#2280 Read.ai is adding 50k users per day

24:07We spent a year building models, really foundational models when it comes to a multimodal approach to communication. And we said, okay, we've gotten a really good spot. Let's publish it. Let's make it free. Because we didn't want to go into like, I'm going to try to sell you into this and then not just ask you to install this and give me your time, but also give me your money. I said more, if we can't get traction when it's a free product, how are we going to get traction if it's a paid product? Okay. So what did you use to build the first version? And just for people's context, I always look at mobile as, what was it, 2007 when the iPhone came out, right?

24:43So that's a good indication of when the start time was. For AI, ChatGPT's launch November 2022 is obviously not the start of it, but it's the start of suddenly the hype. You launched the year before to give people a sense of when this was. What did you use to build the first product? Yeah. So the first product was, we actually looked for meetings that were available in the public domain. So being able to pull in those different conversations. But then we realized if you can't see the other person's reaction when someone is saying something, that creates low value because I'm only analyzing the person that is speaking.

25:16And that's not that interesting. I want to see how people react to the words that are being said. So we had to go out and find additional data sources where we could see reactions. We hired actors to go in and say, like, read this script and read this out. Then we realized that's too expressive because they're acting for a movie. So it's like no one smiles with a giant smile like a joker. You're going to do a subtle smile or a smirk, and we need to pick those things up. So we started to go in and try to find these data sources, create these data sources, and that became our training data set. And then from there, we kind of tuned the models as we went.

25:49Early days, we used things like Mechanical Turk to go in and say, hey, what is this A versus B? Does this look correct from a labeling perspective? Then we started to go in and hire more people to go on the ground and watch full meetings that we generated internally and say like, hey, what do you think about this? What is your takeaway from this? What is the action item that you have in mind? So all those things combined kind of got us to that first model. How'd you get the, where did the first AI models come from? We built them in-house for us. So it was all in-house models that we built. And it was really expensive.

26:19When we started, we were lucky enough to raise a large round. We raised a$10 million round led by Madrona for a seed round. Most startups can't do that. But we had that in place so that we can invest in building that ground truth data, where we can go in and pay people for their data, where we could say, hey, this is the use case that we want because we want to build foundational models around engagement and sentiment. What's your background in AI? How did you know that this was possible? I think my background in AI was more machine learning. So early in my career, I worked at a company called Faircast.

26:52We did airfare price predictions. So in the old days, this is really old days, you'd actually have to have boxes in different areas to actually go on the Internet, pull data down, and you'd build these things called scrapers. And they would go in and say, scrape the airline ticket prices from all these websites on a daily, hourly basis. And we actually built a model that could predict airline ticket prices. And so that was really kind of early days machine learning, a little bit of AI, depending on how you want to look at it, to go in and say we can predict airline prices based on date, time of day, routes, weather, all these different variables that come into play.

27:24Then when we started Place, kind of very similar thing. If you remember the early days of the Apple phone, that blue dot would bounce all over the place. It would normally be a giant blue dot that you're in somewhere in this neighborhood. You go to Chicago, you'd look like you're on the other side of the river. And so with all those noisy data points were signals for us. So we actually built models to go in and say, hey, it says you're across the street, but if I send you a push notification to check into the nearest location that you have, where are you? Oh, you're at the McDonald's, not the law firm across the street.

27:56That's now training data. So we built this check-in aggregation app where you can check into Forceware, Gowalla, Google Places, Facebook Places, and you can be the mayor on all the services. So we had millions of users that were just checking into these places and giving us training data. And for them, it was like a fair trade. It's like, you get my location data so I can become mayor. We get your location data. We can build better models. And get more accurate location than what Apple would have for itself to give to its developers. Correct. Because we could think of what the signals were. Because then you can start, this is the metadata that's really important, where it's going in and saying, hey, it shows you across the street, but the Wi-Fi signal is stronger.

28:34This is 2007 or 2011. The Wi-Fi signal is stronger here, but there's 12 different signals. Which is the strongest signal? So you actually want to collect not just that you're connected to the Wi-Fi or that it's available, but which one has the strongest signal. Hey, the accelerometer and gyroscope and compass, they're moving really fast. So it looks like you're moving at like 60 miles per hour. You're probably not here. you're probably in the train tracks downstairs or in the underground. That's what's happening. So you're able to stitch these things together where we called it dead reckoning.

29:03It's a model or it's an approach where you can go in and say, based on the number of steps that you take, based on the elevation that you go to, based on all these variables, we can predict where you are without actually having your GPS location data. But then if you add GPS on top of that, it becomes even better. So it really is this concept of weak and strong signals to recommend or to predict where you are. And that's what Place did, the company. Place did. And I think it's the same thing for meetings. It's going in and saying, are you looking straight at? Are you looking away? What's your rate of speed in which you're talking?

29:34Five minutes into a conversation, you and I might interrupt each other because we're still getting a cadence. We call it synchrony. But after five minutes, if you're still interrupting each other, that's almost intentional. It's not a good conversation. You both aren't paying attention or you're both kind of irritated or agitated in terms of that conversation. so we can see that. Transcripts don't pick that up for the most part. No. We're able to do that and say, take those strong and weak signals. Here's another example. We do more than just meetings. We do emails, messages, et cetera. If someone emails you and you email them back within five minutes, almost every single time, because this person is important to you, significant other, maybe a client, a prospect, we'll go in and we'll recognize that.

30:15If you're on a call with them and you're paying attention to them, we'll pick that up as well. So we'll tie those two things together. is those two-week signals now become a very strong signal to say, this person is very important in your life. So whenever something comes in from that person, you need to prioritize that up higher. When they're speaking in a meeting, even the one that you're not in, you need to prioritize that up because you care about what they say. And you wouldn't have been able to do that if you didn't stitch together. Are you thinking of doing email too? Like it sounds like you're thinking of making email more useful by taking all these signals and then popping up the more important emails to the top of an inbox, but you're not thinking of creating like a superhuman, are you?

30:54We're not creating a superhuman. So we think you got to work within your existing workflows. We're not trying to change your workflows. That's why we send email updates for your meeting reports. But we are bringing in that context. So in tiny now, for Gmail, as well as for Outlook, you can actually connect your accounts. And we have millions of accounts that people have actually connected. And now we've got all the emails that come in and that you send out on top of the meetings that you have. So imagine you and I are having a meeting, there's four action items. Those action items in a silo of a meeting, unless we meet again, they never got completed.

31:27But in an email, if you followed up and I said, hey, here are the three PowerPoints that you asked for, plus here's a quote on pricing, now that action item from the meeting is complete. So you need that full level of context. So I think going back to your original question around like with the meeting note takers, why won't you have, why do you believe that you're not going to have siloed for product, for sales, for client success. Well, that information has to go to different organizations. So if the sales team goes and says, hey, I'm getting this constant feedback that this feature isn't working correctly, you want product to have easy access to that information.

31:59And in that same concept, you want those emails to have easy access to say, did this thing get actually done? Did the follow-up actually occur? Did that file get created? So you can connect your Google Drive to read now. You can connect your Notion, your Jira, your Confluence, and not just to pull in data, but you create tickets. So one final example I'll give is we have a product called Sales Copilot, which goes in and pulls all of your data from HubSpot, Salesforce, as those are emails, messages, et cetera. And we will actually push you notifications to say, we think based on all these things that have occurred that you should move this deal from 25 to 50%.

32:34We'll give you the reasons why. And you might not have even updated HubSpot or Salesforce yet. And you could say, oh, this actually makes sense. You push a button and it's done. So we've only had that product out for three months, over$50 million in deals have gone through that where we've initiated a change in status. See, that makes total sense to me. But I guess in my mind, it still could be a sales-only tool and another tool could do something else. But you're right. One of the most annoying parts about sales is having to go and update software. It is so freaking frustrating. And the kind of person who's a good salesperson is not the kind of person who wants to sit down and start dragging a thing over or taking a note.

33:15It's so painful. I just want it all to be done. And then to come back to them and say, you know what, it's been a week since you followed up with this person. That's way too long. Go follow up with them. And based on what they said, here's what you should tell them. Ask them about their wedding. Ask them about the thing that they brought up. There's so much there. I would love for also all this stuff to get funneled into a CRM. This is my personal dream. I keep asking people personal questions in meetings, right? Just like I do in my interviews, I would love for all that to go into my even personal CRM so that I remember it, so that I can come back and check in and say, you know, last month you told me that you were going to Hawaii.

33:50How's that trip coming? Did you book it yet? You know? 100%. Not yet. And I think those nudges are that AI or that kind of recommendation or that notification is really what makes AI personal. That's where you need a human in the middle. It's that orchestration way to go in and say, I've got all these data sources. The AI is working for me. It's doing a job. and then it's going in and giving me that. Just say, hey, you should probably reach out. And if you want to reach out, let me draft that email. And we do this today. Let me draft that email for you. Let me put some context in there. But we don't hit the send button.

34:20I think you still need a human in the middle to go in and say, yeah, but that context is a little weird. I'm going to go adjust the sentence and then boom, I'm ready to send it out. But prior to AI, that email would have never gotten sent. I had no idea you did that. I had no idea you did that. You will also draft an email for me. We'll draft an email for you. I think that's the hard part about AI is a lot of it is not just the technology, but it's how you market it, how you distribute it, how you educate people about it. Because features are coming in so fast and furious that for us, I think one thing is like we're constantly launching features, but we need to sometimes take a step back and go in and say, what is the benefit of this?

34:57How do you apply it? How do we remind people that it's available to them? And we've had a number of calls with the search copilot where people have started going and saying like, I didn't know you had this. You should have told me you had this like three weeks ago because I needed it. And now I'm using it every single day. And we're like, OK, that's something we need to work on. Yeah, that is actually a really difficult thing to know what's what's available. And then once you get an answer of no, if it becomes available, you assume that no is still there. Here's here's something I think that is kind of lucky in this space.

35:25Zoom has come out with their own meeting note taker. It's it's I'm going to say you don't have to say because you're partnering with them. I think it's absolute trash there. They're just the basics of the fact that the email that they send you has no taste. It's just a bunch of text. You can't read it and make sense of it. It's not even as easy to read as a Wikipedia entry. And so you just move on. Their AI interaction is kind of confusing. Do you have a platform risk, though? Because they clearly care about this. At some point, Google is clearly going to care about it. I don't know if they have Gemini in there yet or not, but I see them kind of at the edges.

36:00what's the risk of them coming in and creating an okay enough product? I think we already know the answer to that. And that's going in and saying they've got products in place. If you only use Zoom, if you only use Google Meet, if you only use Microsoft Copa, you might get away with like, this is enough for me. But honestly, I think you're exactly right, where the products aren't necessarily focused after launch. It's going in and say, we've launched this thing. Here you go. The continued kind of investment isn't there because that's not their core business. They're working on the next thing to introduce Copilot to more people.

36:31They're working on the next thing to enable Gemini. And the really interesting thing for us has been like, there's not much loyalty on platforms. So for our users on a weekly basis, we see that they use more than one platform for the vast majority of our users. So it's going in and saying like, hey, now I've got a Google Meet meeting with Gemini that has a different voice from Zoom, has a different voice from Microsoft Teams. They don't talk with each other. The agents don't work together. They're just kind of in these silos. So I'm never going to look at this and it's a waste of time. And then the last thing I'll say on that is like from a growth perspective, when Microsoft Copilot launched in about 20, late 2023, 24, everyone was like, oh, you're in trouble guys.

37:11Like they're going to take over and no one's going to ever use that. We've seen more than a 15 X increase in Microsoft team meetings that we've measured since Microsoft Copilot launched for teams meetings. So it's going in and saying, It's almost the best thing if you're in the startup world to go in and say, here's this new technology. Every single major platform is going in and saying, this is available now. People try it out and they're like, this is good enough. There's a population. But there's a group that goes in and says, I need Tableau. I need Databricks. I need Snowflake. And that's the market that we're seeing that you're coming in and saying like, yeah, this is great starting point.

37:45But I need more than this to get my job done. And we don't have to spend anything on marketing when it comes down to that. Like that's going in and saying the platforms are marketing the benefits for us. They're saying, look, here's how a transcript can help you. Here's how I could help you. They the user says, OK, but I want more. And then they go look for more or they show up in a meeting where suddenly there's a read.ai AI agent and they say, let me explore what my friend is using and try that. That's a very viral market. Yeah. And you know what? There is no lock in with Zoom. there is a lot of lock-in with read and the transcript companies because once you have your stuff in there you want to be able to go and search every past meeting and it's just too convenient to stick with your platform so you do have a lot of stickiness built in right hundred percent and the stickiness is built in all the data is there and we're starting to release things like the agents that we're releasing next week where it's going in and actually going in and pushing in these agents that will go in and say hey your week was over last week?

38:44What were all the open action items that you had? What were the deliverables? What was important to you? What was important to your team? And on Monday, it comes in and says like, hey, here's an update on everything that's going on. And if you want more updates, just let me know. But it's automatically doing that job for you. So it's pulling in all that content and saying like, hey, Andrew, heads up. These are the four projects. These two things actually got done. This didn't get done. You might want to follow up on this. And it's giving you those nudges. So it's kind of what we talked about a little bit earlier.

39:10What are you doing to understand how users are using Read and what else they would want and what they're not? I guess what they're not using is easy enough. You check the software and you see what people aren't using and you maybe mark it a little more or understand that it hasn't worked. But the things that they need that they haven't expressed, how do you come up with those? Part of it is the metrics that you had talked about. I think the other part is kind of going in and just getting feedback. So because we have 50 ,000 new users every single day that are coming in, they're constantly trying these things out.

39:38and we add new features. We use a company that recently was acquired, StatsSig from OpenAI, where they're able to let us test different things out to go in and say, hey, should we focus in on Google Drive Connect or SharePoint or OneDrive Connect because people want to search their meetings and the files that they've created to make sure that we can connect the dots there. And so we do these tests. StatsSig lets us do that. We're able to see what the behavior changes are very quickly and we give a winner and then we turn that across the board. Or it's feedback where someone says like, hey, I need this in Portuguese.

40:10Brazil is our top five market for us. We would have never known that had it not been for the customers just coming out and telling us and saying, hey, we need Portuguese. And we're like, okay, are we seeing a big uptick in Portugal? I was like, actually, no, it's Brazilian Portuguese. It's a little bit different. And so we went in and we said, hey, are we going to build two separate models for Portuguese and Brazilian Portuguese? And we're like, yeah, I don't know. The user started saying like, hey, I'm using your European Portuguese version, and there's certain words that's not picking up. Can you make adjustments?

40:39Then we make the adjustment. Now we see where I think we are the biggest note taker in that market today. And we didn't exist a year and a half ago. And these people are creating accounts. Or Columbia is a great example where we never thought Columbia was a market we would go after. This is the country of Columbia, not the university. And so we're like, okay, why are we seeing all this activity for Columbia? And we dove into a little bit. We started asking people, we have about one to 2 % of Columbia's population using read on a daily basis and we're like this is weird like we don't even we have not targeted this up what's happening was university students would bring read into their meetings or their classrooms they would bring it into their study sessions etc they would kind of go in and create multiple accounts and we kind of let them do this because we were trying to say what's going on here and they would every single class now in colombian universities are actually recorded with read they're sharing that across the entire student body they're making it available and they've built this storage of intelligence for the university, for the classes, which kind of drove us to go in and say, okay, if they're using it for that, why don't we make it easier to do that so that you could have this system of record for productivity, for knowledge, for intelligence.

41:47So that if you have a question about a class that you missed, you can just go in and ask it just the same way, like, hey, I missed that meeting or, hey, what's the status of this project? Because I've been out for two weeks. You don't have to ask 12 people. You just go in and ask that system of intelligence. All right. I want to still understand what other possibilities there are for AI companies. You are someone who, I didn't realize this until I looked at your LinkedIn in preparation for this conversation, you invested in dozens and dozens of companies in addition to launching a couple of really successful ones.

42:19What are you seeing that's working for startups that are building on AI or building AI first companies? I think launching very quickly is kind of a thing that people said all the way back to the early days of Ycomber. It'll fail quickly, push it out there. If you launch a product and you're not embarrassed by it, you're not launching it in a timely manner. Those things are all true, but it has never been more true when it comes to AI. You have OpenAI, which has a problem called hallucinations, where it makes stuff up to the question that you were asking. And that application has close to a billion monthly users now.

42:55And so that goes on and says, people are willing to go in and say, I'm okay with mistakes. I'm okay with drastic mistakes. As long as you're clear that this is a new technology, this could happen. So you need to make sure and review these results. But if you were able to give me 98 % of the time, something that looks like magic, I will sacrifice the 2 % of hallucination. So I would go in and say like, be okay that your product is not perfect for early stage founders where you've got smaller products. Now at our stage, we have to be more careful. we have to put checks and balances because now we're the system of record.

43:28But it's not, give me, give me more of these. Like, do you have an example of a smaller business that had launched that did have some, maybe a lot of hallucinations and mistakes in the beginning, but look at where it is now. Yeah. So, any examples of what you've seen? So one is kind of on the creative editing side. So on the ad tech side today, my background is a little bit in the ad tech space. What you found was like AI should be applied to every single ad that goes out there in the world today. Because you can go in and adjust it based on performance, do it in real time. Hey, the shirt should be red instead of blue.

44:00It should be winter, not summer. And do these things. It hasn't caught on because you've got old school companies going in saying like, well, does it align with the brand tenants? Does it align with what we're doing from a messaging perspective? And who's winning is the people that are focused in on performance. These are people where it's like, hey, I'm going to drive results. And if the results prove themselves out, eventually we will go out and deal with like, does it align with the brand goals? Is this right color palette? Those things don't matter. So one has been around kind of video editing.

44:30There's one company, I can't talk specifics on their numbers, but they've done incredibly well where they do video editing. And at first it was kind of like 50-50 chance you'd get something that was great. But the 50 times that was great, it was done four to eight weeks faster than anything that you could do in the current world. So you would just go in and create more and more content. And now they are, I'd say they're probably top 10 video editing app out there. We're talking about Opus, I'm assuming? Not that one. There's another one that you'll probably hear about here very quickly. Has it launched or not?

45:00It's launched. So it's got millions of users. Yeah. Okay. And so it does video editing. And what I've seen with video editing is they'll take a big video and they'll create short clips. I'm imagining that's what you're talking about, right? It's more like you'll take like a eight second clip and it'll extend it out into the video where it's like a 15 second or 30. and you get a lot of weird stuff, right? Like if you've ever messed around a VO3 or VO2.5 or done it in ChatGPT, you get some really interesting things. Now, if you're a startup founder, you'd be like, oh, until it's perfect, I'm not going to launch it.

45:28Well, by the time you get perfect and you launch it, there's going to be a new model that's already going to do that by default. So you need to launch quickly to get it out in the market and people can then make the decision, is this good enough for my use case or not? You don't need to say like, I need it perfect. I just need good enough that people will say, this is valuable and I will take the hits when it comes to the hallucinations, the weird outcomes where it's like, hey, someone jumping up and down when I said sit down. Those are things that you can handle when you're able to get that magic.

45:58What are some ideas that you wish somebody would create in AI that maybe are some simple ones that no one's jumping in and doing? I would say from a simple perspective, it's not simple, but it is going to be table stakes is I've spoken with a few companies in the last election that happened where they thought AI was going to be a bigger contributor to noise in the marketplace. And what ultimately happened was it wasn't. It wasn't there ready for prime time. We're all AI savvy enough that we could say that image looks a little bit wrong. Something looks off there. It's not really happening. Other countries that don't have it, they actually ran into a lot of issues.

46:33The U.S. didn't talk about it that much. But in the U.S. itself, it was very marginal in terms of the impact that it had. I do believe in about the next year or two, it is going to be almost impossible to figure out, is this AI generated or not? And at that point, it's a simple problem in the sense of like, you want a certificate of authenticity. You want to say this is real or this is fake. And the ability to measure that is going to be incredibly important. In the same way with ad impressions right now, you've got viewability. Was that ad viewable or not? I think you're going to have the same thing when it comes to any type of video or text content where there is some kind of seal that says, real, AI generated, or a combination of both.

47:12I think whoever figures that out is going to build a very big company that is going to be something that we see every single day where it's like, hey, on CNN, that clip, there's a little seal. Okay, that's good. I should believe that. Would you have been able to do it if you weren't working on Reed with your background? There's a shot I could do that. There's a shot I could do that. It's a tough one. It's a tough one. And the big thing is that you have to get adoption. That's the biggest thing. You have to get people bought into it early. You have to have enough kind of distribution that people will contribute their content, will validate it, that you can get that level of scale.

47:50Because I've seen a lot of great ideas flame out. If you look at my portfolio on LinkedIn, half the company has been wound down because it just didn't work out for whatever reason. And it's not that they were dumb. It's not that the technology wasn't great. It's that they didn't have distribution. I think that is key at the end of the day. And then the other thing I would say is like, and this is where we are touching a little bit is, I do think there is going to be AI that will fill in the gaps when you're not available. So let's say you're on vacation, you're out for seven days, you don't want to get interrupted because you want to unwind, but there's something that happens at the office and you have the context.

48:25Right now, someone would call our text. He's like, hey, Andrew, I'm sorry to interrupt your vacation. Can you answer this? And you're like, yeah, it only took a minute of my time, but this is now a distraction. Now I feel bad. They feel bad. And I'm not going to give you the best possible answer. AI can fill those gaps where someone is out of office that I can go in and fill that gap. When someone leaves the company, right now, a lot of times when you leave the company with two weeks notice, you're spending a bunch of time documenting everything because it's a big waste of time versus actually doing meaningful work while you're there for the last two weeks.

48:52Right. Now imagine if I can go and say like, Andrew, do whatever you want to do, like get get the business going in a good spot. But the documentation, the training for the person that's going to take over free, we got this. Like those are small tasks that people haven't built out yet because that system of intelligence isn't available. Because if I only took your meetings, I don't have enough contacts. If I only took your emails, I don't have enough contacts. I need to build all of those things to get that full context. That I think is going to happen really quickly, like I'd say the next six to 12 months.

49:20Yeah. For me to give some kind of AI experience where I have access to my email, access to my Slack, access to my Google Docs and Notion with some restrictions so that you can start asking questions that give you private information while you're on my team, I'd be okay with that. But yeah, it feels like it saves you time. Yeah. I think people would go in and say, like, you're not going to bother me. You're not going to call me at two o 'clock in the morning because something broke because my AI assistant can go in and actually answer this for you. And if you think about it, we have out of office messages.

49:51That's the dumbest thing right now. It's like, hey, I'm out of office. Like, OK, what if I do now? Hey, here's four other people you could contact that there aren't going to reply to you in a timely manner. Or I could have my AI just give you that answer. Which is kind of what happens on Intercom and Zendesk, right? Where an AI can look through all the docs and give an 80 % answer 80 % of the time. Oh, yeah. All right. I think I got where this is going. You didn't tell me the exact number of users, right, Unread? We're the millions and we're growing. Like every single month we're hitting new records.

50:23All right. Millions more than 10 million? Not yet. Okay. All right. You also showed me a lot of things that I didn't know were in read. Here's what I like about read. Video. I can't just have audio. I love the audio-only apps. They're phenomenal. Like, Otter is really good. But I need to see the person's face when I said the thing or to clip it for the team and go, look, when I said this, look at how excited they were. That is a really nice feature. Here's what I think is helpful, but it's not changing my life, but it's helpful. The Calendly integration inside. Really nice. here's the part that is the ability to ask questions of my past notes to be able to go in and say what about this what about that um i had i don't know can i get live meeting notes while i'm in a while i'm in a session and start asking questions 100 so we've got a chrome extension that you can actually see a pop-up that comes in and says like hey this meeting is going on do you want to ask any questions do you want to catch up do you want to see can i just go into read.ai app.read.ai once I'm in and see my notes, see my transcript as we're talking.

51:26If it's in Zoom. You have that available today too. Yeah, so you can see a live meeting in there that you can pull in those notes. So this is where we have to do a better job with all the features, making them more easily discoverable. All right. And then there's also the thing that tells me about charisma in the call and all that other analysis, which I like. All right. Thanks so much, David, for doing this. The site is read.ai. I appreciate you. and I'm looking forward to doing more of these AI conversations. No, appreciate the time, Andrew. This was great. See you, man. Bye. Bye, man.