Why Building With AI Got Easy and Maintaining It Got Brutal with Fathom CEO Richard White

0:00Richard White:Most people are reacting to AI. Our guest for this episode built for it three years before it arrived. Two bets. Transcription costs would fall to zero, and AI would get good enough to actually do something with what it heard. Both were contrarian then. Both were right. Fathom is now the top-rated AI note-taker on G2, and Richard is one of the few people who can tell you what actually changed and what didn't. We get into why building software has never been easier and maintaining it has never been harder. Why your years in business are an advantage in this shift, not a liability. And why the real bottleneck right now is not the technology.

0:42Richard White:It's what you can see. If you have been waiting for the right moment to move, this is it. Richard White's background is engineering and product design. This is the Vault Unlocked. Let's unlock it.

1:04Richard, welcome to the show. I'm excited to have you. I just, for the guests and the viewers listening, why don't you tell us a little bit of who you are? I'm excited because I use your product. I love your product. It's in our business today. I've seen you've changed it quite a bit. and I'm excited to have you here. But for the listeners that may not know, tell them who Richard White is. I'd like to think I'm a product designer and kind of technologist. You know, no one's let me write code of production in gosh, maybe 10, 15 years. So I'm not sure I can claim being a technologist as much anymore.

1:40But that was my background. Originally kind of in engineering design, done a couple of startups. I worked at the first patch of Y Combinator. If I want to date myself, I did a product before this called User Voice. But as you kind of alluded to for the last five, almost six years we've spent working on Fathom, which is the number one kind of rated on G2 AI note taker for people on lots of back to back meetings. It's been a really fun ride with a really great team. And the most fun part about it is talking to folks like yourself who love and use the product every day. Yeah, I mean, I've used a lot of different AI note takers and I've used Fathom before and then, you know, we switch and now I'm back to Fathom and I'm, you know, I'm actually I'm sold.

2:20To me, I find it's the most easiest interface, just the usability of it. And I love that. I just feel like it's not overbuilt. It's just built just exactly for what it is. Take us back to where did this start? Where did you see that this was needed? Because the one thing I do know about Fathom was way before this huge, the AI craze and everything. So you saw something way before. Or that's what I'm interested in, like the vision and the strategy you saw and how you brought it together. Sure. Yeah. I mean, it was actually even right just before COVID. Honestly, it was working on a different product, was working on a totally different product and totally different space and just found myself on a ton of Zoom meetings.

3:03Like I think it was like 15 to 20 a day. A lot of them were research sessions, right, where I've got 20 minutes almost back to back to like interview someone, demo something, get their feedback, rinse and repeat. and it's kind of one of those things where like you know if you run into a problem once a day you don't maybe do anything about it you run into it 20 times a day you're like oh my god this is really painful i need to like i don't i need to fix this right and so you know i remember just kind of kind of thinking how kind of crazy it is the way we kind of share knowledge out of like meetings and stuff right it's like oh i meet with someone it's great experience they tell me some really interesting quotes or facts or whatnot and then i heard we scribbled down notes and then try to like clean them after the meeting and remember exactly what they said.

3:46It's a very stressful situation, right? It's like being a court stenographer, right? And also being the, the, the, the lawyer interviewing the person on the stand at the same time, right? It's like, you're kind of doing both and no one likes it, right? No one likes taking notes. No one likes reading notes. Notes are also like, we're like a really poor artifact. I share that my team and a lot got lost. You know, I'd have this amazing conversation. I'd share the notes with my team and they kind of shrugged their shoulders. We're like, okay, right? So I sort of look at this and be like, gosh, there's something, don't we have the technology to fix this at this point, right?

4:17And, you know, if you go back to 2020, there were tools that were doing call recording. Nothing with AI yet, obviously. You know, most of the products are like in the sales space, like companies like Gogg and stuff like that. And they're really expensive and they're candidly kind of mediocre, right? It's like, oh, it took you 30 minutes an hour to get the recording afterwards. It was mostly just a transcript. No one wants a transcript. What I wanted was just like, I get off the meeting, there's instantly some notes. Great. Like, I don't have to do this job sort of thing. And we kind of looked at that space and we kind of had this thought like, gosh, where is this space going?

4:53We kind of had two core hypotheses that really got us excited. It got me excited about what turned into Fathom. That was transcription five years ago, actually pretty, still pretty expensive, right? It was like$3 to$4 an hour to transcribe content, which doesn't sound like a lot. But if you imagine if you build a product in transcription, build a product in meetings, people are easily going to do 10, 20, 30 hours a month on it. Gosh, your hard costs for that product are already like$50 a month, right? So the fact that Gong and Fooksuit were charging$150 a month makes sense in that context, right?

5:25I was like, why is it so expensive? Oh, yeah, the input costs are expensive. And so we kind of looked at it like, well, we think this is kind of commodity. Like when I, we, we tried a bunch of different vendors, like I kind of made a little prototype and tried a bunch of different vendors like Amazon and Google. And I think one was called Rev. I was like, these are all pretty good. They're not great, but pretty good. So with the hypothesis, like transcription costs will go to zero because it's, they're all good enough and it costs are always turning down. We think it'll go to zero and more important than that is like, and we think AI is going to get really good.

5:57And it's kind of funny now because it's kind of an obvious thing, But go back five years, very contrarian take because it's hard to remember. But there was a there was a wave of, quote unquote, AI companies like 2015 to 2020 that were terrible. Right. That promised you the world and delivered almost nothing. Right. And so but because I was like, no one wants a transcript. I don't want to get off a media and read a transcript. I don't want to read. Yeah. Nothing to do with the transcript. But the AI will need a transcript to do all the fun stuff I think it could do in the future. write your notes, write your actions, fill in your CRM, find trends, find themes, alert me when certain things happen.

6:35All that stuff needs a really good, high-quality transcript. So we started the company with those two ideas and said, gosh, if those two things are true, transcription costs go zero and AI gets really good, could we be the first people to give away this product for free? In a space where people are showing you$150 a month, what if we just gave away for free? Because we actually don't think the value is in the meeting itself. It's in building up this database, then you build a bunch of AI features on top of. And so we always have the thesis of, we're going to give away this product for free to individuals with the hope that that gets us into a bunch of companies where we can then sell a different product to the managers of those people, right?

7:09Because the managers have a different problem, which is I'm not in the meeting taking notes. I'm outside the meeting, and I want to know the important things that are happening. I want to know there's a pricing discussion that doesn't go well. There's an argument that happened at the engineering standup. But there's a deadline that slipped three times. But I don't have time to sit and listen to every meeting. And so I got really excited about this business because, one, kind of fit the hypothesis of where I thought the world was going. But two, it had this really awesome kind of two-sidedness to it.

7:37We had one part where you can give away a lot of value for free and feel okay about that because you don't have to charge people later. Because there's this nice kind of complimentary business built on top of that for their managers.

7:47Richard White:That's kind of how we got started, right? And it's kind of funny you mentioned that we were kind of ahead of the curve. And I think that's probably true because we had a third corollary to those hypotheses, which was if you wait to win transcription costs is zero and AI is really good, you'll be two to three years too late to start this business. Right. Kind of obvious to everyone and everyone will jump in. But like any technological revolution, the best companies like build towards a hypothesis a couple of years out and they do all the other stuff. We spent two or three years building all the foundational work and the product experience that you talked about, the good user experience, good usability, the reliability, the distribution channels, all that sort of stuff.

8:24And so it wasn't very much a go to where the puck is going kind of thing, not like wait for it to get there. So when you guys were doing the hypothesis, like this was like back in 2021 COVID days. I don't like to use that word. I hesitate to say it because I feel like we just don't use that word anymore. I don't use it. I hate using it. I mean, it's funny because sometimes my brain, I keep thinking, it was only a couple of years ago, but no, that's almost six and a half years ago now. So today, there's a lot of players in the marketplace, but you've had the market share. So are you seeing competitors coming in and taking over?

9:05Are you guys adapting your product now more with AI? How are you staying in the trends and how do you see where AI is going? I mean, even just with the note taking, let alone what is that next vision that you have for where this can go? Yeah, it's kind of funny. I mean, I feel like it's been a tale of two businesses. I do this whole talk and I show my revenue graph. You could see the point where AI actually shows up. And for us, that was kind of like GPT-4 level of AI. That was a point where the AI could write better notes than a human, right? And that's where we went from being a meeting recording business to being a meeting AI business, right?

9:41And it really takes off. And this whole second act of the business is not about hypotheses and stuff like that. It's actually about how do we get really good at building AI functionality? Because it's actually very fundamentally different than building traditional SaaS or just software. And I can talk about that. So it's been kind of cool. And now we kind of are seeing this, like, you know, the capabilities increase every six to 12 months. And now sometimes it's even like two to three months. Right. And so we're constantly now seeing like, okay, two years ago, state of the art was we can write a really good summary for a meeting and we can figure out the action items.

10:15A year ago it was, oh, we can look across not just one meeting, but every meeting you've had with Acme or, you know, every meeting you've had in this, with this prospect and we can surface like risks. We can surface trends, blah, blah, blah. And now the state of the art is moving to, oh no, no, now we can actually look across every meeting you've had over four or five years, right? Across your work and tell you trends about competitor trends, internal knowledge management type questions. Like, you know, we asked it the other day, like, Hey, we don't like to do a lot of documentation here at Fathom because we just assume that it's a point you just asked the AI, like why generate documentation and maintain it?

10:50Just, if you need to answer a question, you just ask the AI. Now we're at this point where that actually works. And we can be like, Hey, we've got a new engineer and they're wondering about why we built the system the way we did. Can you give me a history of transcription engines at Fathom and it'll go over four years of meetings and it'll write a six page memo, like, you know, 10 minutes. So I think it's pretty cool. We're moving this world where I imagine two fun things are going to happen in meetings. One, I just imagine meetings are going to get really good, right? Like this has been my weird mission for someone who hates meetings.

11:18It's like, how do we make meetings actually fun? One, we remove all the work, right? So like, you don't have to be a stenographer, but also you don't have to get off a meeting and then have more work than when you started. Right. I think that's where this is going. It's like, everyone hates meetings. Even if a great meeting, I still at the end of the meeting, I'm like, crap, now I gotta go do all the stuff we talked about. We're not too far away. You get off that meeting, two thirds of it's already done. The email is drafted. The follow up is scheduled. You know, the, you know, the presentation we went to build out is already stubbed out.

11:48Maybe it's even 80 % built. Right. So one is this kind of magic of you speak things into existence on meetings. And then the other thing that I think where we're going and where the space is going is kind of like information finds you. So the other thing people hate about meetings is they're in a billion of them, right? They're sort of their inability. They're in a billion of meetings, right? Oh, in a billion. Yeah, yeah. Yeah, we're all in tons of meetings. And it's because it's like the primary way we disseminate information in organizations, right? It's like if you weren't there for the meeting, you're not watching the recording.

12:20You lost it, right? Because you don't want to sit through a 30-minute recording or read the transcript. It's just gone. And so if one time you were needed on a meeting, well, shit, now you're going to be on that meeting all the time. I actually think there's a not-too-distant feature here where, hey, we have really small meetings. And if someone not in that meeting needs to know something about that because we talked about a project they're related to or we reference a customer that they're in charge of, that information finds that. I actually imagine a world where like you only have two, three meetings a day, but you have an amazing podcast you listen every morning.

12:50That's basically curated from everything that's been happening on the org yesterday will happen today. And it's telling you, hey, here's some updates around the org. You might want to go talk to Tim about this update or that. And so I kind of think that's a world where you've got, you know, AI native teams. There's smaller teams. There's less meetings, but there's actually paradoxically less meetings, but more shared context throughout the org. And so I think those two things, the work gets done for you and the information finds you, puts us in this like really exciting world where people can get out of meetings, get back to building stuff again.

13:21Right. And doing work. Is that what you're is that what you're working on? Is that the that's is that like so is that a different company or is that what fathoms? No, that's that's that's our state of mission. Right. Our mission is to like make meetings amazing by kind of continuing down their source of intelligence that finds you and we do the work for you. Or we, a lot of times now we partner with agents that'll do it for you. Right. So we have API MCP, all that stuff. So it can do some of those actions for you. It's interesting because I was just thinking, so basically all of these people, I just say all the different, all the different departments, all the different roles are having meetings throughout the day.

13:59I just want to understand this because I think it's wow. And then at the end of the day, all of that's curated into a 20, 30 minute podcast, maybe so in the morning all employees basically or anybody can like hey what's going on in the company you listen to it you have full idea of what's going on all the departments and i love what you said information finds you so if something is happening on the department in a meeting you're not even part of and your name is mentioned or whatever it might be you would get a notification saying hey even though you had nothing to do with it to either be ahead of it to understand what's going on, whatever that might be.

14:38And you're actually, you guys are working towards that right now. Yeah. I mean, we already have a version of this today where you can put in what we call trackers and it's not like a keyword. It's just like, Hey, I want to know anytime a pricing discussion doesn't go well, or I want to know anytime there was a heated debate in like an engineering standup or it understands tone, understand semantics, and it will compile all those clips together and either daily or weekly. It'd be like, okay, here's every competitor mentioned. Here's every pricing discussion that go well. Here's the themes of, of, of what these topics were, right?

15:06In cases like that. And so we already have today that I can go find you, but you have to kind of declare what things you care about, right? We'll opt you into a standard set, but like, but I imagine we're, we're going to keep going beyond that to like, not only do you just kind of explicitly say, here's the types of moments I'm interested in, but the AI eventually just, you know, looks at the, the job title on your badge and kind of says like, ah, given you, and I know the projects you're working on, like I'll go set up a bunch of these myself, right? Like, and I'll listen to all these trackers and then I'll synthesize them and give them to you.

15:32So kind of like kind of like a meeting notes themselves. Like, I think we've got the V1 today, but I think where it's going is going to be kind of mind blowing. Yeah, I was just as you're thinking, as you were saying all that, I was thinking the next layer, too, is it could be an intelligence for the business owner, like for the owner or the, you know, the board of going, what what's the energy like in the company? what you know are people happy in the company are people's dissatisfied i mean obviously people watch what they say on uh the meetings but there is tonality there's facial expressions there's things that are happening that as a business owner you can just get a report at the end of the week and be like hey you might you know your engineering team there's there's a there's an issue here like right this thing's about to explode yeah there's not a lot of folks speaking up there's you know very contentious meetings there's a lot of stuff and i i'm glad you mentioned tonality because you know, we first got in this business, everyone wanted to do just like sentiment analysis on transcripts.

16:28And I'm like, so much is lost when you don't have tone, right? Like, especially in business, right? In business, it's all about tone, right? I, my background is engineering, but I ran our sales team for a minute, my last startup and you know, tone is everything in sales. Yeah. Yeah. They said they're going to buy, play me that clip of them saying that, right? Like you'll know from that clip, like whether they're going to do it or not. Right. So yeah, it's pretty impressive what they can do now. And we're not doing it yet, but I also imagine, yes, facial recognition, like, you know, how engaged are people and stuff like that is something we'll look at in the future as well.

17:01I haven't done research. Like how big is Fathom now? Like the company itself? By employees, about a hundred. Okay. Wow. Okay. But we're also kind of, you know, one of my internal goals is I would like us to get to a hundred million revenue with less than 150 people. I actually have a lot. I think actually like we're now in this era where it used to be that, you know, no one wants to talk about the revenue. No one's like going to be like, oh, here's where revenue added. Here's where growth. So I've heard this uses employees as a proxy, but I feel like that proxy is getting broken, right? Because so many companies now are like, gosh, I don't need a 300 person sales team now to get$200 million in revenue sort of thing.

17:37No, no, you don't. Again, that's the power. Like, I mean, as an engineer is someone who's incorporating ai into your product and you've been incorporating obviously at the next level where are you seeing the like where where does the ai stop at some point because the one thing i've realized is like as great as it is today it's still like i don't care what anybody say it's still not there like if you ever had it like if you ever had to actually ask whether it's claude or cloud code or gpt to actually do something it doesn't get it right every time like you sit there fighting with it where do you think it gets to the point where like you don't even you just kind You're just talking and it's literally listening and it's literally building.

18:16And when does that stop? What's the negative impact of that? I mean, I kind of look at it as like going back to the command line versus some package software. I think we're getting to points where anyone can open the command line that's a Claude or JetGPT and get decent outcomes, especially for personal requests, stuff like that. But there's still a lot of room to basically engineer a better answer or a better output by being really intentional about which models you use and which order and whatnot. Right. And so I think like we're seeing is kind of the, you know, the clause in our great general purpose solutions.

18:54When you're like, I know I want this specific thing, you can get better speed, better accuracy, whatnot out of purpose, still purpose built systems. Maybe we'll get to 0.5, 10 years where it won't matter. Right. And there's like, ah, there's a general brain. It's good at everything. Right. But at least for the next handful of years, there's still a lot of value. And I think vendors like ourselves where we have a whole team that is nothing but R &D lab that's constantly figuring out, you know, everyone thinks, you know, all the time people are like, hey, give me all my transcripts. I'm going to throw them all in the cloud and I'm going to ask it some trending questions.

19:25I'm like, you could do that. It will not succeed. Here's your transcripts. Like for us to get to the things I was talking about earlier, like those tracker concepts and be able to like basically give answers across tens of thousands of meetings. there's a lot of engineers a big pipeline of different ai steps we have to take right it's not like one agent's doing this think about like it's a whole team of agents that are taking on different parts of this task i i i understand uh yeah what you're saying it's not as easy as just throwing it up but let's talk about that so people understand because i know people do that they would throw up all a bunch of their transcripts in a say claude and say give me the you know the feedback but it's, but it breaks and there's a nuance that misses and, uh, and it hallucinates.

20:06I mean, I mean, I'm working with it right now and it's like just nonstop hallucinating and I'm catching it. But for some people that don't know how to use AI properly, like it's a, it's not as perfect as people think it is today. Yeah. And that was one of the biggest challenges, you know, even us kind of proctizing things like this was, you know, when you're asking questions like, Hey, tell me every time there's a pricing discussion that doesn't go well. Well, how many of your meetings have that? Maybe 1 %? Hopefully. Hopefully it's not like 20%, right? I would say it's like 0.2%. Well, gosh, then you don't need a really high hallucination rate for most of the content you get back to be hallucinated, right?

20:42Like the more you're looking for needles in a haystack, the more likely, more painful the hallucination problem becomes. And so, and so, you know, it's kind of funny, uh, GPT five last year, um, was kind of viewed, I think commercially as like not a very impactful major release. But it was actually really important to us because one thing they fixed in that release was hallucinations. They dropped hallucinations by like 85%. And that actually opened up a whole bunch of use cases where it's like all of our use cases, a lot of the interesting ones are needle in the haystack type problems. And that's why the dropping all your transcripts in cloud doesn't work is because one, the longer the context window gets, the less quality it gets.

21:23But two, sometimes 10 ,000 meetings just not going to fit into that context window. And so you have to employ a multi-step process. And if one of those steps involves some agent that might hallucinate a lot, well, everything downstream from that part of that process is going to be terrible. Right? right yeah yeah so it's funny you're talking about the new models i just noticed i don't know if i'm like i just woke up one day and opus 4.8 is now out like it's great like it was sauna it's it the the speed at which ai is is being produced and building i've never seen it before no me either in and i think and i feel like people are like not really like seeing it i does i try to explain to people like it it's scary if you're not understanding it and you're just sitting back and thinking that like we're going to live in a world that is like that you think is going to exist it's not like the new world we don't i'm sure you can agree like even you as being such a visionary and seeing the future like very hard to see what this world is going to be in the next five years there's going to be new jobs new role new new new things that we don't even have an idea or concept of that we're going to be doing.

22:32Do you have any suspicions or any, have you thought of any ideas of things that you can see how it would be different for us in the next five, 10 years? I mean, I think there's a couple of shifts. I mean, one, my buddy Emmett, who's on Twitch and now runs this AI company called Softmax, talks about, I think there's a really good analogy where he describes models as kind of like, you know, certain level of education where it's like GPT-3 was like an eighth grader, right? Yeah. Yeah. GPT-4 was like a high school student. GPT-5, you know, like, you know, it kind of says like, you know, again, four years ago, we were at eighth graders doing things.

23:09OK, what stuff would we delegate to an eighth grader? Not a ton. Right. OK. High school student. OK. Now we're at kind of like kind of like unlimited grad students kind of thing. Right. It's kind of like the state of the art. Right. And so I think if you think about like truly think about this as what would you hire a grad student intern to do? it really shifts your, your mindset on a lot of these things. Right. Um, you know, they're still going to make mistakes. And that's where I think the one interesting part is like, what does the grad student lack? It lacks business experience, business acumen.