# Stop Adding Lazy AI to Your SaaS

Channel: Rob Walling
Video: https://www.youtube.com/watch?v=sIqHAtaqMw8
Duration: 52 min
Language: English
Words: 10652
Transcript page: https://viewrankai.com/tools/youtube-transcript/sIqHAtaqMw8

---

[0:00] AI's everywhere. We're using it to run our companies. We're thinking about implementing it in our SaaS or maybe we are. Um but the question is where does it belong? And that's the framework I'm going to talk about today in the latter half of the talk. Topics I'm covering today, I'm breaking my typical three-act structure. I think every perhaps every MicroConf talk I've ever given has three

[0:19] acts and I just uh this is two acts. We're going to talk I'm going to talk about the five rules of AI in a SaaS company. So in the company itself, not the product. And then basically the last two-thirds of the talk is all about adding AI as a feature and how I am thinking about this along with um you know, the other hundreds of Tiny Seed companies. And my hope is is like any talk that I do uh that gives you a framework for making better decisions. I'm going to talk about AI and the roles of it in uh your SaaS company first because the biggest

[0:48] thing that I see with conversations around AI is we get confused and we say, "Well, how how are you using AI today?" And it's like, "Well, I'm using it in support and operations and also I'm implementing a feature." And those are two different things, right? So I'm going to first talk about rules of AI in a SaaS company and I'm only going to spend I don't even know five, six minutes about that just to give you a framework for it. Potentially someone should do a complete talk on that

[1:14] someday. Maybe me, maybe not. Um so the five ways that I've seen SaaS founders using AI in the company is where AI is your entire product, right? The product doesn't exist without AI. Where AI is a feature, that's what most of the talk today is going to be talking about. So I'm going to add AI retroactively. I already have a product, it already provides value, it already does stuff, but I want to add

[1:39] augmentation. It's where AI using AI to build your product which most of you are familiar with. AI for growth sales and marketing and internal operations. So those are the buckets I put stuff into. To give you examples of each of these um there's a couple different ways AI could be your product. One I do not recommend bootstrappers do and the other one bootstrappers are having good good success with. So, the idea of AI being your product is that it wasn't possible to build 5 or 6 years ago, right? Before AI, you just

[2:13] couldn't have built this this product. In other words, if AI goes away, your product would go away. One example of this that bootstrappers I I don't know a single bootstrapper that's doing this is where you're actually building LLMs. You're building models. We, Tinyseed, backed someone building their own model about 3 or 4 years ago. It was before ChatGPT hit. And we were like, Einar and I were like, man, this is so cool. He has his own AI model. This is a is a moat. So, we backed him and then all these models just steamrolled everybody doing that. You just can't you can't outrun that. So,

[2:47] this is a not recommended um approach. The [snorts] approach that I have seen folks do and there's maybe a couple in this room um is where AI is a is a such a core feature of your product that you couldn't have built it 2 years ago, right? So, Fiscal.ai is a Tinyseed company and AI is they have a bunch of data about public equities, stocks and such, and the AI they have a chat interface, they have analysis, they have all kinds of stuff that AI really makes it possible. Lead truffle, many of you might know Brian uh from uh X Twitter. I thought he might be here,

[3:20] but I don't think he's here this this year. Um, but Lead truffle, you'll see example later, is a texting interface for home services company. So, if you're lawn care um and you're out actually literally mowing lawns and chopping stuff, um you have an AI agent that is chatting with someone uh I guess in a call, but they text, right? And they're like, hey, can you come and give me a quote for my lawn? And Lead truffle is the AI underneath that and that just wouldn't

[3:44] have existed without LLMs. Rosie, it's Jordan Gall's company. Many of you know Jordan. Similar to Lead truffle, but uh more on the phone call side. Also has texting. And then Podsqueeze. This is where AI is we feed it our MP3s for the podcast and it transcribes the whole thing, does our show notes, and puts in the little the timestamps and such. And Assembly

[4:09] does this for city and state government. So, having AI, you know, again, building an LLM and doing underlying models is danger, danger. This model is fast, fast growth right now because there's so much opportunity in it. Okay? So, that that's where AI is is your product. AI as a feature, which is the bulk of this talk. Examples of this, probably obvious, but like Notion, right? Notion existed, Notion had a moat, we're all using Notion, some of us, and then AI it added AI writing assistant. Zoom added meeting summaries, just a feature. Loom added AI titles and chapters. If AI stopped existing tomorrow, these products would be fine

[4:48] because we were using them before AI. Podscan, which is Arvid calls, right? It does AI transcripts and filtering. And Figma, text and image generation. So, that's is a feature or is your is your whole product and as a feature. And then for building, well, there's Claude code, I think is the one that's not on this slide, but everything else you guys know that AI for building. Yeah, I know. Um And uh And then AI for growth, just some examples of this if I'm sure folks are using Apollo and Jasper and Copy and Regi, you know, all these there's tons

[5:23] of examples of this now. It's funny, 6 months ago when I put this slide together, you know, I could I could get a few dozen of those and now there are hundreds and hundreds and many of them don't live up to the promise that they're promising, but AI for growth is something everybody should be thinking

[5:40] about. And then AI for operations, this is internal stuff, right? So, this is your support, your legal, knowledge base, HR, bookkeeping, etc., etc. So, the reason I walked through those, and I know I walked through them quickly, and these are just examples, I want to get everyone on the same page of what we're talking about. Using AI in the operations of your SaaS company is important. I think everyone had it's I think it's there's a lot of low-hanging fruit if you're not thinking about it. If you've already thought about it, um hopefully this has given you some more ideas of how to do it, but now we're going to switch it up, and

[6:14] we're going to talk about AI. Let me see if this works. Yeah, there it goes, cool. Uh AI as a feature, because all or most of us in here are SaaS builders. And I think the AI as internal operations and in growth and as as building a product, I think there's a lot of obvious ways to implement that. A

[6:35] lot of people are talking about that. I don't hear as many people talking about how to think about a mental model of what feature should I build? How should I think about implementing it into my product? And that's what we're going to talk about today. The The big thing, the mistake that I see folks making, I mean there's there's

[6:51] several mistakes, right? Um I'm going to give an example of this, but I see uh like lazy lazy AI, cuz it's easy to think, well, the first interface we all knew was this chat interface, this this language interface. And so, can I just add a little sparkle to a text box, and it just generates for you? And it's like, you can, but should you? Does that make a

[7:14] difference at all? And so, I want to get you in the mindset as we go through this of lazy AI versus outcome-driven AI as a feature in your SaaS, all right? The idea is don't add AI cuz you can, add it because it matters. Just cuz we can build things quickly doesn't mean we should build them. Um example of lazy AI is this game. So, if you haven't played this game, it's

[7:36] really good. It's called Battle of Hoth. It uses the Memoir 44 uh uh battle engine, so it's a quite simplified. I got this game [snorts] and I'm pretending that I'm selling it on eBay here cuz I sell a bunch of on eBay all the time, uh, old games and such. And eBay is such a great example of someone who could use AI in a very intelligent way and instead they gave me an a, uh, sparkle text box

[8:03] or a sparkle next to my description. It's awful. So, if I were to list the Battle of Hoth game, I took a picture of it, right? And [snorts] I uploaded it to eBay and then when I go to actually describe what the game is, I click a button and it like spits out, "The Star Wars Battle of Hoth game by Days of Wonder is a thrilling and strategic game designed for two players." The challenge is, AI, A, this is not that great, but B, I could just go to ChatGPT and get this. It saves me like 30 seconds of time. It saves me no

[8:28] time. So, there's no value here for me. What I really want, the outcome that I want, is when I'm If you've ever filled out an eBay listing, you need the UPC, the the manufacturer, the year of manufacture. It's all these nitty-gritty details that you're putting in and I want to take a picture of my Battle of Hoth game and then I want it to fill all of that out, right? UPC, game title, game type, year, theme, type. I'm still

[8:53] doing that by hand. That makes no sense. It's such low-hanging fruit. Now, maybe it's hard to build, you know, I'm sure the eBay code base is what, 30 years old now cuz it started in '94 or whatever. I get it. There's technical issues, but this is the example of lazy or ineffective or unhelpful cuz eBay, some project manager, product manager somewhere [snorts] said, "We need to add AI. What should we do?" And someone said, "Let's put a text box and let's put a sparkle next to it, right? Don't don't do this. I'm just going to discourage you, right? It's a waste of your time and B, it's just it's

[9:24] not helpful, right?" So, um, this is the kind of thinking that we're going to have today is how can you actually save your users a lot of time? We're all using ChatGPT and Claude and and all the other LLMs. So, assume that's happening and then assume how can I take them to the next level with my features. So,

[9:43] this is the meat. Six uses of AI in your product. Used to be three, then it was five, now it's six. And again, the reason I like to think about this in these ways is my brain, when I think of a problem or a concept, until I know what all the outside edges are, the whole universe of possibilities, I struggle to like think

[10:02] through it. When I think of if I had if I was running Drip still today, like how would I add AI to it? It's like, uh I don't know. How would I? Like I just it's overwhelming to me. But if you tell me, "You know what? There's six categories. There's only six things you can do in a SaaS app with AI." It's like, ah I settle. I feel good. And so, that's what I'm trying to do is get [clears throat] us get us into that zen state of like, "Hey, it's it's kind of a checklist. Like start with this list and go down and be like, is it this one,

[10:30] this one, this one? Which of these works, right?" I did [snorts] the same thing. I kept hearing on the internet um uh my goodness about platform risk. There was a lot of no-code folks saying there Everybody has platform risk. You either have Everybody has it. Everybody has it. And it's like, "Yeah, true, but it's not just binary. There's not platform risk or no platform risk." And so, I sat down for a few weeks and thought about it. And did I do a talk on this? It might have just been a YouTube video. But uh there's there's eight levels of platform risk. If you ever want to hear about this, go to YouTube,

[10:57] go to the podcast. This will help you think through platform risk. This is a slide from an old deck, I think. Um but this helped me get my head around. Now, when we talk about platform risk, at least whether you agree or disagree with these eight levels, at least we have a starting point to not just say, "Oh, I have platform risk or I don't." I did the same thing with SaaS plateaus a couple years ago. This is a slide from an old deck. Because people would say on the interwebs, "Hey, I plateaued. What should I do?" And my first question was always, "Well, why are you plateaued?" Because

[11:26] until we know that, I don't know what the you should do. So, then getting conversations and I I decided how many ways are there to plateau? Let's figure it out. And I did something similar to this where I came up with six on my own. Tiny Seed Companies gave me a seventh. Did a talk on this. So this is again getting our head around the ocean that is that is this. Product market fit the same way. I have five phases of bootstrapped product market fit. Not that important, but

[11:49] um that's what we're doing here. So here it is. This is the slide. And then I have 24 examples, real-life examples of either Tiny Seed Companies or companies that y'all know like Zoom and Notion and ChatGPT. That's the rest of this talk is to give you the six uses. And originally I just had some uses and without examples it's like very amorphous, right? So I'm giving you examples from this research uh project

[12:14] that I did. So the six uses that I see. And if you come up with a seventh, let me know after the talk. First one is a conversational interface. First I called this chat. Right? We're used to ChatGPT, Claude. We talk, it talks back. What I realized is it's chat is technically not correct and so I couldn't put it there because there is one example in here that is is not chat cuz it doesn't talk back, but it is a way for me to ask in plain English and get something back. Okay? So conversational interface. Number two is generation. This is where it generates

[12:46] for you. ChatGPT, write a letter so I can resign from my job today. ChatGPT, give me a you know a Dolly, give me a generate an image that does XYZ. Generation, we're all super familiar with it. This is one of the first that I came up with when I started thinking about this. This is what eBay

[13:03] is doing, right? Categorization. This is where we take a bunch of data and we put it into categories. We sort it. And we used to do it by hand. Maybe you had tags or folders. AI is pretty good at this. Ingestion. This is one I missed in my first pass through. I hadn't realized how many SaaS founders are especially during onboarding are ingesting dirty not dirty in the way you're thinking about, but like, you know, kludgy data, um shitty data that's like a screenshot of something and it's like ingesting that and turning it into strongly typed objects in your database, right? So, ingestion is one that I'm glad that I

[13:41] stumbled upon. Analysis, we'll get into it. You kind of need to see examples for this one. And then the sixth that Anthony told me and two other people is agentic interfaces. So, this is MCP and CLI. And this that's that's the new part of this talk that I I think is is a hot

[13:58] topic and it's ever-changing. And I have some good good thoughts on that. If you're enjoying this, you should make sure you're subscribed to the MicroConf YouTube channel. Over there, you can watch the rest of the talks from MicroConf Portland. Jason Cohen breaks down why your SaaS has a hidden growth ceiling and how to bust through it. Craig Hewitt covers what the agent economy means for bootstrap SaaS and how to get your team ready. And Amanda Natividad walks you through what to do

[14:21] when Google stops sending you traffic. We put a lot of work into the MicroConf Portland lineup and it's one of the best we've ever had. Go subscribe so you don't miss the rest. The link is in the description. So, the question as we go through this that I would be thinking in your shoes is, "Okay, Rob's going to talk through these six." And what I want to think is, "How can I or should I How

[14:40] can I apply each of these to my product? And should I apply each of these?" Generation is usually the easiest one cuz everybody thinks about it, but at the other ones and there are a lot of clever uses that that I want to point out here. So, let's define them, give you examples. Chat interface, which again is conversational

[14:56] interfaces I would have called it. Anyways, chat So, ChatGPT, you know, "What's the best class to play in Dungeons and Dragons and why is it the cleric?" And uh ChatGPT tells us stuff. We're used to this. This is fine. Another example, this is a conversational interface. Find email, which is a tiny seed company, he's out in France. Um he is B2B email and phone data for modern sales teams. So, if you're going to do cold outbound, um you you get a ton of uh of leads and such and info from Find Email. And I tell you that I was trying to find a screenshot of his

[15:28] search screen or it was like a filtering screen and it cuz it was gnarly. It was like 15 fields, I think, 16, 17, like a lot of fields and you configured them and it's like I want it from this source and I want size of the business should it be greater or less than and there's like radio buttons and there's all this stuff. Well, the interface now that Find Email uses is I want Facebook ad agencies in the US with less than 10 employees. Like this is so much better than the prior one and you click find companies and then it gives you a list,

[15:55] you know, gives you a list of companies. Um the reason this is not chat is cuz it doesn't talk back, right? It's a conversational interface, but being able to ask in plain English instead of having to use the complex mode is is really nice. Lead Truffle, I mentioned them earlier. As I said, this is at the core of their business. If you're a lawn care or um uh what are you? Lawn care or you're an electrician or something and you're out on a call, you can't pick up phone calls and you can't text and the AI uh will

[16:26] talk back and forth. And then of course Chat PDF. What's the best class to play in D&D and why is it the cleric? Um Those are interfaces, yeah. So the question here question here is, you know, what's a task that would be dramatically simpler if someone could ask in plain language. And the big models I'm seeing are either if there's a back and forth or if I had a big ass setting screen, a big search, a big filtering thing. I think about, you know, again an example in Drip or any segmentation engine is you always have a bunch of ifs, ands, or thens and like lay people don't really

[17:01] know what ands and ors do. I know probably most of us in this room do, but like that would be such a better way of give me all my subscribers who have bought something for me from me that was more than $100 in the last 30 days and are still subscribed. Like that would be such a better way to do it these days in a way that just wouldn't have been possible, you know, back when we built that interface. So, if you have a big filtering interface, I think it's a really really good time to

[17:23] use it. Uh number two, we're going to get into generation. Um and this yeah, I talked about this earlier where you generate uh images and or text. This is an interesting one. SessionLab has a couple good examples in here um in the rest of this talk, but SessionLab is for uh online and in-person workshops and like planning

[17:44] agendas and keeping everything on track. And so, you can imagine going into their software and you build out your agenda, even for a conference, right? 1 hour, it's an icebreaker. Uh 1 hour is to find a challenge, blah blah blah. They added their little ask AI thing that you can see in the middle of that screen. And the cool part about it is they didn't just put the sparkle in a text box, but they actually, I haven't seen this in a ton of places. I really like the way that they put the auto It's like an auto-completer or a suggestion,

[18:11] right? So, cuz if you click it, you don't really know what to to uh the the the downside of doing this kind of thing and just putting the the sparkle is that you don't often know what the instructions are, right? You don't even know the possibilities of what I could potentially type. And so, you get this blank page syndrome where you like freeze up. But if you can give folks uh ideas like add categories or uh what can be improved or generate feedback form or whatever, that's a nice I think it's a it's a positive design pattern for this

[18:40] type of of approach. And this generation, of course, helps people beef up their agendas. PodSqueeze, I mentioned earlier, we send them an MP3. They send us our show notes. Um they transcribe it, then they put all the timestamps in it and the the subheadings and such. Um and it's not perfect cuz it's AI, but it's pretty damn close. And it saves us

[19:02] quite a bit of time. Hey Harris, you made a slide. That's you, man. Yeah. If uh this is TinySeed Tales season 5. Um Harris is here. Wave your hand. If you If you like that season, you should go up to Harris and high-five him and just be like, way to go, man. Cuz he went He went through the gauntlet recording that with me. Lots of

[19:18] ups and downs. [snorts] All right. Team Fluence is um it monitors LinkedIn and monitors engagement and it helps you reach out in a warm fashion. So, on the left is a Team Fluence thing dashboard saying, "Oh, this prospect has been Well, they've been moved to a lead. They've been moved to a prospect." And then they visited These are Team Fluence team members. They visited Thomas's profile page. They visited uh Steven, who's a co-founder, profile page. Um so, that shows their history in Team Fluence of like, "Oh, someone's kind of interested in my software or potentially or they're looking to poach me and hire

[19:57] me uh if they're coming to my LinkedIn." And so, Team Fluence generates from that an email that can say, "Hey Lacy, Thomas mentioned you've been looking at his LinkedIn profile. Saw you also connected with me recently recently, so I checked out Ledger Bennett." That's her company. "Are you not writing any cold email campaigns, right?" The It's It's not just tracking the data, but it's completely generating at least a

[20:22] template, even if this isn't perfect. Um, it gets you there and I like that idea. And of course, Zoom is generating meeting notes and summaries. Notion is generating all kinds of stuff for us. So, generation is an interesting one that I think we're still building building out really good patterns for, but um the question I would ask is like, where can I actually save someone X amount of minutes? I don't know what X

[20:44] should be. I put 15. Maybe it's 10. Maybe it's 30. It's not two. If you can save someone two minutes, don't do it. Like this is the eBay lazy thing, right? So, there's some minimum X that I think we should be thinking about as product builders. The third one is categorization. And this is where you dump in a bunch of data and it puts it into buckets. So, Prodigy is a well, it's a tool that helps categorize stuff for AI for like LLMs and such. It's kind of a a meta thing, but

[21:16] you can feed in a ton of information. This this is uh obviously a um like an essay of some kind. This film is narrated by Ben Jago, a corrupt police detective who is prepared to manipulate those around him and blah blah blah. But what you can see is AI is actually categorizing things, right? Ben Jago, it says person. Metropolitan Police, it

[21:33] says org. Pretty neat. Don't need it in every app, but your app might. Anchor Point, Tiny Seed Company, and they provide asset management digital asset management for like 3D modelers and 3D animators. And so, you can imagine just gigs and gigs and terabytes and terabytes of audio video, mostly video and still images and

[21:55] 3D models of things. And [snorts] they have added this cool feature that where you can ask AI to categorize it all. And so, you can see from the upper left, it's like there's a brass pan, and it says pots, liquid, frying. AI added that. There's a camera, and it adds film camera and camera. There's a cannon,

[22:14] and it adds cannon and gun. So, categorization is another way to think about um this is just sorting. A hot hawk is cold email. And um you can create a a thing you couldn't have done 5 years ago before AI, but you basically say, "Hey, apply this particular label," says MicroConf, "if the contents of this email is related to MicroConf." And lo and behold, you see the MicroConf label at the top. So, it's kind of it feels trivial, but it's like actually super helpful if your customers need it. So, where can

[22:44] you save folks? At least 15 minutes. At least X minutes. Ingestion, as I said. So, the first three are ones that I came up with kind of I'll say on my own or like if they were just so obvious, it was like, "Yeah, I'm going to include these." Ingestion was one I missed, and this one I think can save a lot of time. So, there's ingestion during onboarding of dirty data and there is I think ongoing ingestion of stuff and some

[23:08] really interesting examples in this one. [snorts] So, let's go back to SessionLab, right? They provide you workshop agendas and you know, stuff to help you run conferences, workshops, etc. So, you can either start with a blank agenda, which of course no one wants to do, or you can import an agenda. And [snorts] you used to have to have it

[23:26] in a very specific file format, right? We can imagine if our data set is like, well, it's probably a CSV and then there's all these columns and you have to use exactly the column names that I have, otherwise it's going to choke on it. Well, with AI, you click import agenda there, highlighted in red, and that is literally a screenshot. It's a static PNG or a JPEG of just some, you know, crap someone put up on a on a whiteboard or built in in a tool. And they can import that and actually turn it into an agenda now using AI. So, that's saves their folks time either on

[23:58] the front end of migrating into SessionLab, where it's like, well, we have these 20 agendas we use, or on the ongoing every time we plan, we our process is we use a whiteboard or we use this software and we don't actually want to build it in yours. This is such a genius way, I think, to to do this and it wasn't that hard to build, says the

[24:16] guy who didn't have to build it. All right. And then Reimbi, David David's here in uh in the room and he gave me the example of Reimbi does uh reimbursements for it, let's say you fly a candidate into uh your office and you interview them and then you want to reimburse them, right? Through PayPal or Venmo or well, it's like AP, accounts payable um accounts payable is not set up to do that. You have to add them as a vendor, it's a big pain in the ass. So, Reimbi helps make that a a lot easier. And this one's kind of clever cuz they have to upload

[24:49] receipts for stuff and you don't just want an image of a receipt, but you want the categories, you want the name of the business, you want all the stuff. And so, uh hopefully you can see that. Yeah, it's all right. So, you know, it's the expense type is a meal. It pulls from this this static image. The expense date, the city it was purchased in, the amount, the description, breakfast at Biscuit Love with scrambled egg. I mean, it like pulls everything out. And the vendor is Biscuit Love. So, again, ingestion

[25:13] helps you do it a little faster. eBay, I wish they would build this. It's another example. eBay, if you hear me, I know eBay's here in the audience. [laughter] Watching this on my YouTube channel. And [snorts] then the last, let's say I got Yeah, I got two more. These are the ones The each of these kind of kept ramping up and this one I thought was so cool. So, JD didn't make it to MicroConf this year, but he's been to He and his co-founder been to a bunch. They [snorts] have software for senior living placement Senior placement agents. So, they're kind of like real estate agents

[25:46] or realtors for senior placement. It's a thing that's here in the US. And they're very low technical, very non-technical customers. He says, "When our customers are interacting with their clients, they find the computers get in the way." It's like, all right. "They like to take notes on paper. So, we literally they built a way to print out

[26:02] each customer's" It's like a realtor. "Their customized form fields." So, these print it out and they can handwrite notes in a meeting. And they just take a picture of it and it fills out their profile. Think about that. Like, think about 5 years ago you're building a web app and you're like, "I'm going to build this whole beautiful web interface and then I'm going to give someone the ability to write on a piece of paper and scan it in." It's like, it's preposterous. I'm never going to spend the time to do that. But AI gives us a way to think about this in such a such a new paradigm, right? Such a new

[26:31] approach to it. And lastly, also with non-technical users, Maui is for contractors like construction firms in Costa Rica. And he said, "Field users, which are like foreman, right? Or electricians." You imagine someone out there. They send audio with material requests through WhatsApp. So, they They a voice note and they're like, "I'm out of nails, and also send me hammers." AI converts No, no one ever said send me hammers. AI converts the audio to text and then to a strongly typed like a database object, right? It automatically matches the material to the list of materials. I would never have thought of these, so I appreciated that Tiny Seed companies bring it out

[27:08] because I think this is where we get creative. Talk about eBay earlier being lazy is maybe a mean way to say it, but eBay kind of phoning it in. This is like really listening to your customers and really giving a way to get get data into your app in the way that they want to. So, where can you remove the most friction by ingesting it in custom

[27:26] customer native formats? All right. Analysis and then a gen tick and then we're done. So, analysis is I bucketed a couple things, but realistically, Gong is a sales tool, right? You record all your sales calls of you and your sales people, and it will actually give you um a scorecard. Now, it doesn't I think it should give you like a one-to-five rating. I don't see that here, but it shows who talked to the most on the bottom, but in the upper right, it says, "Did the rep discuss the customer's goals and review the results correlated

[27:58] to the outcome?" It says, "Yes," and AI says the rep did tie the customer's goal of increasing forecast accuracy to the outcome. So, uh it's anal- it's analyzing something. It's not generating it. It's not ingesting it, you know, it's actually taking something and giving kind of a quasi human answer that again, 5 years

[28:14] ago, a human would have needed to do. So, is analysis something that you can do. We don't do this, but I kind of want to implement it. So, we get hundreds and hundreds of applicants. We've analyzed thousands of Tiny Seed applicants. And we hand rate them. Uh we have There's four, sometimes five of us who

[28:32] are hand rating one to five. And I feel like at a certain point, AI could potentially be our first pass. So, that'd be an example of analysis that we have not implemented. And then another one is also from Team Fluence. Remember there was the the LinkedIn um it's where you have the timeline and

[28:49] then it would spit out an email to them. In this case, you can set up um it's qualifying criteria. And there's hard criteria, there's a bunch of different criteria here, but the cool part is this is a lead in the system and you look down at the bottom that I've highlighted in red. It says it gives them a score of four out of five and then it gives a rationale for that. And that's I think a big important piece of this, right? Is not just scoring, but it says Remote First has this many employees, is headquartered here, is a B2B company that sells blah blah blah blah. So,

[29:25] they're a four out of five for your leads. This is nice analysis, right? It gives you a two out of five, you you sort sort descending by the uh by the old ranking here and I think it's pretty powerful. Uh last one is Front End Mentor, also a tiny seed company, and Front End Mentor allows you to um it's like for junior devs, front end devs to learn, get better, and I don't even know how they used to grade these before uh AI, but if you look at the top in purple, that's an AI analysis summary of the person's code, and it says [snorts] you've dominated demonstrated strong

[29:59] organization patterns, but here's some recommendations and it has three things to make your code better. I should ask Matt, the co-founder, how they did it cuz I don't know if they had people review. I mean, this is this is pretty cool, but it's the same thing, right? It's analysis. It's actually analyzing it like a human would. So, where can you move bottlenecks

[30:19] through AI analysis? So, those are those are the five that kind of are UI based and therefore the ones that were most obvious to me. But now we're going to do Agentic and the trippy thing about MCP and CLI is in Istanbul all three people who came up to me and said, you know, interfaces MCP They said MCP. MCP's the thing. MCP, we should build MCP. And I was like, yeah, yeah, cool. So, when I went to write this talk, I was like, so it's MCP. And then I went to Twitter,

[30:47] and turns out it may or may not be MCP. And that's what's up in the air right now. We are at a uh you know, a a time that is in flux. And this is the debate. Do you build MCP or do you build CLI? Or John Knox, do you build them both? All right, both is the maybe the answer. So, of course, I went on to Twitter and asked this question. This was 2 weeks ago. If you've built or planning to build an agentic interface to your SAS, which approach did you

[31:15] take? And you'll notice the winner is both. But it doesn't win by much. It's the 33%. CLI only is 31%. MCP only is 22. So, it's it's muddy, and I'm not going to leave you with a conclusion other than if I were to do it today, I can tell you what what I would do. But I think this is still shaking out, and that's why this is really interesting, like cutting-edge,

[31:38] bleeding-edge opinions. So, I have what do I have? Four, five examples here. And agentic interfaces are not they don't screenshots don't work. So, it is just just quotes from people. So, SessionLab co-founder said, Oh, here's the other thing. Before I get into these actually, I went into TinySeed Slack to do this section like 2, 3 weeks ago. And I said, who has built MCP, CLI, whatever? And there was even a debate and discussion in there of like, oh, don't build CLI. Oh, you

[32:04] should. Oh, this and that. And um uh and everyone who weighed in on that thread said, my my MCP's about to go live. Or we built it 3 months ago, no one's using it. Or we're currently in beta. There was no one in there who had a definitive like, A, this is the way to do it, and B, tons of people are using it. It is

[32:27] really, really cutting-edge, you know? And maybe in a good way, maybe in not a good way, to be honest. Meaning, I don't want you to build something nobody uses. So, that's that's how I think about this topic is it's like still up in the air and I think if there are some specific app types where I think MCP and CLI are really

[32:45] important. Um and it's especially where I think of like SavvyCal, right? Which is the best scheduling link on the internet. And if you imagine um not having an agentic interface to that, that's a problem because SavvyCal is a tool that you're going to want an agent to be able to ping in and and send and

[33:07] book. Um there's a lot of other apps that I think for now need to be holding off and thinking about this. So, let's let's have any of the examples. Um SessionLab, which is the workshop uh agendas. Philip says, "We have a few users." I think they were the furthest along. "A few users using our MCP server. I think our clientele prefers MCP over CLI cuz MCPs have native support in many tools like Cloud Desktop where you have MC tools available and you can set permissions, etc. With CLI, which is similar to it's same as our API, it's much more flexible, but it requires more technical skill. And

[33:46] that's probably the best summary of it. We're going to see competing opinions, but that's where I think it sounds like more consumer prosumers, they want MCPs. I think enterprises want MCPs and then in the middle it seems like more technical and in the middle folks want CLI. So, that's kind of a the state. Um Ellie, who's a uh founder of Inbox Zero, which is AI that keeps you know, in your in your inbox keeps you to zero. Uh we launched a CLI to manage your account and get stats. We also have a CLI to make self-hosting easier. There's a heavy trend towards CLIs for Cloud Coder

[34:19] any AI you run yourself, but if you're using a connector on Cloud Web, it's an MCP. So, once again, kind of doing I'm starting with a CLI, but I bet he'll be moving to MCP. [snorts] And then Check out Page said um checkout page is similar. You can like build a card or build a you know, a page

[34:36] where people can buy stuff. We Today we had our first company actually request our MCP. So, this is what I mean is I kept waiting for folks to be like, "I have dozens of users." Today we had our first company. For them, it would primarily be a way to improve their customer support workflow.

[34:49] Their customers ask questions purchases. Like, "I haven't received my order." Blah, blah, blah. They have Cloud Co-work set up to draft replies. Um but, they need Cloud Co-work to be able to access order history and invoices. So, it's a way for MCP less technical users. And lastly, before we wrap, call in with StatusGator, which is a status page you can set up for your app and it shows not only your status, but the status uh it relies on all the things that you rely on, all the APIs and everything. So, if you have AWS and Facebook, whatever, it'll show you. One of our large enterprise customers is testing our MCP

[35:24] server and loving it so far. They connected it to an IT automation service, and this was the main use case we had in mind when building it. Skipping down, there's a huge amount of fear out there over the death of SaaS. And I put this one last cuz I don't believe SaaS is dying, but I do believe that we're going to need some agentic interfaces and we do need to keep up with it. There's a huge amount of fear over the death of SaaS death of SaaS death of jobs. But, our hope is to at least make StatusGator a part of the AI-powered solution that people

[35:52] are building. It's a massive push to automate more IT work like ticket response and our goal is to ride that wave and be an essential data provider. And that's the same case for a tool like SavvyCal is to be a tool that agents can use. So, what would I do today? Well, it would depend on my customers.

[36:11] If I had prosumers, consumers, I'd MCP. If I had enterprise, I'd probably do MCP. In the middle, I'd do CLI or maybe I'd do both, which is what a third of people are doing. AI makes it faster to build after all. Um I This one is going to shake out. And if it's funny, if I gave it in 6 months, I bet they'll be even more conflicting and

[36:31] confusing data cuz that's AI today. So, I'm going to leave you with this slide. Six uses, six ways. And that's it for me for today. Thanks for having me. [applause] Rob Daves. Um how do you think about pricing when you look at all these tools? Ooh, AI price like pricing with AI. Yeah, cuz sometimes these things should be like easy like that eBay one shouldn't add pricing, but some

[36:56] certainly should. Yeah. Yeah, I've seen three models. And the key is to when you launch these, you launch them with a small subset of your customers first to see what the actual cost is going to be. I had a whole section in here that I took out for time and cuz it wasn't very good if I'm honest, but it was kind of pitfalls of putting AI into your product, and cost was one of them. Uh

[37:18] compliance was another. I've seen three models, right? What The first is you absorb it, and that's maybe the It's the most common for kind of the It's one-time ingestion. It's as you said the eBay thing wouldn't be that big of a deal. So, you either absorb it, you have um you include it in tiers, a certain amount, so you have to go up a tier to do X amount of calls to the AI or X amount of of you know records processed or whatever, or you have it as an like an add-on [clears throat] like SMS would be, right? You have AI

[37:52] tokens. The problem with making it an add-on is the the best It's the best way to do it for you as the business cuz it's like, well, I'm all I'm in the clear that I'm never going to eat eat money on this, but your users don't know what AI usually don't know what AI tokens are. So, you're like, great, buy 1,000 tokens. And it's like, well, what does that get me? Cuz we know when we're losing Claude Claude code, I put like 20 bucks into Claude code the other day, and I don't know how long that's going to last me. Like I have no It's just a

[38:14] black box. Um but those are the three models, and I don't I don't have a recommendation. I think it depends, you know, on which which I would go, the specifics of it. So, similar, I mean, I get to ask this question a lot about SMS and how you should do that cuz it's expensive, right? Um usually you don't absorb it, but you either put in the tiers or you have a kind of is an adder and there's then there's overages, does it roll over to the next All of this is someone should maybe do a talk about that at

[38:42] some point. Sarah. I think I may have a seventh for you. Um although I don't know if this fits into your model or not. What are your thoughts on headless SaaS where the head is literally AI talking to the CLI, the MCP, something like that. So that there's no UI for user, but it's it's really AI driven. So in in my use case here, deploying websites, for example, I have a CLI tool that I use, but an AI can just use it as easily and I I don't necessarily need a UI to click around and push buttons. Do you think that's a separate category or does

[39:21] that fit into your How does that work with you? Yeah, that's a good question. I I would my gut instinct, without digging too far into it, but I might want to ask you later, is that it would just be in the It would be in category six. It would be like an API almost. So like Jordan, who just asked this question, has an API to get Secretary of State data from all the 50 states, right? And Cobalt Intelligence is a company and and you you have a web UI you can log into to like if you're a consumer prosumer to download a CSV, but you also then have an API that you can

[39:53] hit. And let's say we removed that web app that like the web interface, I would still call that SaaS and I would still call it It would just be a MCP CLI interface, I think. Um yeah, I I don't want to keep going, but I I I want to It seems like you you have to have some user interface, right? Cuz don't I have a login? Like don't I need

[40:12] to create a token to to hit your API? There's some type of login. Is it just very simple? You can just do OAuth, right? You can just have an OAuth flow that just gives you the token and that's it. Yeah, yeah. Okay. Who's that? Me or you? All right. There's a little feedback. Yeah, yeah.

[40:28] I'd put it in sex, I think I would. Yeah, back here. Hey Rob, uh I'd be hard-pressed to not say this, but I built the eBay thing you're talking about almost to a T. Wait, what? in all the aspects kidding me? No, it's kind of fun. And I'm also Minneapolis based, so hello. What? Um [laughter]

[40:44] What is happening here right now? Okay. Okay. Can you approach me later? And yeah, I I want to hear this. Yeah. My question is more is less about the technical aspect, but I find when I talk to people of on all ends of spectrum about, you know, some are really technical and some are kind of out there thinking about AI. I have a hard time um not hard time, but it's I find it quite interesting when you say AI, it means something entirely different to one person to the next. So, they How do you

[41:13] translate that to your customer? Mhm. May see AI as this endless thing and some may be technical. Like, where do you balance that? Ooh, this is a good question. Out and outside the scope of this talk, but still a really good question. I would love to answer this like spend 15 minutes on the podcast just talking

[41:30] about this. Because I made a prediction for 2026. And if if you followed any of my predictions on the podcast, I'm like a two out of 10. Uh as usually, I'm about 20 30% accurate. So, it's just funny. You should have the anti-portfolio of me. Bet bet that everything I say is not going to happen. But one of the ones that I said was I thought AI as like AI and H1s and AI as a positioning and a product descriptor was just going to go

[41:53] away. Cuz everything's going to be AI. So, it's going to mean nothing to put it in. And I think relates to this. Um I think that if you I think it does depend on your customer type. And I you know, I'm imagining uh imagining my mom or my brother who like use chat GPT and that is AI to them. And so if you were to say take a picture of your Battle of Hoth game and turn that into the listing and you and I know that's AI and that it'd be very hard to do without it,

[42:29] that wouldn't even connect with them. You know what I mean? So, that's a non-answer but it is that's the path that I would start down is like I would start with my customers and I would want to figure out is AI mean anything to them or is it the result that they want that they're going after and then figuring out how to do I brand it as my own term of like it's smart import, you know,

[42:50] it's it like do I do that? I'd toy around with that or if AI means something to them. Yeah, it's a checkbox right now. I mean, I will say well, here's something else though. I had Jason Cohen and he was on the podcast Startups For the Rest of Us what? 4 months ago? And one very smart thing he said among many smart things he said was that there especially with enterprises their budget there's budget line items for AI now, right? So, you have all this stuff you can spend this and that and then AI million bucks we can spend. So, in that case would I put AI in my H1? I

[43:21] probably would if I knew that my customers had budget and were specifically trying to do that. So, this is where it depends and it's a you know, a longer longer conversation but I think this is a really good question to be honest. What do we got? I I have the box over here. I don't know if did anybody have the yellow box want

[43:37] to go first? Anthony Some trouble maker. This is the guy that came up and told me MCP in Istanbul. You're welcome. Um, I have a I have a question for you regarding the attitudes of customers and other folks to the companies in Tiny Seed who are starting to include AI and have Did you talk to them all about pushback, negative pushback from people who either are against the way that AI works right now in terms of the the amount of power that it uses, the just basically the impact on humanity, the negative impacts, and is it impacting them in any negative way by including AI

[44:13] in various places in their software? No, I haven't asked them. I I could write a talk about that. No, I I like the sentiment cuz there's definitely it's not as bad as crypto in terms of the polarity, but there are certainly folks, you know, like my 19-year-old hates AI for the creative the artistic reasons cuz they're like artistic and for the power and for all that for those reasons. Um but I have not asked, but you're right. It sounds like there are two ways that AI can kind of bite you in the ass there. It's A, if your customer base just doesn't generally like AI as a as a concept, and

[44:50] B, if the AI is lazily or poorly implemented, right? Of like, why did you add all these stupid sparkles everywhere? Yeah, there's No, there's only two and there will never be more than two. Heard that before. Any other questions? I got one here. All right, yeah, Craig. Uh-oh. Oh, no. This guy. So, Craig is Craig he's trying he's trying to troll

[45:14] me. He's like, "No, ask the question." [laughter] He's doing a talk later today about AI in SAS. So, it's I think we're going to be I don't know if we're going to be counterpoint, but I do think we're going to be complimentary. Yeah, yeah. I I think so. Uh really really just a comment from we had like a little knowledge share session this morning and the the topic [snorts] of

[45:32] liability came up. Mhm. Uh I think from Andy um around like uh or Victor like if AI is used in software development, there are clauses saying like, "Hey, with this is not violating any copyright," which is absurd, of course, but also like uh at what point does um liability and human in the loop come into play with AI? So so like I really don't have an answer, but it's kind of just something for us all to think about is like, we're using all of this AI to create a bunch of At what point are we introducing liability to ourselves, our business, uh potentially our customers

[46:05] in using it? Uh I don't have an answer. It's really just a really interesting thought that I spend kind of all day every day in this and I hadn't thought of, so I just wanted to share it kind of with everybody cuz it's relevant to how you're using sat AI in your product maybe in a not good way.

[46:19] Yeah, thank you. But Yep. Yeah, and that's the thing, right? Is is we all feel like we need to stuff AI in really quick, and what are the downsides of it? There's liability, there's potential costs, there's potential negative implication, you know? So, these are the These are the flip side of like I'm not saying do it slow, but at least be deliberate about it and think about it. If you're going to run this product for 5 more years or 10 more years, you want to, you know, you want to be smart

[46:42] about it.

[46:47] Oh, do you Yeah. Hi, up. Thanks for talk. Um what do you think about agents uh live inside the app? Like agents which do the real work, like Notion agent, for example, which is um summarize, ingest data, and call external tools. What are your thoughts, Tony? Agents like building agents inside Uh agents live inside your app. Yeah. That you build your That you build

[47:12] yourself. Like Like you you want to So, it's not an external agent coming in hitting an MCP. It's at You just named it. Eighth God damn it. That's an eighth, isn't it? Building an actual agent full-blown inside your app. Yeah, which uh helps your um Like, for example, my user create an agent in my app and uh like my app do some work for for user and feels like, you know, like employee uh which uh summarize, ingest data, call external tools. What What What are your thoughts,

[47:43] Tony? Uh Well, that's like a A that truly is probably an eighth example of this. I'm not doing this talk again. Um I'm not expe- I'm just going to be like And that was what it was at the time. It's um What it sounds like is that an agent built into your SaaS would then do these different six things. Yeah, did I say six? Yeah, it would be seven. Anyways, yeah, agents would do generation, and then it would do

[48:10] analysis, and then it would do these. So, uh A it's technically probably not another category, but B I think right now, unless your users are ready for this, like maybe we should slow down and and look at these six and do some MCP and CLI. Unless your you feel like you're you're losing deals because of this, or you feel like it

[48:32] could be a real competitive advantage. That's the only reason I would do it right now. Um I think it's the future. I don't know if it's the present. At least not from what I've seen. I mean, the fact that I thought MC was I thought people were building MCPs and CLIs all over the place, and everybody cuz everybody's talking about them on Twitter. And then when I saw that like nobody had any users, it was like, oh, we're not there yet, you know? So, I bet if you ask this question in a year, I'd be like, well, [laughter] obviously we should be doing that. But right now it just feels like I I I don't

[49:01] know is kind of my ultimate answer, you know? I don't think we're there yet. We have any more? Yeah, Harris. Hello. Um question for you. I'm seeing and like some mid-market deals, procurement people are saying that they need to be able to opt out of any of their data like touching AI. Uh Part of me is like, hey pal, your email just like went to Google, went to Gemini, like it's already like

[49:24] Yeah, yeah. [laughter] So, I guess we're sort of everyone's like ignoring that and I guess talking about their like core company data. And so, my question to you is as you think about building features, Yeah. do you need to think about isolating it from the beginning, or do you say to those people like, you know, That's really interesting. I mean, kind of it So, what I used to when Derek and I were building Drip, I would often tell him or we'd we'd talk about collaborating on a feature, and we'd say, we're going to build it really simply. The UI is going to have like a text box and a check box and

[49:56] that's it. But the data model, cuz data models are so hard to change, the data model is going to have all these many-to-manys in case we ever want to do something. So, we wouldn't gold plate it per se, but we really gave it room for expansion. I was building it today, yes, I would have some I wouldn't probably wouldn't build it in the UI, probably wouldn't offer it as a first thing. But [snorts] that just so you heard the question, it's like, if you build these AI features and someone wants to opt out of all of them, that's a trip. Because it's not just having a a hook or an if then else

[50:24] statement or whatever it is, different pricing plan. It's like the app just isn't as good. Some of these things really make it a lot better. So, um my answer is I bet we will see it become more prominent the more of these features that get built, you know? Also, today doesn't feel like a today thing. It wouldn't be number one priority. But would I want some hooks, some data, something, some stuff in my business objects to maybe be able to handle it in

[50:48] the future? I would. It's about it. I like that this started as like a talk about building AI features and we brought up like five reasons not to do it, right? Or like hang-ups. This is cool though. This is the MicroConf thing. Like this is real time. This shows you that this topic is not so evolved. It's not so um it's still so nascent that we're still all trying to figure it out, right? It's like real time. You heard this last week, right? So, that's business intelligence that everybody I didn't know that and the rest of the people in here may have not have even thought

[51:16] about that. So, I hope you enjoyed that talk. And if you want to join us for our next MicroConf, we'll be in Iceland from September 21st through the 23rd of 2026. And then we'll be in Austin, Texas April of 2027. Get all the details and your ticket at microconf.com/events. Use promo code rob50 to let the team know I sent you. After watching my talk, you might be thinking about how to get your team on board. Maybe you have some stragglers on your team who aren't

[51:44] leaning into AI as as as you would hope. In this next talk over on the MicroConf YouTube channel, Craig Hewitt has some blunt advice for founders. Go check it out.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt