Beyond note-taking with Fireflies

0:29I'll see you next. drops, behind-the-scenes content, and AI insights. You can learn more at practicalai.fm. Now, on to the show.

0:48Welcome to another episode of the Practical AI podcast. This is Daniel Whitenack. I am CEO at prediction guard. And I am joined as always by my co-host, Chris Benson, who is a principal AI research engineer at Lockheed Martin. How are you doing, Chris? I am doing very well. Got to spend the morning with some folks down at Georgia Tech talking AI. And you're headed my way. After this, we're going to meet up at the Midwest AI Summit, which is as of recording this happening tomorrow and also as recording this still experiencing government shutdown and travel problems in the US. So hopefully, I'm hoping you make it our way.

1:31But yeah, excited for the summit. It's going to be going to be a fun time to be together in person and meet some listeners, meet some AI enthusiasts and AI curious folks. So I'm excited for that. I'm encouraged because our guest was on a plane earlier today and he got through. So I think we should dive right into the conversation here. Yeah, yeah. Sounds good. I am very excited because we have with us Chris Schramannini, who is co-founder and CEO at Fireflies AI. Welcome. Hi, Dan. Hi, Chris. It's great to be here. Thanks for having me. Yeah, it's really, really great to have you on. I see, of course, Fireflies join my meetings all the time, which is cool, sort of Firefly call assistance and recording and that sort of thing.

2:21Of course, this is probably something like it was somewhat early on, at least in the public perception of ways people were using AI that kind of impacted maybe their day-to-day were these kind of meeting assistants that have taken various forms and people have tried various things with these. So as we're sitting here in November 2025, like what is the kind of like, how would you pick apart that industry from your perspective? Like what are people trying? What are various approaches to that? What is the state of that technology, etc? For me to really appreciate how far we've come as a company and as an industry, it's good to even look just five years back and maybe even go back a few more years before that, where we started in 2016.

3:11We were working on a whole host of different tools all around the general AI space. And remember, this was before ChatGPT existed. This was before any LLM or crazy AI tools were available. It was when me and my co-founder were reading articles on deep learning and sequence-to-sequence applications and how to make it work. We were using crude technologies in today's standards like BERT and really manual natural language processing libraries around that time. So it was a very different time when everyone wished and hoped that you could have AI understand conversations. And that was the essence of what we were trying to build.

3:55At the time, we didn't even call it an AI note taker. We were trying to build this general AI secretary or EA. and we needed to pay our bills. So we said, we've tried a bunch of different products in this space. So we need to figure out what's the thing that people are willing to pay for. And wouldn't it be a great idea if someone for$100 a month could get an executive assistant with a human in the loop and some AI in the background. So we went about building that, testing that out, learned a lot and realized having a human in the loop is there's no way this is going to scale. If we want to eventually build this into a platform used by millions of people every day, taking notes for you 24-7, capturing all those transcripts, there was just no way that human in the loop was going to build this sort of company.

4:49It was also right around the same time companies like Scale.ai were created where they were doing data labeling and the human in the loop processes. So we took inspiration from our peers at the time, but our ambitions were, I think, a little too grandiose. And we said, do we really want to build this business? Like the max we might be able to support is like 5 ,000 people, 10 ,000 people. And it's going to be an operations heavy business rather than an AI business. So that's when me and my co-founder said, you know what, we validated there's a market need for this. And within like trying it out for 10 of our close friends who were paying us enough money to pay for rent in San Francisco at the time.

5:33I feel like they were just, they felt bad for us. My co-founder got a place for$750 in SF. That's probably the biggest hustle at the time. And we were like, this is the only way to make things work is we just have to like rent our time. So as we built that, we validated it very quickly. And we did it without writing a line of code. After we did that experiment with like 10 of our friends, we said, if we really want to build a serious business, we know there's a like a product market need here. Before we would write code for six months, we would ship something and no one would want to use it. And we would be wasting a lot of time.

6:11This is the first time where before we built anything or wrote a line of code, we validated the market and we ourselves were the product, right? So fast forward, we ended up building the product. There was no category for AI note taker at the time. And that's when we created Fred, the AI note taker. We started with simple promises of, hey, one click capture your meetings. You don't have to deal with like native recording on Zoom because every user, when they were using like local recordings, there was a limit to how much they could record on Zoom and other platforms. Some platforms like Google Meet didn't even offer an ability to capture meetings.

6:48So we started with a basic function of how can I record my calls so I can go back and play through them. Then we focused on transcription, but it was really expensive at the time and not very accurate. And we said, that's okay. Even if transcription is not perfect, as long as I can just search, index through the conversations and just go back to the general timeframe when we talked about dates or months or something, some important keyword that's good enough. That's how bad the tech was back then. So we built like a way to record your meetings, a way to search back through your meetings. Then as it got better, we built the transcription layer that you can actually read.

7:27Initially, there was no dashboard. It was just an email that we would send out. And then afterwards, we started building our own task detection and action item and keynote detection engine by hand. That was brutal. So we wrote the code for that using some off the shelf stuff and some custom scripts that I wrote at the time. And fast forward, we launch in 2020. So it was a four year struggle of trying to figure out what we were trying to do. We launched in 2020 still no chat GPT around that time. But the core product of being able to capture search and see some general bullets or sentence fragments was interesting enough that we got our first couple paid customers by the end of 2020.

8:14So it wasn't until 2021 that we actually started making revenue. And then fast forward, COVID happens, right? And so we get like this proliferation of free users, and then some of them turning into paid users. And then you accelerate through to 2022, we get early access to OpenAI's GBT 3.5. Vinod Khosla, who is an investor, also happened to invest in OpenAI. That really opened the floodgates for us. Again, LLMs were very expensive at the time, but we said, you know what, let's go ahead and bring this in. It can't be that bad, right? Like whatever we're doing today, it's probably going to be better.

8:54It blew our mind. it changed the complete technology. And then from November 2022, so exactly three years ago to today, it's been an absolute rocket ship. We've never looked back. Company has grown, accelerated, crossed seven figures, and then eight figures in revenue and scaling beyond. So it's been very transformational. So I think it's fair to say we were lucky to be at the right place at the right time. But the more honest answer was, we just showed up a little too early and we just tried to survive for what it's worth. And eventually the bus came around and we were able to hop on the bus.

9:34So, and then create this whole category of AI note-taking. And now it seems like the most obvious generative AI use case. Like if you're a company is working on something and you need to pivot because it failed, the first thing you look at is maybe we should just build another AI note taker. Seems like it's working for Fireflies. We should try to copy them. So it's a very different space now. I'm curious. So that's a fantastic, I love the history of that. And I love that the kind of the incremental development of both the business and the technology, as you described it. One of the things I'm wondering is like, it seems like early on, you know, you talked about that in the beginning, you would write code for six months and ship it and people weren't using it and you learned the lesson of kind of going and making sure you had a market before you did a major commitment on that.

10:26And so, and as you've talked about the evolution from that point on, at what point did you start to see users changing their own behaviors in response to how you're, like whatever state your product was in at the time? Like what were the things that you noticed where users would respond and they would change and adjust their own behavior based on the integration of the product into their own workflow? And how did that behavior change evolve over time as your product got more and more sophisticated and capable? All the other products that we built were easily buildable, but there wasn't a clear market demand end for it.

11:07And we always looked at it from, is this feasible from an engineering point of view, rather than is this something that the market cares about? Is this solving a killer pain point for customers? So when we decided to build Firefly's the note taker, we said, let's forget what is even theoretically possible right now. Let's figure out what customers want and we'll work backwards from there. That was a fundamental shift. So when we had this experiment where we bootstrapped it with just my co-founder and I and a dozen friends where we were testing it for them, like we were the software. We were everything at that moment in time.

11:47And that was before there was a company really built. There was no business plan, but it was, is this something that's valuable enough to do? You validate it, but it's one thing to validate it with humans and now another thing to build the technology. Both my co-founder and I were technical. We both liked solving hard engineering problems. And credit to Sam, my co-founder and CTO, went to MIT, studied aeronautical engineering. And for him, computer science is like the easy stuff, right? The aeronautical engineering, he was working on drones, autonomous vehicles, unmanned vehicles back in 2015, 2016.

12:25So I think like this is like, if I can solve that problem, how hard can this like machine learning problem be? That was like, again, we got a little carried away, but it was a very hard problem. But I think a lot of the best technologists in the world, they worry about the technology part afterwards. They're like, great, if it's a hard problem, we're going to find a lot of really smart people that want to get behind it and solve it. And we did the best we could with the technology that was available at the time. So to even be able to capture calls, stream them into the cloud, record them, store them, index them, search them, that itself was a hard problem with the resources that we had back then.

13:08And it was also very expensive to do all of this stuff. Transcription has gone down like 10x since that time we launched. So that's been an enabling factor for us. we've scaled our own infrastructure over the last five years where we manage our own bare metal servers. It's the volume is insane, right? Like we run maybe a great, we're a multi-cloud platform, but we run more of our traffic off of our own servers than on Google or AWS. So we've gotten to this economies of scale and done all of this hard work where the end user, they just look at Fireflies and says, oh, wow, it's$10 per month. And I'm getting this note taker that's going to join all my meetings all the time, every single day around the clock.

13:54Doing that in the past would have cost thousands and thousands of dollars to support. Like it doesn't make sense to charge someone$10 if it's going to cost you$1 ,000 to support that user. So the business innovation, the scaling itself was a, I think a miracle. That was like a masterclass in how the engineering team got behind and solved it. Cause not only you have to optimize for quality and reliability and uptime, but you also have to make it affordable enough where someone's going to be willing to pay 10 bucks a month. And that was like a big feat. And since 2023, we've been profitable ever since we've started hyperskilling.

14:33So we've done all of this while being profitable. And we didn't touch any of our series A funding that we had raised along the way. It was just our seed round. And And that initial seed round was enough to like seed the initial free users, figure out our monetization strategy, build out our own infrastructure and get it to a point where we found product market fit, we found customers and we found revenue. So, yeah, that was the different thing this time around was we did not stop ourselves or limit our dreams from what was technologically possible at the time. And we absolutely got lucky because we might not have been able to reach the success that we have today had it not been for LLMs.

15:17But one way or the other, maybe we would have been two years delayed, right? Eventually, this sort of technology would come out because that's how fast the industry was moving. And again, the credit was definitely to Sam on the forethought because every time we went to investors, they would tell us there's no way transcription is going to be accurate enough. There's no way transcription is going to be cheap enough. In fact, many investors told us, you can make good money with this human in the loop business. Why don't you just go turn that into an actual business? Like you ran this cool experiment with your friends.

15:47You should turn that into a business. And we were very strongly opposed to that. We said we want to build a software company. We don't want to deal with operations. And this is also from some battle scars from the past when we tried to build a food delivery app in college and the logistics behind it. So we said, never again are we going to do anything that requires like logistics and on the ground stuff. So it was a series of optimistic decisions. And the technology ended up catching up to where we were vindicated.

16:37So friends, you know that feeling when your team has brilliant ideas and you're excited, but these ideas, this innovation gets stuck behind endless meetings, scattered documentation. And when you're moving at this speed of bureaucracy instead of the speed of innovation, we've all been there. The gap between idea, execution, impact, all that good stuff, it is killing your team's progress. But here's the thing, simply throwing AI at a problem doesn't help. It just makes things messier. That's exactly why you should check out Miro. Miro is powered by AI and it transforms teamwork that normally takes weeks into days.

17:14Now, I've used Miro to plan different ideas out, map complex workflows, and I've even generated fresh ideas from interviews all in one single place. And that is the power of Miro. It is like having a second brain with AI that actually thinks like you, like your team. No more switching between 10 different tools. Help yourself, help your team get great things done with Miro. Check out Miro.com to find out how. That is M-I-R-O dot com. Again, Miro.com.

17:50Well, Krish, I'm wondering, you kind of alluded to the evolution of the technology, which of course has played a key role in that. And for our kind of AI practitioners and those that are really curious about more of the kind of AI engineering side in our audience, could you just kind of in generic terms, I'm sure you can't share everything about how everything works, but you had mentioned this evolution where it was kind of at first these transcription, earlier transcription models, which weren't that accurate. and maybe you training or building your own kind of set of other models that would do kind of task specific things like, you know, action item detector or whatever that is.

18:35So how does that kind, how does the approach to something like an AI note taker now with the, you know, the models that are available differ both in terms of like what is easier, but maybe also there's certain things that are challenging now, which you didn't have to think about before because of how the technology has evolved. We'd love to hear that side as well. If you're using transcription engines today, like Whisper or one of the ASR providers today, and even some of the big ASR providers from the Googles of the world, Microsofts of the world, you don't have to massage the output as much.

19:13That's exactly the word that we had to think about because the raw output was pretty crappy and we had to fix the grammar. We had to fix the punctuation. We had to figure out speaker identification at the time. So you're just getting a, you know, garbled up string of text and you're trying to make sense of that. And you also have to figure out how to identify certain type of words because for certain industries, certain work domains, there are different types of words that are out there? How do you deal with filler words, pauses, ums? Like we speak a lot of filler words. I feel like working on Fireflies has improved the way I speak and helped reduce the number of filler words I use because the first versions of Fireflies, every time I would read the ums and the ahs from my own meeting notes and recordings and transcripts, I'd be like, wow, I said it that many times.

20:05So we ended up building a little filter back in the day that told me like how many filler words I used in a meeting. And I would instantly like learn from that and say, oh, wow, I need to learn to speak slower. I need to be more enunciated. So the original text that was generated from transcription engines wasn't good enough. So we had to build a bunch of our own filters on top to clean up the output that was generated. And then we had to build other layers on top that would classify different parts of the meetings and then pull out what it thought were action items what it thought were bullet point notes, what it thought were key ideas.

20:43And at the time, it was very extractive, meaning you're just extracting pieces from this garbled up text, and then you're calling that notes. And it's a shame, but at the time, that was good enough for a lot of people. So we had to build, rebuild, re-architect things many times over, whereas if a person was doing something today off the shelf, you don't have to do all of that. Like you can get to like 80 % pretty solid just using off the shelf parts, right? And that is something that we didn't have the luxury of. Now, you may ask, well, Krish, like if it's so easy now to build something 80%, 85 % off the shelf, like what's the competitive mode?

21:27Like everyone should be able to do this. And I think the part there that we were fortunate enough, one is going really deep on the problem because the other 15 % actually takes a long time. That differentiates good versus great. And you have to really polish the product. And those things take a very, very long time. And you have to learn from your users. The roadmap that we've been able to build out beyond the core tech for teams, for enterprises, that takes a lot of work. You have to think about access controls, privacy, storage. How do you think about multi-tenant? We're one of the first companies to get SOC 2 compliance in the AI meeting assistance space.

22:10And then HIPAA for doctors that want to use Fireflies. We offer this option for enterprises called private storage, where you can store Fireflies meetings inside your own server storage containers, which was another architectural change. If you don't architect it correctly in the beginning, it makes it very, very difficult. all this administrative sharing features, team features, which makes Fireflies more valuable as more and more people inside your organization start using it. Search was a really big problem that you have to solve over time as well. So all of this is enabled by great transcription and being able to understand language.

22:48I absolutely agree with that. But then all the use cases you have to build on top of it, like our Ask Fred example, where you can, instead of reviewing the notes or the meeting, you can just ask Fred questions about the meeting and it will catch you up on everything that happened. Right. So that required a lot of stuff like figuring out search from, from scratch and making that really effective. So, yeah, I think it was a big journey in terms of building out like these core building blocks. But as we were able to do that, when we started, it was a blue ocean and going back to, I'm like Pogo sticking back between the commercials and technology part because both are really important.

23:29If you want to be a great artist, you don't have to think about the commercialization of what you're building. You can just think about great art. But to build a great business, you have to be a good artist and you also have to figure out, can someone pay me money for this? We were fortunate enough at the time to establish the AI note taker brand and be one of the first companies to start championing that word. And now it's an entire category. Build that AI note taker, get to distribution, get to millions of people using Fireflies. Today, like tens of millions of people get notes every month from Fireflies.

24:07That in itself, right, distribution is something that is super important. And because we started early when it was less obvious, we were able to make the most of it. Like there's no point in going to the gold rush after it's been announced and after someone's made a ton of money off of it or a killing off of it. I think like being there a little early helped us maximize the impact because for every end note taker, every additional note taker that comes out, it becomes progressively harder for them to stand out in the crowd. And there is something to be said about these markets where distribution is one of the most important things to build in a PLG flywheel.

24:50That's why you don't see so many Calendly competitors. People still use Calendly. Like it's one of the de facto platforms that everyone uses. In our space, there's probably like three or so big players with Fireflies being one of them. And then each has its own like merits. Fireflies, for example, is very much focused on teams and businesses. And if you need robust integrations and workflows and admin controls, you come to Fireflies. Whereas if you're looking for more of a prosumer type product, you'll go to one of the other platforms. So we had to pick what we wanted to do and go really deep in there.

25:25A good parallel I like to use also is the project management space or the CRM space, where you had Salesforce build a massive enterprise-scale business. And then HubSpot comes along and takes a huge market out of the SMB business. Both are tens of billions of dollars in market value that's been captured. In project management, you have the Asanas of the world. You have the Monday.coms of the world. You have all of these other project management systems that follow. So we've found our niche. I wouldn't even say it's a niche because our niche is anyone that's a knowledge worker that works inside of a team.

26:00So we specifically say no to consumer grade use cases like university students, teaching, like those sort of stuff. Like our bread and butter has always been like, if you're a team, and it doesn't have to be in tech either. Like a lot of our customers are outside of tech, which was also really fascinating to see. But yeah, it's been a very long, tedious journey. I wouldn't have believed three years ago we'd be where we are today. But to answer your question on tech, yeah, it's been one of the most fascinating things about this space because now it feels super easy and everyone can do it. You hit on so many things there that I'm interested in.

26:42One of the things I think that I just learned something from you definitely on and that I will probably try to share. I'd like to generalize it towards not strictly a Firefly thing because I think there's a lot of people in our audience that can learn from this in whatever industry they're in. Is your pursuit of a sustainable competitive advantage and what that meant to you. And it sounds like in your case, being there early and having to solve the whole problem and not, unlike people that would come in later and get to that 80 % easily, and you have that last 15 % or 20%, which is really hard.

27:20you were well positioned to develop the expertise in your organization and to find the niches that you wanted to service and get their first and best with that expertise to do that. So that even as the space has developed other competitors, you were able to hold them off and hold your niche and be a powerhouse in that way. Is that a fair representation of kind of what you were just saying? Yeah, this was one of those situations where being early definitely helped. And then having the time to keep building and refining it and listening to your users over time. A lot of people claim that SaaS is dead, everyone will build their own SaaS.