use it, they'll use it for everything, right? This, that's so powerful because you might think of it as a bit of an enhancement, right? Or to automate some of what you do. Like we, we're thinking about this from an automation or efficiency perspective, but it's not just automating 40% of trouble tickets. It's all the things people will want to do with your AI that they would never do without it, right? That they think to do it. So, It's pretty cool. Okay, maybe just a couple last points here that surprise— these are just my surprises. They might not be surprises to some of you or to Simon. I'm surprised how much people loved our AI. I didn't know, right? I didn't know. And, you know, this one on the bottom is pretty funny. We build agents, so I shouldn't be shocked, but I'm shocked how good this digital Jason M. Lemkin is, even though he knows how it all works. right? And, you know, I mean, I've had a few conversations with wicked smart engineers that at first kind of make fun of how good I think this is, right? And they tell me how it all works and how I'm not really understanding Transformers and the future of this and that. And I'm like, okay, slow down, just like be a normal human and like ask your SaaS question, right? I just raised $4 million at Demo Day, but I've only got 2 people at the team. I'm growing like a rocket ship. But I have this competition. What should I do with my startup? I bet the answer is pretty good. I bet it's pretty good. So I was just shocked that people love this AI. They love it. Love, like, really smart people. They come, like, it's hard. There's some limitations to this, I'm realizing. I didn't put in this. One interesting learning is how many folks want to go anonymous. So we don't know exactly how many folks come back, right? We don't know how many unique sessions there are, but a lot of folks come back again and again to this AI. Which is the ultimate test, right? If people, if people come back and my learning is if you do this well, you, Simon, you kind of teased on this in the last point, but it really can, if you do it well, it really can connect you with your customers and users in a way you can't. It can add scalability to your relationships with your community, right? Your community of buyers, of prospects, of users. If it's great, right? That's the learning. People love it. Right? So. I don't know anything to add to this, Simon.

There is this potential to give people magic moments that like they were surprised and didn't expect the AI to be this good.

Yes.

And it doesn't really happen. Magic moments don't really happen anymore with like traditional B2B SaaS dashboards.

I haven't— I can't think of one since Canva that was pre-AI. Yeah, I can't think of a magic moment.

Yeah. But they do with AI chatbots, and that's really cool. And it's like a, probably just a novelty effect. At some point we'll get used to computers being just as good as humans and yeah, uh, it'll be normal. But, uh, for the time being, it's a really cool way to, uh, get people to use your product. And, um, it, it, but I think it's super cool is like the fact that people build different kinds of relationships. It's not like Yeah, leads to higher conversion rates and like makes everything that you already do a bit better perhaps, but it also unlocks new use cases, new, new custom product for new, in new ways, interacting with your content in new ways that like wasn't even possible before.

And absolutely, absolutely.

That's also really cool.

Yeah. Yeah. I just didn't know people would love it. Right. So if you can. I think it's— that's the learning. And maybe you bring up a good challenge, at least today. At least today. Even if you're slightly skeptical of what AI can do for your product, flip it around. Is there something magical you can do with it? Like start using it, trying it in your product. Is there a magical moment that you can make happen today, right, with it? Then it's worth it, right? It's worth it. Okay. This next one, this is a meta issue. I've written about this a lot in SaaS because I have pretty different views than a lot of folks on social media here. It's not my perspective, it's just what I'm observing, um, which is that this stuff's a good deal. So, um, you know, the, the craziest ones, I think— I mean, all this stuff's such a good deal. So if I look at what we're using, Delphi, $29 to $99 a month, I mean considering not only is it a great— like, okay, if we thought about it as a chat tool, is it a good deal for a chat? It's probably mediocre, right? It's mediocre. It's fine, right? But look at all the— all we talked about, the 6,000 additional interactions we've had, right? The happiness, the surprise and delight, the, the talking on video with me and Brian Halligan. I mean, this is crazy stuff, right? What you get for your dollar. Um, and you know, the ones I— the bigger ones when I think of— now listen, what you can get— there are rate limits, but when you look at what you get for like Cursor and Windsurf in software development for $20 a month, it's insane. It's insane. Now, you can't— not everyone can get away with $20 a month. If you're a full-time developer doing a lot of coding, I don't think you can get away with $20 a month. But, but if you're only doing a little bit and Cursor or Windsurf can write half your code for $20, and your like social media app is $60 a month to do 3— schedule 3 tweets, it's a terrible deal. It's a terrible deal. And Also, all of these, like Opus Clip, where I invested, that it's like the number one way to create social media clips from video. Like, the free version is so good and it's so cheap. I don't even know what I pay. Like, it doesn't matter, right? It's not the first time this has happened. I think apps like Canva and others are— and Slack in the day was incredibly cheap, right? Zoom. Zoom was incredibly cheap. But man, this stuff's almost too good of a deal. And I guess this is just a reminder, and I know I don't know what all the costs are under the hood for everybody. I know from a few folks that I've invested in, I know that you can make the math work. I know a lot of folks are skeptical that Cursor and Windsurf aren't just huge wrappers on, on, on, you know, on Anthropic and Gemini and losing money. Right. Maybe. Right. I know Opus doesn't lose money on its video rendering, but I do think this was a reminder that just be aware that you have to provide much more value going forward in your app than you did a year ago. Much more value. People are expecting more. They're expecting— when instead of your chat just like being here, when you can develop relationships with humans and it knows 12 years of data and it follows up and it's a BDR, people like you, you, your app has to be like so much better than last year is, is just my reminder. And I don't know that you can charge $50 more a month for— Zendesk charges $50 extra a month for its AI. Okay. Freshdesk is more than twice as much to add its agent. Are we really going to be able to get away with this forever when this is $30 or Cursor's $20? Like, I don't know. I'm skeptical. I'm skeptical that we're going to be able to, to, to charge like we hope we can charge in B2B for this AI. We have been trained that everything is incredibly high value. So I don't know if you have any learnings on this one, Simon, but it's just a reminder to how— what good value these tools are. They're like insane value, I think.

I think you touched on the point that probably every SaaS app going forward is going to have an AI chatbot. And well, every SaaS app going forward is probably going to be agentic and it's going to do more for you.

Yeah.

So we're entering a new era where these apps don't just let you manage data and present it in a better way, but also get things done. For you by just accepting natural language instructions. That's like the era of, uh, SaaS apps that we're going into where it's like agentic SaaS apps, vertical AI agents doing things for you. And it's— I think it's important to start early on to like build, build the right dataset to be able to support these use cases. Yeah, I think it's, I think it's really good to jump on the trend and add some sort of AI interface in your app to some sort of natural language interface in your app to see what sort of questions people ask and also train your users to use more agentic patterns. Yes. And like grow with that because these things take a while to built, it takes a while to get to product market fit because it's almost a new product. And I definitely think that the user expectations of like what sort of value you get for software is going to change. We're now suddenly able to automate a lot more work with software. And so the typical software that you pay for is going to do a lot more for you. And to your point, I think it's definitely going to be deflationary in the sense that like that for the amount of value provided, where the expectations are going to change since computers can just do much more than they used to.

Right. Yeah. I just think maybe to pick on that and then we'll go to the last point for another time. I guess my challenge to anybody watching or listening or thinking is just for— don't view— maybe, maybe it was okay last year. I don't know if it was, but today, don't view AI as an upsell. Don't view it as something you're going to charge extra for. View it as how can I make my product 10x, part of my product 10x better? And how can I do that today? Because the competition is going to be doing it, right? It makes— if your product is not better because of AI, it's going to look like more and more of a ripoff, more and more of a ripoff. And if you squint at a lot of public SaaS companies, they're not adding any net new customers anymore. They're flat. PagerDuty, no net new customers. Okay, it's happening again and again. And listen, if you keep— if you view AI as a feature to charge more for, you will get more money from your customers because they're already customers. They'll try it if they trust you or they even like you. They'll try it. They'll pay you the money. But that's— don't view it as an upsell. View it as core to your product going forward. Rethink it. And that's, I think, the future. And if you get your sales team and your marketing team, your product together, and all they're talking about is upsell for AI, I think the competition is going to kill you. I think they're going to kill you because they're going to do 10 times as much for the same dollar or less. Okay, last point, and then we'll run out of time. We can do another session with Simon later. And if you guys want, we'll ask you and we can do a deep dive on everything he's seen across all of, all of his, all the folks on open source and otherwise, uh, at AssistUI, because he has tons of data. But the last meta question I have around is what's defensible? And again, I just, I tried this app Delphi because the way we find most things, brand. I saw that Lenny and Brian Halligan used it. So when I finally had a spare moment, 4 months went by, I had a minute. I'm like, let me try it.

Okay.

This is how most of us in the real world discover apps. I see it. I see someone on the internet. Simon says a tweet that it's great. Jason says we've run, we use Rippling or Gusto and 80% of the world does not have time to like try out 82 apps. So that's how I tried it. I actually didn't know if it was the best. I was actually, all of this has outperformed my expectations.

Okay.

And all of it. But then afterwards I grabbed another app. Someone said, hey, try this. This is, this is another app that can ingest all your data, create a chatbot and do the same thing. I'm like, well, that's interesting. I'll just try it to learn. I'm not gonna move, but it has an RSS feed, it has scraping, it has all this. It just didn't work. It just didn't work. Okay. It, it just gave me generic Claude-like answers. So that's today. And I think it's a reminder that the product matters more than prob— not only the, the, the tool underneath, but probably the model. But in, at the same time, all of this is public. Anyone should be able to build a clone of this AI, right? In fact, anyone could fire up the same application, attach the same feeds, and outside of the personal training I've done, which isn't much, is reproduce it. So I don't know what's defensible here. I don't know what is— and any metal— this is the— I'm just learning. I know a lot now that I know. Now I know less about defensibility because I'm realizing as amazing as our AI is, jaw-dropping, like it's really great, you could build your own in a week and it would be just as good because 99% of this is on the internet. You have to put in the work, right? But you could clone it in a week. So that's the one I'm still processing the learnings from. As great as it is, someone smart could build this in a week and possibly an hour.

So I mean, final thoughts, Simon? I think that's true for most B2B SaaS apps that you could probably—

No, you cannot build Workday or Salesforce in an hour, no matter what anybody says. You cannot fire up Lovable and build Salesforce in an hour. You cannot.

There's probably a couple of B2B SaaS apps that are worth $1 billion that you could build over a weekend. And it's just like to theoretically build over a weekend, but like it's actually a lot more work to get all the details right and build the customer relationships and the sales motion and everything. I think the same applies here. The same rules of B2B SaaS continue working here. AI itself is not going to be defensible, just the technology.

That's the thing. That's the reminder. The AI itself is not defensible, right?

What I would recommend, uh, you do if you implement AI is to not see AI as an R&D, uh, thing, but rather you're just using what's off the shelf available to you and just building a good product around it. And I think that's a much better way to view AI than to say you're building your own AI models and doing everything in-house.

Yeah, the, none, none of that is defensible. I mean, we know it's not defensible, but now that I've seen it in our AI, I'm triply convinced it's not defensible. Right? Yeah.

Yeah. So AI is just another feature in your product.

So wrapping, I know we're at time. Do you, do you then, given all of that, given what we've learned about the defensibility here, do you think that plays to the strength of folks that have the massive datasets, to the HubSpots, to the Salesforce, to the Workdays, to the folks where people have already, are already investing and investing data on thousands and thousands of interactions of humans? Uh, there's a lot of discussion on, on who benefits right now. The answer seems to be both startups and the big guys, right? But any meta learnings from this on defensibility? Is this, is this the, the, the renaissance of HubSpot and Salesforce and Workday?

Uh, no, I don't think so. So like maybe temporarily, but, uh, the amount of data you need for to train AI is going down, is converging as the general models become even better. So the total sum of data you need to create a really good AI agent is going down. And that's kind of democratizing. That means that everyone can build really good AI agents.

And it means you might have 10 times more competitors than you had last year. That's the other thing I'm seeing a lot, right?

That's true as well.

All right. Thank you, everybody. I know we're at the hour, slightly past. Amelia, thank you for guiding us. Simon, thank you for joining us. If you guys want to do more of a deep dive with Simon, ask how things really work. We're going to do more stuff like this. And again, about half our content at SaaStr Annual in May is at the intersection of AI and B2B, where we'll dig into all this kind of stuff. So see everybody next week. We're going to have, uh, Chemistry Ventures talk about Series A funding. And then again, see everybody in May for Saastr Annual. Thanks, you guys.