Spelade

  • Hi Hitchhikers,

    I’m excited to share another interview from my podcast, this time with David Kossnick, Product Manager at Coda. Coda is a collaborative document tool combining the power of a document, spreadsheet, app, and database.

    Before diving into the interview, I have an update on Parcha, the AI startup I recently co-founded. We’re building AI Agents that supercharge fintech compliance and operations teams. Our agents can carry out manual workflows by using the same policies, procedures, and tools that humans use. We’re applying AI in real-world use-cases with real customers and we’re hiring an applied AI engineer and a founding designer to join our team. If you are interested in learning more, please email [email protected].

    Also don’t forget to subscribe to The Hitchhiker’s Guide to AI:

    Now, onto the interview...

    Interview: Supercharging your team with Coda AI | David Kossnick

    I use Coda daily to organize my work, so I was thrilled to chat with David Kossnick, the PM leading Coda’s AI efforts. We discussed how Coda built AI capabilities into their product, and their vision for the future of AI in workspaces, and he gave me some practical tips on how to use AI to speed up my founder-led sales process.

    Here are the highlights:

    * The story behind Coda’s AI features: Coda started by allowing developers to build “packs” to integrate with their product. A developer created an OpenAI pack that became very popular, showing Coda the potential for AI. At a hackathon, Coda explored many AI ideas and invested in native AI capabilities. They started with GPT-3, building specific AI features, then gained more flexibility with ChatGPT.

    * Focusing on input and flexibility: Coda designed flexible AI to work in many contexts. They focused on providing good “input” to guide users. The AI understands a workspace’s data and connections. Coda wants AI to feel like another teammate—able to answer questions but needing to be taught.

    * Saving time and enabling impact: Coda sees AI enabling teams to spend less time on busywork and more time on impact. David demonstrated how Coda’s AI can summarize transcripts, categorize feedback, draft PRDs, take meeting notes, and personalize outreach.

    * Tips for developing AI products: Start with an open-ended prompt to see how people use it, then build specific features for valuable use cases. Expect models and capabilities to change. Focus on providing good "input" to guide users. Launching AI requires figuring out model strengths, setting proper expectations, and crafting the right UX.

    * How AI can improve team collaboration: David shared a practical example of how AI can help product teams share insights, summarize meetings and even kick-start spec writing.

    * Using AI for founder-led sales: David also helped me set up an AI-powered Coda template for managing my startup's sales process. The AI can help qualify leads and draft personalized outreach emails.

    * The future of AI in workspaces: David is excited about AI enabling smarter workspaces and reducing busywork. He sees AI agents as capable teammates that understand companies and workflows. Imagine asking a workspace about a project's status or what you missed on vacation and getting a perfect summary.

    * From alpha to beta: Coda’s AI just launched in beta with more templates and resources. You can try it for free here: http://coda.io/ai

    David’s insights on developing and launching AI products were really valuable. Coda built an innovative product, and I'm excited to see how their AI capabilities progress.

    Thanks for reading The Hitchhiker's Guide to AI! Subscribe for free to receive new posts and support my work.

    Episode Links

    Coda’s new AI features are available in Beta starting today and you can check them out here: http://coda.io/ai.

    You can also check out the founder-led sales CRM I build using Coda here: Supercharging Founder-led Sales with AI

    Transcript

    HGAI: Coda AI w/ David Kossnick

    Intro

    David Kossnick: ,One of our biggest choices was to make AI a building block initially. And so it can be plugged in lots of different places. There's a writing assistant, but there's also AI, you can use in a column. And so you can use it to fill in data, you can use it to write for you to categorize, for you, to summarize for you and so forth across many different types of content.

    David Kossnick: Having that customizability and flexibility is really important. I'd say the other piece more broadly is there's been a lot of focus across the industry on what, how to make good output from AI models and benchmarks and what good output is and when do AI models hallucinate and lie to you and these types of things.

    David Kossnick: I think there's been considerably less focus on good input. And what I mean by that is like, how do you teach people what to do with this thing? It's incredibly powerful, but also writing natural language is really imprecise and really hard.

    AJ Asver: Hey everyone, and welcome to another episode of the Hitchhikers Guide to ai. I'm your host, AJ Asver and in this podcast I speak to creators, builders, and researchers in artificial intelligence to understand how it's going to change the way we live, work, and play. Now, You might have read in my newsletter that I just started a new AI startup

    AJ Asver: since starting this startup a few months ago, a big part of my job has been attracting our first set of customers. I love talking to customers and demoing our product, but when it comes to running a founder-led sales process, prospecting, qualifying leads, And synthesizing all of those notes can be really time consuming, and that's exactly why I decided it was time to use AI to help me speed up the process and be way more productive with my time.

    AJ Asver: And to do that, I'm gonna use my favorite productivity tool, Coda. Now, if you haven't heard of Coda, it's a collaborative document editing tool that's a mashup of a doc, a wiki, a spreadsheet, and a database.

    AJ Asver: In this week's episode, I'm joined by David Kossnick, who's the product manager that leads Coda's AI efforts. David's going to share the story behind Coda adding AI to their product. Show us how their new AI features work, and give me some tips on how I can use AI in Coda.

    AJ Asver: By the way, I've included a template for the AI powered sales CRM I built in the show notes, so you can check it out for yourself.

    AJ Asver: But before I jump into this episode, I wanted to share a quick update on my new startup At Parcha, we're on a mission to eliminate boring work. Our AI agents make it possible to automate repetitive manual workflows that slow down businesses today.

    AJ Asver: And we're starting with FinTech in compliance and operations. Now, if you're excited by the idea of working on cutting edge autonomous AI and you're a talented applied AI engineer or designer based in the Bay Area, we would love to hear from you. Please reach out to [email protected] if you wanna learn more about our company and our team.

    AJ Asver: Now, let's get back to the episode. Join me as I hear the story behind Coda's latest AI features in the Hitchhikers Guide to AI.

    AJ Asver: hey David, how's it going? Thank you so much for joining me for this episode.

    David Kossnick: It's going great. Thanks for having me on today.

    What is Coda?

    AJ Asver: I am so excited to, go deeper into Coda's AI features with you. As I was saying at the beginning of this episode, I've been using Coda's AI features for the last month. It's been kind of a preview and it's been really cool to see, it's capable of. I'm already a massive Coda fan, as you know. I used it previously at Brex. I used it to organize my podcast and my newsletter, and most recently it's kind of running behind the scenes at our startup as well for all sorts of different use cases. But in this episode, I'd love to jump in and really understand why you guys decided to build this and what really was the story behind Coda's AI tools and how it's gonna help everyone be more productive.

    AJ Asver: So maybe would you describe and what exactly it does?

    David Kossnick: Coda was founded with a thesis that the way people work is overly siloed. So if you think about the most common productivity tools, you have your doc, you have your spreadsheet, and you have your app. And these things don't really talk to each other. And the reality is often you want a paragraph and then a table, and then another paragraph, and then a filter view of the table, and then some context in an app that relates to that table.

    David Kossnick: And it's just really hard to do that. And so you have people falling back to the familiar doc, but litter with screenshots and half broken embeds. So Coda said, what if we made something where all these things could fit in one doc and they worked together perfectly? And that's what Coda is.

    David Kossnick: It's a modern, document that allows you to have a ton of flexibility and integrate with over 600 different tools, uh, for your team.

    AJ Asver: Yeah, I think that idea of Coda of being able to one, integrate with different tools would also be both a doc that can become a table and then have a mashup of all this different type of data is something I've really valued about it. I think, especially when I was at Brex and we used to run our team meetings on Coda, it was really great to be able to have like the action items really formatted well in the table, but also have the notes and more freeform and then combine that with kind of follow ups.

    AJ Asver: And we even had this crazy table my product team where we would post like weekly photos and that's like really hard to do or in an organized way in a doc, and you'd never wanna do that in a spreadsheet. So, um, I love the fact that Coda enables you to combine all that different type of data together. So, Coda has that. And then it also has packs, which you mentioned too, right? And these are these integrations that allow you to like take data from lots of different places and put it all together.

    Story behind Coda AI

    AJ Asver: And from what I understand, Coda's AI products started as a pack, right? It was like this pack that someone had, Coda had built more as kind of like a hack project to get people to use the OpenAI capabilities inside Coda.

    AJ Asver: And I'd actually tried this too, and that's kind of how you guys decided to actually build this into a real, uh, native integration.

    David Kossnick: Yeah, totally. I'd say, you know, the first bet Coda made was on packs as a platform. And so maybe about a year ago now, we released an SDK where anyone can build a pack for Coda in their browser, in JavaScript and compile it and publish it. And so it can pull data from places, push data to places. And it's really been incredible to see people do this for all sorts of use cases we never even thought of.

    David Kossnick: And it made possible people, starting to experiment with AI in a much more effortless manner. And so someone did a kind of weekend project and published the first OpenAI pack and it really took off starting to see it get used for all sorts of different cases. And it got us inspired for thinking, Hey, you know what?

    David Kossnick: If we did something native that you didn't have to think about authenticating to external services, what if it could go much deeper in what context it had about your doc and your workspace in order to help you save time?

    AJ Asver: One of the things I really loved here is kind of a product management thing is like seeing this nascent behavior right, happening on the platform and then deciding that it was something worth investing in further. So what point did you guys decide like, oh, this is becoming big enough or popular enough where we should make an investment in, and how was that decision made?

    AJ Asver: Is that like a decision that like the CEO makes or is it more like kind of bubbled up from the bottom where like there was a team that saw this happening and was like, hey, we'd like to invest in this further. I'm really curious. Like give us a bit of like the inside baseball of how that happened.

    David Kossnick: There were a few moments. Uh, the first moment was kinda the weekend hackathon where a few people, I think it actually started on Friday afternoon. Some of the DALL-E API had been released by OpenAI, and they really wanted to start generating images in a doc. And so it took like two hours, uh, to basically create a new pack from scratch and have it fully workable inside of a doc.

    David Kossnick: And then we had a weekend blitz to basically ship it on Product Hunt. And Reid Hoffman, who's on Coda's board, big fan of OpenAI, was actually kind enough to hunt it on Product Hunt for us. Um, a bunch of really cool templates people had quickly built for it. And then about a month and a half later, we had a company-wide hackathon. This must have been back in December. And there was a ton of enthusiasm on about AI, partially from, um, the OpenAI pack. And so we explored like a dozen different ideas. Um, I think we won the People's Choice Award, the company votes on all the different hackathon pitches at the end of it.

    David Kossnick: And then, uh, at January we had, a board meeting and we showed off some of the, thoughts from the hackathon as well as some of, what people had already been doing in the community. And the board was really excited about it, and so we started up a much bigger effort within the company.

    AJ Asver: It's a really cool story to hear that, you know, what started as like a project that became like a hackathon, that, that became like a pack that was put together really quickly and just kind of an experiment then became like this big investment for the company. Right? And now when I look at the product, which, you know, you're gonna talk a bit more about, I can see how like it's, it could really become like a core kind of primitive of the Coda experience, just like a table and a dock and a canvases as well.

    AJ Asver: So, that to me is like really inspiring, especially for other folks that are like working in product that, companies that are like Coda stage that you can really like come up with these experimental ideas and they can end up becoming products. And also I think it was really encouraging to see that you guys kind of explored AI and like integrating into the product pretty early, right?

    ChatGPT and Coda AI

    AJ Asver: Like this was like pre-chat g p t hype, like, or just like as ChatGPT was becoming popular, but before GPT-4 came out, Christmas at least, I, I feel like the hype was just starting to simmer them, but not nec necessarily boil over like it is right now. talk me through what it was like developing the AI product into what it is today and how you guys kind of built into the product and, and how it works.

    David Kossnick: I look back on those early days and it's sort of, uh, amazing how much chaos there was in the market. You know, ChatGPT had just come out and it was incredible, and we were dying to get our hands on a chat based API but at the time, the only backend available to us was GPT-3.

    David Kossnick: They hadn't released an API for ChatGPT. There was nothing else really at that quality level on the market. And so I remember going and chatting with every major developer, API, AI company, every platform and being like, Hey, like, what have you got? What can we use? And we had a bunch of different ideas on UX, but we were kind of bottlenecked on what's possible.

    David Kossnick: So we started by building something with GPT-3 actually. And we said, okay, Chat's gonna come at some point. We don't know if it's a month out or a year out. Uh, we, you know, we gotta start moving now. Um, and we did a few experiments on it. Some just straight outta the box. And some that were very specific.

    David Kossnick: So actually one of my favorite was Coda has a formula language. It's incredibly powerful. People love it. It's also, um, a little complicated for people who are starting to learn it. It's a lot like Google Sheets or Excel. And so we had said, what if you could have natural language that just turned into a Coda formula?

    David Kossnick: And so we, um, we collected a data set for that. We actually crowdsourced it within the company. We took all of the company's internal staging environment, data of quota formulas, and we had people annotate what the natural language equivalent was. and we fine tuned GPT-3 for it.

    David Kossnick: And we built a little thing that would basically, you know, convert text to formulas. We were like, wow, that's actually pretty good. We realized, you know, some of the hard cases are really hard, but some of the average cases it does quite well on. Um, but it was definitely a mode where, because there was no sort of generic chat backend, we had to think like, feature by feature, what would we do for this exact scenario?

    David Kossnick: What are the prompts we would create, uh, and so forth. Um, and we got, you know, decently far down that path. Uh, at which point, you know, ChatGPT's API, which was called Turbo 3.5, was released and unlocked kind of a whole set of other use cases.

    AJ Asver: I think for people that, you know, may have forgotten by now, cause it happened so quickly, right? GPT wasn't available through chat until November. So you basically had to just provide a prompt and it would do a completion, but it wasn't fine-tuned in the same way it was right now. It wasn't, um, basically it wasn't as good at following instructions right as it is now. And so you had to do a lot more work to get it working, and then of course chat landed. Did it feel like you kind of were given this like gift where you'd be like, oh, this is gonna make it way easier. And then how did that kind of lead to where the product ended up?

    David Kossnick: was definitely a gift. We were super excited. It's also, as I'm sure you know, is like, a very double-sided, uh, sword. Um, you know, prompt engineering is hard. It's brittle. You sort of make a tweak and move. Sideways in, forward and backwards all simultaneously for different set of things.

    David Kossnick: Um, and so there's definitely a new muscle on the team as we moved into sort of turbo and GPT-4 on, how do we really evaluate which things it's doing well on, which is doing poorly on how do we make it really good for those use cases, both by setting user expectations and by, by changing the input, when we actually want GPT-3 with something fine-tuned.

    David Kossnick: And so it sort of opened up a whole kinda, new worms in the problem space, which was super exciting. And I think one of the things that, that got me really revved up about, uh, you know, Coda specifically as a uniquely great surface for AI is there's so many different ways people use Coda. So many different personas, so many scenarios.

    David Kossnick: It's an incredibly flexible tool. And so having a backend like, ChatGPT is really, really useful for a fallback. For any sort of long tail and unusual, surprising request, cuz it does really well at the random thing. And one thing we've discovered is, you know, for the very narrow set of most common things, it does pretty well too, but not as good as the more specialized thing.

    Making Coda AI work for lots of use cases

    AJ Asver: So as you were talking about, Coda being used for lots of different use cases, I noticed that because there's so many different templates in the gallery and so many different ways Coda has been used, there's like pretty big community of Codens, right, that are building these different types of Coda docs. How did you think about it when it came to adding AI into Coda to make sure it's versatile, versatile enough to be used in many different ways. much did that impact kind of the end design and the user experience?

    David Kossnick: you know, One of our biggest choices was to make AI a building block initially. And so it can be plugged in lots of different places. So you'll see as we get to a demo a bit later, there's a writing assistant, but there's also AI, you can use in a column. And so you can use it to fill in data, you can use it to write for you to categorize, for you, to summarize for you and so forth across many different types of content.

    David Kossnick: Having that customizability and flexibility is really important. I'd say the other piece more broadly is there's been a lot of focus across the industry on what, how to make good output from AI models and benchmarks and what good output is and when do AI models hallucinate and lie to you and these types of things.

    David Kossnick: I think there's been considerably less focus on good input. And what I mean by that is like, how do you teach people what to do with this thing? It's incredibly powerful, but also writing natural language is really imprecise and really hard.

    AJ Asver: Mm-hmm.

    David Kossnick: We had a user study early on. I remember it was super surprising.

    David Kossnick: Someone asked, our AI block how much money was in its bank account when it was in the person's bank account. And I was just blown away that they, it felt so knowledgeable and powerful. They assumed. Know that even though they never authenticated their bank account, they just like forgotten. It just felt like something that they, that we would expect it to do.

    David Kossnick: Um, and so how do you remind people sort of the universe of what's possible or not, or what it's good at or not? We have something very simple. You know, a lot like Google and Auto Complete as you type, you get suggestions underneath it. But a surprising amount of effort went into that piece in particular.

    David Kossnick: What do we guide people towards? Which specific types of prompts, how do we make the defaults really good? How do we expose the right kinds of levers to show people what's possible and make them think about those? Um, and I think as an industry, we're still pretty early on that for these large language models, like I think we're gonna see a wave of innovation on how do you teach and inspire people how to interact to to have good input in order to get the good output.

    AJ Asver: So we've talked a lot about that story behind Coda and AI, and it's really interesting to hear how you guys developed kind of the thesis around it and put into the product. Um, I think for folks that aren't familiar with Coda especially, I'd love to just jump in and for you to show us a little bit about how Coda works with AI and, and what it can actually do.

    Demo: superchaging your team with Coda AI

    David Kossnick: That sounds great. Yeah. Maybe I can walk you through a quick story about a team working together in a team hub with AI. And so this team hub is a place where different functions come together on a project, and have shared context.

    David Kossnick: So, a super common way people use Coda is collecting feedback. Um, all sorts of feedback, support, tickets, customer feedback, uh, sales calls. Um, and so we have lots of integrations that do this. They pull in Zoom transcripts and content looks a lot like this. It's really rich.

    David Kossnick: There's so much context in here, but it's really hard to turn this into something that's valuable for the whole organization. Um, I've spent many hours of my life writing up summaries and tagging things, and so wouldn't it be great if AI could just do this for me? Uh, and so here's a quick example. I'm gonna ask AI to summarize this transcript in a sentence, and here we go.

    AJ Asver: That was really cool. And I think what was really magical for me about it is not just that you can summarize, because obviously you can take the transcript, you can put it in GPT-4 and summarize. And there are other tools that do summaries too. But I think what's magical is when you combine that AI in the column, but with the magic of Codar's existing integration.

    AJ Asver: So like you have connected it to, I think it's Zoom, right? There's like a zoom

    AJ Asver: pack that's already outputting all the transcripts. So you've automated that bit and then you create this formula that basically runs on that column and then every time a new transcript comes in, I presume it just automatically summarizes it.

    AJ Asver: So that like piece of like connecting all those dots together, that's why I love Coda and that's why I think this is a really cool example of where kind of Coda shines.

    David Kossnick: That's exactly right. And one of my favorite pieces here for dealing with large amounts of data is just categorizing things. There's so many times I'm going through a large table of data, picking a category, so wouldn't it be great if based on this transcript I could just automatically tag. What type of feature request it was and boom, there it is.

    AJ Asver: where's it getting those feature requests from,

    David: Yeah,

    AJ Asver: tags? Is it like kind of making those or?

    David Kossnick: In this case, I already have a, a schema for what are the types of feature requests that I've seen, and so it's just going ahead and tagging all those things.

    AJ Asver: That, that's a really interesting feature too there, by the way, because now what you're doing is you're taking kind of the open-ended feature tagging problem where really GPT could like generate any feature tag at once and you're constraining it with this select list. And that's another good example of where if you did this in ChatGPT, you may end up with lots of different types of feature tags, but by constraining, you now end up with this format when you can now I, I assume go and like, organize these transcripts by feature tag because they're like actual each of those little data chips, right? And so it's now like segmentable, like

    David Kossnick: Yeah, and one of the cool things about it is you'll see this one is blank. That's actually intentional. That means it, it either couldn't find a match or didn't know what a match was. As there's plenty of cases where there's, you know, there's no real feature request or doesn't really know what to do with it, and it just won't tag anything either.

    AJ Asver: Very cool. Very cool. Okay, what What else you got?

    David Kossnick: So very common. Maybe the support team does an amazing job summarizing all the tickets, the feedback that's coming in, even tags things for you. And then as a PM you have to sit and read it all and think about it and say, okay, how should this influence my roadmap? Wouldn't it be nice if you could use AI to get started?

    David Kossnick: And so imagine you say, create a PRD for new image editor based on the problems in all this user feedback. Here we go.

    AJ Asver: Okay. No need for PMs

    David: Yay.

    AJ Asver: what are you gonna do after? this goes into production?

    David Kossnick: Well, of course it's a first draft. You know, you should always, uh, proofread. You should always change it. And so maybe AI can help with that too. Say, make this a little longer, um, and have it help me edit this PRD before I send it off. Um, there we go.

    AJ Asver: One of the things I love about this is for me, often when I was PMing, it's that cold start problem. It's like you are at like Tuesday, you know, you've got your no meeting Wednesday and you've gotta write this PRD in time for like a review deadline on Thursday because it's gonna go in product review on Friday.

    AJ Asver: Right? And you've gotta start it and you just keep putting it off cuz you've got back to back meetings. And then you get to Wednesday and you're staring at a blank screen and then you're like, Oh, maybe I need to go check my email. Right. I just think like more than anything else, this will just solve that cold start problem of just getting something down that you can start iterating on so you can just make progress faster.

    AJ Asver: And now what would've taken you three or four hours to get from like blank screen to PRD first draft is now probably gonna be a couple of hours because you've got that first version, you can kind of iterate on it from there. So I think this is gonna be a huge help for PMs. And, the cool thing about it is you're taking all this structured data right, from different places and bringing it into one place.

    AJ Asver: And one of the things we often did when we used Coda at Brex is like we would have, you know, like you had like kind of place that had customer feedback, Or you'd be aggregating different feature ideas in like a brainstorming, section and then you're kind of bringing them in here and turning them into a PRD.

    AJ Asver: So that's pretty cool.

    David Kossnick: Thanks. So another super common scenario is you have a team meeting at Coda. We do these inside of a doc with structured data, which I really love. They let people vote on which agenda item to talk about first, and you can even have notes attached to it here about what's happening. But again, what do I do after the meeting?

    David Kossnick: As a PM I spend a ton of time writing up next steps, but oh yeah, I can do that for me. That's awesome. Uh, one of the other things I do all the time is write up summaries. Um, what if instead I could ask ai, AI to do that too? And then of course I can send that out to Slack direct from Coda.

    David Kossnick: So outreach. One that we've already started doing at Coda is personalizing messages to key accounts. Um, and so let's say we have this, uh, launch message about this new image editor feature. We wanna tailor it based on the title and the company.

    David Kossnick: Uh, we can go ahead and get a, an easy first draft to start with here. Boom. Let's say we don't like one of these. I'm just gonna refresh this one. We'll get another example. Um, and maybe I want to go in and change this a little bit. Um, hope your family is doing well and using our Gmail integration. I'll just go ahead and send that email.

    AJ Asver: And that's actually gonna send that email now to the, person that you wanna do outreach to. And you just basically generated that email based on kind of the context of the person's job

    David Kossnick: Totally. And one of the really fun parts of this is it's super flexible. So imagine you had another column that's like family member names, or hobby or other things like that that you're not gonna find in, your favorite, uh, sales tool. Um, AI is really good incorporating that context. So having all your stuff here in the team hub, being able to pull that context in, um, is really powerful.

    AJ Asver: Does it generate tables too? How how does

    David Kossnick: You know, we showed a simple example here, which generates the table of target audiences. Um, but actually one, maybe I'll just show real quick, um, that I've been doing in my personal life. So very different kind of template. Um, we use meal planning, so I'm a vegetarian, so meal planning can be sometimes a bit of a pain with two kids, uh, making sure everyone gets exactly what they need.

    David Kossnick: Um, and so I made a quick template. Everyone in my family can go in and add. Their favorite ingredients and get out both a bunch, a bunch of ideas as well as, uh, specific dishes. And so this is an interesting case where it's just a normal table here. And I'll say, uh, AJ, what's your favorite ingredient?

    AJ Asver: Well, you know what? My kids love

    David Kossnick: Broccoli. Wow. Nicely done. cool. And then I'll go ahead and, uh, or auto update it here. Uh, so it has a bunch of different meal ideas and yeah, let's say I'll take, uh, spinach and cheese quesadilla. Um, and I'll add that one in here. Spinach and cheese quesadilla. Um, and AI is gonna start generating, uh, what ingredients are needed for that as well as a, a recipe and about how long it would take.

    AJ Asver: That's a really, really awesome hack. I think I need to start doing that, as well to just use it to generate ideas for, meals as well and meal planning. That's, that's very, very cool. And this is like a good example of also like how it can be helpful in like a personal setting too.

    AJ Asver: Right?

    David Kossnick: For sure.,

    Using AI to speed up lead qualification

    AJ Asver: I was not gonna bring you onto this podcast without you helping me with something. So one of the reasons I wanted to bring you on here is because now that we've started this startup and we're trying to get everyone to be very excited about our AI agents, I am doing what's called founder-led sales, which is where we kind of work out, okay, who are the companies that we wanna target?

    AJ Asver: And then at the very early stages we're trying to find like five design partners that we can work with. And they're kind of like enterprise FinTech companies, similar to like Brex where I used to work. And my job is to do the sales. Cause there's only two of us right now and Miguel's busy building the product. And so I gotta work out which companies will be a good fit and then we work out, okay, how do we get an intro to them? Maybe it's through an investor, maybe it's through a mutual contact. Maybe we, outbound to them because you know, maybe it's a company that we worked at before or something. And so I've been trying to work out how to use Coda to help me do this.

    AJ Asver: And I was wondering if you might be able to help me make my Coda CRM better with AI.

    David Kossnick: Sounds amazing. Let's do it.

    AJ Asver: I'm very excited about this because this is gonna save me a lot of time.

    AJ Asver: Over here is my Coda CRM and very high level. I have this list that I've got, and I think I got it from Crunchbase, um, with just a bunch of FinTech companies. And some of them might be a good fit, some of them might not. Oh, that's definitely not a FinTech company. Let's take that one out. Um, and I, and I'm trying to work out kind of one, how big are these FinTech companies? Either they kind of size where they would be an interesting fit for our, for our product, we generally try to focus on kind of growth stage FinTech companies, and then two, Would they qualify or not? Based on do we think they might need the product we're trying to build? And so I was wondering, maybe the first thing is I have this kind of list of the different, um, tiers of, of customers that I might want or the different types of companies that I might want to organize 'em into.

    AJ Asver: So for example, there might be growth stage FinTech companies, there might be early stage FinTech companies. And I wanna work out how I can take these FinTech companies that I have in this list and kind of categorize them by that tier. maybe we could, start there.

    David Kossnick: Yeah, sounds great. What kind of categories are you thinking about?

    AJ Asver: I think maybe we can start with kind of early stage FinTech growth stage FinTech and a financial institution. So maybe first I need to add like a column, right? Is that correct?

    David Kossnick: Yeah, sounds good.

    AJ Asver: Okay. So we can go do that, have a column after this one. And then do I make this a select list? Is that

    David Kossnick: Yeah, that's exactly right. So you could add, uh, select list

    AJ Asver: Great.

    David Kossnick: sort of a, a list of types or a list of items.

    AJ Asver: Okay, so then let's, let's pick a few different stages. So it's growth stage, FinTech, um, maybe series C +, and then maybe early stage FinTech Seed Series C and then maybe kind of traditional financial institution. Cause I think there's a few of those in this list. And then I know there's probably some crypto startups in here too, so maybe I'll put like, crypto is another one too. Okay. So I've got my select list. What do I need to do next?

    David Kossnick: Um, cool. So the, uh, actually before you forget, yeah, maybe rename the, the column there.

    AJ Asver: So we call this customer segment. Great.

    David Kossnick: So next, if you could just go to, uh, right click on that column and do add AI. Um, so the question is, what do you think would determine it? Is it, you know, one thing we could do is just give it the company name. many that are well known, it might do actually a pretty good job.

    David Kossnick: Um, you could also try giving it the, um, the description as well. And you could say basically, you know, what, what category does this belong to?

    AJ Asver: I kind of like the idea of doing it by description. That seems like a good way of doing it. So then would I just do this like, um, pick is like kind of pick the correct customer segment based on the company description

    David Kossnick: Mm-hmm.

    AJ Asver: provided?

    David Kossnick: That's perfect.

    AJ Asver: that work?

    David Kossnick: And then if you do at, and then the column name should pull it in for you.

    AJ Asver: Okay. And then let's see, what's this company

    David: I think it was info.

    David Kossnick: Yeah.

    AJ Asver: Yeah. Great. And then how does it know which one to pick? Oh, it just kind of knows from what's available, right? Is that how it

    David Kossnick: Yeah. The AI knows in a select list or look up the AI knows to use that set of options.

    AJ Asver: Ah, smart. All right. So let's, let's do it. Okay. Wow. so it's already started. Um, but what happens if we think one of them might be wrong? So, for example, Affirm. I don't know if maybe it needs a bit more data.

    AJ Asver: So I'm wondering if a good one could be that we look at funding.

    AJ Asver: And then last funding round will probably help. Um, and then maybe like employee count. I feel like this should all help work it out.

    David Kossnick: Yeah. Sounds great.

    AJ Asver: Okay. So then if we do that, um, then we need to go back to here and go to this segment thing. We close this. Okay. And then we go, okay. Provided. And then this is kind of my prompt engineering now I usually like to do it like this. So company, maybe we'll do company name cuz that might be helpful too, right?

    AJ Asver: Might know some of this stuff already. Company description, just info. Last funding round. Total funding. Okay. And now let's give this another try.

    David Kossnick: All righty.

    AJ Asver: Um, Fill.

    David Kossnick: Wow. Very different result.

    AJ Asver: Great. That's way better. Right . Now it like correctly categorized firm, which is awesome. um, and Agora and Airwallex Crypto for Anchorage, which is awesome. Apex, this is great. Now I have some qualification.

    AJ Asver: I just ran the qualifier and it's actually started qualifying these leads, which is really cool. And I guess in the case where it doesn't have information, it'll tell me, if the company's a traditional financial institution, might work that out. So I guess I could go through these and check these all later. But the cool thing is I now have some qualified beads that I can start, um, working out which ones to focus on for my sales.

    David Kossnick: That's awesome.

    AJ Asver: And reach out.

    David Kossnick: Nice.

    Coda's long term strategy with AI

    AJ Asver: I am really excited to start, um, using my new. AI powered lead qualifier cuz I'm a one man sales team and I'm already thinking of other ideas. Like earlier when you showed me generating emails, I think I'm gonna start trying that as well, like generating outbound emails or even introduction emails to, to the right customers, based on who the right investor connection is.

    AJ Asver: And that stuff always just takes a lot of time. And I'll say like when I'm in front of a customer and pitching, I'm like the most energized. And I think through the sales process when I'm like in the CRM and trying to write emails and then qualifying leads, I'm like the least energized. So having AI helped me there is gonna be really great to, to gimme a bit of a boost. Um, I'm curious, where do you guys see the long-term kind of impact of AI being for Coda and how you guys think about it for long-term strategy and also when will this be available for other people to try out?

    David Kossnick: We are launching, the beta very, very soon. The beta will be very similar, but there'll be a lot more templates and resources, um, and we'll be letting in way more folks and so would love feedback.

    David Kossnick: Please try it out. And in terms of where we're going with AI, I'm really excited about it. I mean, as I mentioned at the start, Coda has kind of a uniquely great place for AI because it brings so many different pieces of a team together in one place. And that context is so helpful for AI. And so things that get me really excited is being able to ask your workspace a question about how a project is going, and it's able to just answer because it knows all the tools you're connected to, all the notes people are taking about every project.

    David Kossnick: Um, you know, imagine you come back from being off for a week on vacation, you're like, what did I miss? And you get the perfect summary and you can drill into more detail and granularity on anything that you're curious about. Um, you know, imagine it felt more like having a teammate when you used AI.

    David Kossnick: Who was able to engage on, um, on your projects and give feedback and have details. And obviously it's not exactly a teammate and you're gonna have to, to teach it about each, each case. Um, but I think the, the vision of having less busy work and more impact is really exciting.

    AJ Asver: I, I think that potential of, AI in Coda that you mentioned beyond just a single doc, but when you're using AI to really run your company is gonna be a really, really, really powerful one. Because if you have kind of sold on using AI as your wiki and to run your projects and your docks and your spreadsheets, then you guys basically have all the information as you mentioned that's required in kind of an internal knowledge base to answer these complex questions.

    AJ Asver: Like what's the status of a project, who is working on what parts of the project? And so. personally, that's, uh, something I'm very excited before because we use Coda for everything and we're a very small team right now. But I imagine as we get bigger and more people get involved, being, being able to ask those questions and being able to answer 'em is really, really cool.

    AJ Asver: And so it's gonna be in beta soon, which is really, really awesome. And how are you, as a PM thinking about, you know, launching this Beta and what it means to kind of bring this into production? And do you have any tips for other PMs that are working on AI products? Because you are, you, you've been at this now for six months, right? Um, which puts you, I would say, in like the early percentage of PMs that are working with, you know, GPT. Um, curious what your tips are for, for other folks trying to bring a product from that hackathon to a beta and then a GA.

    David Kossnick: know, I'd say a lot of people start, like we did, of just sort of, you know, throwing AI behind a prompt, a prompt box in your product. I think it's a great starting point to learn and see what people are doing. And I think as you develop a sort of deeper sense of what use cases are really valuable, um, you're gonna build something much more specific for them.

    David Kossnick: One of the really fun things about working with AI is you don't know exactly how it's gonna go. You know, the models are a moving target in terms of what they're really good at, what they're really bad at, what the, what people expect of them as well. And so, uh, you know, a very common process I've seen a lot of teams do, us included is have some generic prompt box, some entry point into AI in their product and sort of see what people use and gravitate towards in their scenario.

    David Kossnick: I think that's a great starting point to learn. As you have deeper conviction, building purpose-built things for those use cases is really, really valuable cuz it's just a lot less effort at the end of the day. Writing prompts is really helpful for a really specific thing, but it's a lot of work for something you just want done quickly.

    David Kossnick: And so some of the stuff we've been working on at the last month or two is, exacto knives, really specific things to open up exactly what you want in one scenario. and people have been loving them, which is great.

    AJ Asver: I have one feature request, which is please automatically work out what type of column I need. Based on like the description of the column That seems like an easy one for AI to, to solve.

    David Kossnick: You know, that's an interesting one cuz we, we actually do do it today, but in a subtle way actually. And yeah, this is like really in the weeds, but the text column type in Coda is actually an inferred column type. If you put a number in there or any other kind of structure piece of data, it will actually be able to, you know, operate on that data as if it were that type.

    AJ Asver: Well, I am very, very much looking forward to seeing more AI in my Coda docs and also very excited to see where this all goes. think what you guys are doing with Coda and AI is really, really, really, really cool and also just very helpful. Like it saved me a lot of time and I think other people too. Um, once it's available in beta, I'm sure there'll be lots of new use cases that we haven't even seen yet.

    David Kossnick: I did have one last question for you as well, which is, uh, you know, since you're working on AI every day on agents in particular, I'm curious how you think about sort of the future of agents in productivity. Like, what do you imagine agents are gonna do in finance and in every vertical, uh, you know, in collaboration with people?

    AI agents

    AJ Asver: That is something I spend a lot of time thinking about, um, in between like building actual agents. And right now I would say that we are vastly underestimating what's possible based on what we've been able to achieve at departure with our agents and the fact that they can just follow a set of instructions from a Google doc and carry out, you know, a very complex compliance task, like,

    David Kossnick: It's not a Coda doc? AJ!

    AJ Asver: That's true. We should be using Coda docs. Yes. A Coda doc, sorry. being able to follow the instructions and just carry out a task with the tools given. That's like a very novel and pretty awesome, uh, thing to be able to do because now you can automate a lot of repetitive tasks that have to be done very manually today. And so I imagine we're gonna see a lot of. This idea of intelligent automation as we've been calling it, where you aren't just doing the kind of robotic process automation or like workflow automation that you did before where you're connecting things together with conditional rules. But you're actually now using essentially prompt engineering to automate a task fully with lots of different steps and lots of different tools being used just by providing the instructions.

    AJ Asver: In a way, you're, you are used to doing it already if you are solving this manually with the team, which is essentially a user manual, a set of, um, kind of procedures. And so you are gonna think about the many different use cases for that, not just in finance, but in processing forms in healthcare or insurance or in all kinds of other places where these manual, workflows are being done.

    AJ Asver: I think it's gonna free up people to have a lot more time to do more interesting creative, strategic work, than doing this more repetitive kind of, Tedious work that exists today. So we're very excited to see where that goes. And we're at the very early stages of it today, but, um, I I think it's gonna move very quickly.

    David Kossnick: Amazing.

    Wrapup

    AJ Asver: David, really appreciate it. you for taking the time to demo Coda AI, for talking to us a little bit about the kind of story behind how this feature came about and then for also helping me be more productive, with my Coda workflows as well. I'm really excited to see, product launch soon, beta and just a reminder for everyone where can they find, AI if they wanna sign up for it.

    David Kossnick: coda.io/ai.

    AJ Asver: Awesome. Well thank you very much David, and you everyone for listen in to this week's episode of the Hitchhikers Guide to ai and I will see you on the next one.



    This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit guidetoai.parcha.com