#4: Zapier

That's where this human in the loop comes in, right? So kind of in terms of actual legal work and reviewing the contract, you do want an attorney to come and review the red line that AI generated. The better, the quality of the red line AI is generating, the less work that the attorney has to do. And over time, the agent and the AI will learn to get better at that red lining. The more you use it and more you give a seat back, the better it'll get. [MUSIC] Hello, everybody, and welcome to another episode of the A to Z of AI. We are super delighted to welcome Suk to our podcast today. Suk is currently the general counsel at Zapier. And Suk, we just want to say thank you. We all know each other via the L suite. You've been a member for a number of years. You and I have had a chance to meet at a number of our events. And as you know, I'm a big fan of Zapier and really, really excited to get to chat with you today. So before we dive into how you're using AI, Suk, if you wouldn't mind, if you can just introduce yourself to everyone, tell us a little bit about Zapier for the one person who may not have heard of the company. And then final question is, what did your work look like three years ago before generative AI? Awesome. Well, it's great to be part of this podcast. Thank you for inviting me here. I am Suk Kim, General Counsel at Zapier. We are a AI orchestration automation company. Before AI or AI, we're thinking automation, right? So there is an AI component to automation, but before that, it's really kind of if and this, right? Automating business processes, automating repeated workflows. So we were doing a lot of automation within our legal team already because of the company that we are. But with AI, then it really did kind of get to another level of what we could automate and how we apply intelligence to automation. It's been really fun working within our group. It's kind of our legal team is six attorneys, including myself. And we actually kind of all the stuff that I'm going to show today, demo today, it's just been built by the myself or somebody on the legal team. And kind of their weave, you know, we're all attorneys with our different experiences. But in terms of kind of doing automation and using AI, all of us are just slightly different levels, but all of us are using it in our work. So I'm happy to share that today. That's awesome. I'm curious tonight, followed as you know, from afar, both as a Zapier fan and also in full transparency Zapier customers, I followed how Wade and other executives, including yourself have in many ways been such voices and proponents of this current AI age. Would love to hear from you around how Wade has maybe led that internally. And what's the culture at the company been around using generative AI, both within your own product, as well as just generally across what you do? Yeah, I think the one thing about Zapier to begin with and that starts from Wade, but the other two founders as well, Brian and Mike and rest of kind of all the folks working here at Zapier is that we are a very open culture and we're very much kind of fast innovative culture. When AI and kind of the technology came along, it was something that we very early on thought, gosh, this could be really transformative, not just for Zapier, but really for our customers and what they can do. And our customers have done amazing things with our platform, even before AI. And so for us, it's really a click like, oh my god, imagine what would open up for our customers. If we can make AI, it really easy for our customers to use, right? But still at the same time, really powerful and creating really a strong business impact for them. And so I think that's when we saw the moment and we said, hey, we really need to get on top of this and that's really for everybody on the team. And we're seeing that more and more now is that AI and it's kind of, its ability is really unlocks. What already to go back, Zapier to begin with was a no-code platform, right? And so our kind of cool was to make automation be available and be able to be used by everyone, not just coders, right? And so because that was already our company mission, AI fits so nicely into it because now we even kind of, you know, 10x that ability to everybody. So I think when we see how it's developing and how it's unlocking different capabilities of folks who are not coders, to be able to do really amazing things that we didn't know that was possible, right? And it's only, it's kind of unlocking more and more. We're still at the beginning stages of AI and so that's exciting to see kind of what capabilities or what people are able to do with it. >> Yeah, so that's awesome to hear. One thing I heard you say is you described the culture as Zapier as being a very open one. And I think you know, a recurring theme when we have these conversations is how much culture affects the way that a company is able to adopt and integrate these tools well. When I think about what kinds of companies tends to immediately thrive when they start adopting AI, its companies that already are fairly committed to not being siloed, to being internally transparent, to having data accessible across teams, real visibility. And I love that as a tool, AI really encourages teams to move more towards a transparent company that structures their data well in a way that team members can really access it because that's the environment that AI tools thrive in. >> Yeah, yeah, that's exactly right, that's exactly right. >> Well, I think we're both pretty excited to jump in and see you share some of the ways that you and your team are using these tools. So can we just see some of your examples? >> I'm going to start with actually this one. It's basically this AI regulations monitor agent. It's built using the Zapier agent product and it's an agent product where you can use natural language with the agent to prompt and have the agent build all the automations for it to work. So right now I have it triggering every week and the purpose is here's the problem I'm trying to solve with this. There's so much regulation happening, not just in the US at the federal level, state level, but across the globe. There's also cases that are ongoing and of course we all get email alerts from law firms but they're really hard to keep track of. And so basically this agent is, I've said, hey, here are 13 or 14 law firms that I work with or I've seen that their material is really solid. And so I kind of said, hey, on a weekly basis, go look at all the websites and summarize if there are any updates to regulations. So it does that on a kind of I think Sunday evening. It goes through that and it gives me on the Slack message a really kind of summary view. And then in the thread, because I don't like it, taking up all the Slack space, in thread, then it goes into a lot more detail. So for example, this one is from Taylor Westing and the other thing that's really nice about this agent is I've told it to actually, because my team, right, they're different folks on my team responsible for different subject areas. So the AI is looking at the topic and then choosing who on my legal team should be reviewing this. So this one they thought I should be reviewing it. So it assigns it to me and tags me. The next one is Ogletree one and Laura is an employment council on our team. So they thought Laura should take a look at it. Benjamin is our product council and they thought, okay, the AI thought, Benjamin should look at it. So you don't go through, right? Here's kind of a state level AI regulation. It also looked at corporate disclosure and security implications, right? And so it's tagging people, which is really nice. Then the other thing that I told it to do is because we all have a ton of things to do. So I told it basically, hey, when somebody is tagged, goal and schedule some time on their calendar so that they reviewed us and do that within kind of the next week, find the time available and this is really great because now I know people on my team are assigned when they need to and they're actually this one I only hooked it up the mine but in the one that we have live running, it'll hook into everybody's and it'll say like, hey, you need to go take a look at it. So this is kind of that one we have running and you could run this and you could easily make this for any regulations or any geography that you're looking into. And if you don't want to go into kind of law firm, you could actually hook it into actual kind of updates from regulatory agencies. You just need a lot more hookup for it, but there's so much that I love about this. I want to call out two things that I think are fantastic here. So the first is you and your team have really right sized the delegation that you're doing here to your very last point. You could have potentially gone through the work of trying to get it to actually summarize in detail all of the proposals and rules and regulations as they come out. But instead you said, you know what, the last mile is always really difficult. Let's make sure that that last mile is still handled by our team. Here's reliable sources go pull that. And so-- So you're really mitigating the opportunities for error and giving your team exactly what they need while cutting out all the busy work. And then the second thing is one of the last things you said to you, actually getting time on your team's calendar, you're using these tools as an effective executive assistant as well as everything else. So those two things, it's remarkable because there are things that these tools are very capable of, but they also require a lot of thoughtfulness to get them to do those particular things, right? The failure, you've dodged the failure modes in other words. And it's a nice combination of the automation that requires the deterministic flows with the intelligence layer. It's a nice one of that. Another one I'll show you here is our NDA, and I ran it this morning so that we wouldn't have to spend time here. It's basically we have an NDA workflow and we have this hooked up into our CLM, but for purposes of this demo, we just hooked it up directly to Slack. And so basically Kyle is a counsel on my team, so normally this wouldn't come from me, this would come from like a salesperson or somebody on our procurement team, right? Say, hey, Kyle, can you review this NDA? They're what it does. We have a bot, an NDA bot that has our playbook and is revealing that NDA and it's doing two things. It's going to red line it. Then it's also going to categorize it with just simple red, yellow, green so that it can message back now. It says, hi, super. It would be the salesperson or the procurement person, whoever requested it, says, hey, review this. The AI by Zapier has reviewed it and it's categorized it as red. And it's going to take and here are the reasons, but it's going to take a little bit of time for Kyle to review the red line it generated. So then it's letting the requester know it's not sitting in legal queue somewhere, right? We do know that the attorney has to come and look at it, but at least they know. And then so the first message will be this one. And then once then Kyle goes in, because on the backside, it's also the bot is also for pinging Kyle on a DM and says, hey, Kyle, because we have in our Zapier, human in the loop step. And that will pinging Kyle and you say, hey, Kyle, here's a step you need to come and look at. And Kyle goes and looks at it. He can review the red line and he says, okay, that's done. I reviewed it. And then I approved it and now I'll share it. Now if we're using a CLM, it'll directly go out to the counter party for purposes of this demo. We just kind of said, like, hey, here's the red line. And then whoever is requested, they can also see the red line, right? Because it connects up. What's really nice is that it includes a human in the loop step. It cuts a ton of time on our attorneys because it takes a lot of time to manually communicate. But we want to be good partners with our stakeholders and want to let folks know, but this lets them know right away. And which is kind of a nice process for us. That's awesome. So I'm curious for both of these. Did you on the Zapier end? Use the natural language box, so to speak, to start building these or did you actually go into the workflow and set them up? How did you get started here? We have within Zapier. You can choose. It's your adventure. Kyle built this one. And he actually really likes workflows. And so he kind of will build it out that way. If I were to build this, I would probably start with the Zapier Co-Pilot and ask where we also have other tools that help folks figure out what steps they need. Yeah. And I love that too, right? I've done both. I think with workflows, what's great is it's really good at setting things up for you. But I think sometimes in Zapier you hit on this earlier as well, right? Like sometimes we can stumble over this need of AI needing to be perfect. So I think as lawyers, maybe that's, I'm not one, but Zapier you are. I'm curious to hear how do you then balance that need for perfection when it comes to building workflows for your team? Here it's sharing information about new regulations. Maybe you risk level as low. Perhaps here, there could be a little bit more of a concern. So how do you and your team think about that? What's your approach to it? For us, it's kind of, I think what's really helpful is to understand where, kind of like, when I think about in different buckets, when I think about AI use cases, there's the one I showed here. There's workflows that are deterministic and you've got kind of rule based this, that, right? And that you don't want any kind of additional intelligence applied because you just want things to work the way they're supposed to, right? It's triggering actions across many different systems. You don't want that to fail. Where you have some AI layer, you want to make sure that there's some control and that's where kind of playbooks and everything come in the way. And that's where this human in the loop comes in, right? So kind of in terms of actual legal work and reviewing the contract, you do want an attorney to come and review the red line that the AI generated, the better the quality of the red line AI is generating, the less work that the attorney has to do. And over time, the agent and the AI will learn to get better at that redlining, the more you use it and more you give a seat back, the better it'll get. And you can also automate updating the playbook as you update kind of your responses to the red line. So it gets better. But in terms of the risk appetite, it's kind of really looking at things of like, where are you okay, just the AI responding? Where do you want the human in the loop step? Those are good checkpoints. Yeah, makes sense. I know you were going to share one more with us. Do you want to do that? Yeah, let's do that. This one, I just copied and paste it because it was kind of hard to share. We're a pretty transparent company. And so we have just a hashtag legal Slack channel that most across the company, people just ask questions unless it's personal or private or confidential or sensitive. We have an agent that's responding. And so here is an example of one where, hey, somebody's asking, this is somebody across Zapier, right? Some person. It's in a case, customers are asking for data subject requests. And this one, because the agent, it was pretty specific on the answers and it had the exact right information, it gave that and said, okay, if you're a status fighter, give a complete, then it's done. If you actually need something more, it actually tags into a gerotech, it'll open up a gerotech and assign an attorney if it requires more. Here's another one from a legal channel. This one, you can kind of see this one, the bot responded, but it didn't have all the information. So then it tagged in the appropriate person on our team to come and take a look at it. So then it tagged Kyle, then Kyle came in and said, oh, yeah, here's kind of the bot was pretty good, but here are the things where you're just additional information. So kind of what you're saying about risk level, you could set it like, for example, we have it set. So if the bot is clear, we have a Zapier table running and those are deterministic, pre-approved responses. If the agent is pulling a response from that table, you don't need anything more, you don't need to tag the person where if you're kind of pulling in question, but it's not definitive, then it's going to tag the appropriate person. If it's really complicated and the agent doesn't know, it won't even even post. It'll actually just do a human in the loop step and say, hey, Suke, here's my draft response. Can you look at it before I respond? You've said it a few times and I just want to reiterate because I think one of the things that our audience often worries about, of course, is things like hallucinations. I just want to underscore this marriage of the Generative AI tool and something deterministic, which is to say, as you said earlier, very if dead. If the question looks like this, then you are pulling from a table, pre-drafted responses, something that is not hallucinable and that combination is really what gives this the power and the reliability that your team has. That's really incredible. I think it's easy to forget that it is possible to marry these Generative tools with very concrete information. Yeah. I think that's the key is just kind of understanding what are you using the AI and automation for? Deterministic flows, deterministic with some intelligence applied and how much do you want human steps in there? The other thing that I didn't talk about is I think there's kind of this whole other use case where it's almost like people call it like the digital twin or your kind of your AI helper or whatever. And it's basically giving the AI the relevant knowledge sources so that you could use it as a thought partner. And that is almost kind of a personal use case. And so you don't need to be as worried about deterministic responses. You're actually feeding it as much context or in log or skills that you feed it so that it can really be a true thought partner and you can feed it kind of how you like to think about it. You can feed it your Slack messages or kind of your other documents, memos, you've written and things like that so that it has that context. And it can really kind of be a good thought partner. But I've actually just told my team was hey, you could actually build yourself a super and put in like all the feedback I've given you and kind of my thinking so that when you want to bring something like ask that bot and it kind of probably give you some things that I might give feedback and you could probably also So use it to build your case. Like, how if-- So it says, there's what are you going to say back? Like, you can prepare for your reading in that way. And so I think there's just a lot of fun additional ways to use it. That's less about having workflows and completely different use case. So I love that example. I actually did that with my supervisor. I had Gemini take a look at emails from a particular time period. And I said, hey, study this. And then I use it as a thought partner when I'm sending a proposal or a pitch. Let me think through how to make it make this case in the way he would think. And it's been incredibly valuable, right? I'm curious to ask you. So now post AGI in the era we find ourselves in. How has all of this changed the way you do your work? And what do you find yourself spending more time on given all that AGI is doing for us? I spend more time on systems thinking. And you still have your important legal work that you need to do, right? Those don't go away. But I spend more time systems thinking and saying, because you have really talented smart people on the legal team. And I have to think, OK, how can I automate or use AI for a lot of the stuff that they do on a more routine basis? So that actually they can go spend their experience and smarts on really tackling some really complex things. And even in doing that, I think it also helps them look at it and think, like, yeah, if I can shift-- I think the NDA. I think we did ROI calculation for Kyle. And in terms of the number of NDAs and the time he saved, because AIs doing the first review and sending all the messages, and it's kind of a good 20% of his work time for last year, he saved just from that one alone, just because we all get so many NDAs. And you still have to look at it, right? So it's just simple things like that. And then he actually is somebody who really loves to build workflows and automation. And so I said, hey, if you can save time in that way, guess what? You can use that 20% checking out different AI tools and seeing and seeing how different ways that we can do better in terms of using AI or other kind of tools for to improve our work. That's fantastic. I mean, once again, culture is so important, empowering your team to say, hey, if you end up spending way less time on NDAs, that's a success for you. And there are other things of the company that you can do to add enormous value in. You can experiment with these tools more. You can create new efficiencies. I'm being able to empower our teams and ourselves to do that without the fear that we're going to make ourselves obsolete and say goodbye to ourselves. That's so important. So I'm curious. You're talking about ways that you and your team have saved time. Are there things that you all have given your best shot to solve with AI? That it ultimately didn't work. And it ended up remaining kind of in the solidly human camp. I think one of the things that we're still, I'll say solidly in the human camp, I'd say we don't give up that easily. And so there's certainly things that we thought, how should we do this or not? And we said, hey, let's not do that. Only because like, we don't know how courts will respond. Or kind of like, hey, we do have to think about attorney climb privilege. Like there are things that we're like, yeah, let's not go down that road. And so there's definitely that. And one where we're trying right now is some of us are playing around with cursor to see like, hey, many in-house lawyer legal teams do this. Like we use, as you saw, we use JIRA for our ticketing. I'm like, oh, it'd be really nice to have like a nice looking UI on top of it. And so it still kind of remains all the ticketing of assigning on JIRA. But can you report on the SLA? Can you report on who's doing what, right? And kind of doing-- and JIRA does have that. Right? But can you do kind of a nice looking UI on top? And so that's kind of what we're working on right now. And these are lawyers and just playing. And again, that's kind of the thing. Like that's your side project. If it works great, if it doesn't, OK, it doesn't. But it's kind of-- look, if you add an hour or two and you want to work on something different, and that's what you can work on. Yeah, it's amazing. I was just going to say, I love that we all are in this era where we can do that now. Versus going to our engineering and product teams take and you build this, we can all just do this. And you're right, we don't have to get it perfect, right? But we can at least try and learn something through that. So thank you for sharing that. That's awesome. We always like to close this out. We call this the A to Z moment. And we're going to ask you a few questions. And then I know Zach will close this out finally here. But curious to hear from you, as you look ahead-- and I know even over the past few weeks, so much has happened. But what would you say excites you most about AI? And let's say, looking at the next six months. Yeah. I think AI, even more-- I kind of mentioned this before-- allows folks like us, for non-developers, to be able to do more. So it's almost kind of like your imagination is the limit. It's like the AI and the tools are becoming so accessible to non-coders like us that it's so important to get your hand on keys. It's so much thing to kind of hear about it, talk about it. But it's really different when you start playing around with it and not be intimidated to like go into cursor and try it. Go into some other coding product and try it. Or go into Zapier and start to build your first workflow. Or go into Zapier agent. That's really pretty easy. Entry point for folks who haven't done workflows and try to build kind of like the bot that I showed. I think what's interesting is that it's evolving so quickly. So we have no idea what's going to happen in six months. But I do think what's true is I think it's going to get more and more accessible to people. Well, that is a perfect segue to the question that we always end with, which is, if someone listening or watching, try just one thing that you mentioned today, or maybe something that you haven't, what should that one thing be? I would say the really nice starting point is the regular Tory bot, just because you could just open it and start telling the agent, hey, I want to build this. And then it'll start doing it and you just test it and get it to the way you want it. But it's really an easy experience. So this has been amazing. Thank you for taking the time. It's always a pleasure to see you in person, but getting the chance to talk to you about how you're using AI, how your team is using AI. I know this will hopefully inspire in all of us and all of our listeners some ideas of what we can add to our list of things to do. Excited to see all that's going to happen at Zapier and we'll probably need to bring you on in the coming months to chat further about where we're headed. So thank you so much for today. Yeah. Awesome. Well, thank you so much for having me. And I should do want to say, anybody's interested in they go try and they're kind of hitting up against what they're able to do. They can totally email me. They can email me and I'll pass it to whoever can help them. But we really love kind of getting legal teams using these tools. Well, it's Zapier or some other tool and using AI. So we're here to kind of bounce around ideas. And I get a lot of ideas from our customers and legal teams on our customers. And so I'm always happy to hear and they help. That's awesome. Thank you again. Really appreciate it. Thank you.