[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_262":263,"_264":265},"routes/_app.x.$handle.$name",{"_98":97,"_33":46,"_106":88,"_28":56,"_13":14,"_107":-5,"_11":108,"_109":110,"_111":112,"_113":114,"_115":116,"_20":-5,"_117":-5,"_118":-5,"_119":27,"_120":121,"_122":25,"_123":-5,"_124":-5,"_125":126},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_44":45,"_33":46,"_39":47,"_48":-5,"_49":27,"_50":51,"_52":51,"_53":54,"_55":56,"_18":57,"_58":59,"_60":25,"_61":25,"_62":-5,"_63":64,"_65":66,"_67":-5,"_68":-5,"_69":-5,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":81,"_82":83,"_84":85,"_86":27},"version",26,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":27,"_28":29,"_30":29},"file",{"_31":32,"_33":29,"_34":35,"_36":38,"_39":40,"_41":14,"_42":43,"_24":27},"previous",{"_31":32,"_33":29,"_34":35,"_36":37},"deleted",true,"deletedHere",false,"path","beehiiv.ts","latestPath","fileId","60415d48-79de-11f0-8034-0224a6c84d84","name","fileType","script","content","export async function createBeehiivDraft(type: string, title: string, content: string): Promise {\n const apiKey = Deno.env.get(\"BEEHIIV_API_KEY\");\n const publicationId = Deno.env.get(\"BEEHIIV_PUBLICATION_ID\");\n \n if (!apiKey || !publicationId) {\n console.log(\"Beehiiv API key or publication ID not configured\");\n return;\n }\n\n await fetch(`https://api.beehiiv.com/v2/publications/${publicationId}/posts`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${apiKey}`,\n },\n body: JSON.stringify({\n title,\n content_html: content,\n status: \"draft\"\n }),\n });\n}","export async function createBeehiivDraft(type: string, title: string, content: string): Promise<{ success: boolean; error?: string }> {\n const apiKey = Deno.env.get(\"BEEHIIV_API_KEY\");\n const publicationId = Deno.env.get(\"BEEHIIV_PUBLICATION_ID\");\n \n if (!apiKey || !publicationId) {\n console.log(\"Beehiiv API key or publication ID not configured\");\n return { success: false, error: \"API credentials not configured\" };\n }\n\n console.log(\"Creating beehiiv draft with title:\", title);\n \n const response = await fetch(`https://api.beehiiv.com/v2/publications/${publicationId}/posts`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${apiKey}`,\n },\n body: JSON.stringify({\n title,\n content_html: content,\n status: \"draft\"\n }),\n });\n\n console.log(\"Beehiiv API response status:\", response.status);\n \n if (!response.ok) {\n const errorText = await response.text();\n console.log(\"Beehiiv API error:\", errorText);\n return { success: false, error: `API error: ${response.status}` };\n }\n\n console.log(\"Beehiiv draft created successfully\");\n return { success: true };\n}","createdAt",["D",1755267043013],"branchVersion","branchId","29c6e9e6-7938-11f0-ab1e-0224a6c84d84","id","29c68fe6-7938-11f0-ab1e-0224a6c84d84","Transcript-Processor",["D",1755194257445],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","Transcript -> Blog Post / Newsletter","imageUrl","","c156eb0e-8b9d-4104-a055-502e9a0eb441","pinnedFileId","604cb698-79de-11f0-a2d3-0224a6c84d84","pinPreview","pinMigrated","screenshotVersion","likeCount",0,"orgId","2019e55b-c541-4c51-bef0-008eef10ef88","tursoName","tursoHostname","tursoMeasuredAt","pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_98":97,"_105":101},"org",{"_98":97,"_99":-5,"_100":101,"_18":57,"_102":103,"_104":27},"branch",{"_13":87,"_33":88,"_44":43,"_89":45,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":46,"_96":97},"isOwned",73,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","charmaine","handle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvdXBsb2FkZWQvaW1nXzJ4cFpFME5pSXNZamJNd296dlZSYVJ0OVF1RSJ9","tier","pro","canRestrictVals","profileImageUrl","branchName","adoptOrgId",{"_44":45,"_33":46,"_39":258,"_48":-5,"_49":27,"_50":51,"_52":51,"_53":54,"_55":56,"_18":57,"_58":59,"_60":25,"_61":25,"_62":-5,"_63":64,"_65":66,"_67":-5,"_68":-5,"_69":-5,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":259,"_82":260,"_84":261,"_86":27},"tree",[212,214,215,216,217,218,219,220,221],"files",[212,213,214,215,216,217,218,219,220,221],"counts",{"_201":64,"_202":203,"_204":64,"_205":206,"_207":208},"contents",[127,128,129,130,131,132,133,134,135],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles","grants","orgs","customDomains",{},{"_44":194,"_33":195,"_31":196,"_139":-5,"_42":43,"_41":197,"_141":27,"_39":198,"_13":64,"_34":199,"_36":-5,"_24":27,"_144":-5,"_28":200},{"_44":187,"_33":188,"_31":189,"_139":-5,"_42":43,"_41":190,"_141":27,"_39":191,"_13":64,"_34":20,"_36":192,"_24":27,"_144":-5,"_28":193},{"_44":183,"_33":29,"_31":32,"_139":-5,"_42":43,"_41":14,"_141":27,"_39":184,"_13":185,"_34":35,"_36":38,"_24":27,"_144":-5,"_28":186},{"_44":176,"_33":177,"_31":178,"_139":-5,"_42":43,"_41":179,"_141":25,"_39":180,"_13":157,"_34":35,"_36":181,"_24":27,"_144":-5,"_28":182},{"_44":169,"_33":170,"_31":171,"_139":-5,"_42":43,"_41":172,"_141":27,"_39":173,"_13":64,"_34":20,"_36":174,"_24":27,"_144":-5,"_28":175},{"_44":162,"_33":163,"_31":164,"_139":-5,"_42":43,"_41":165,"_141":27,"_39":166,"_13":157,"_34":20,"_36":167,"_24":27,"_144":-5,"_28":168},{"_44":153,"_33":154,"_31":59,"_139":-5,"_42":43,"_41":155,"_141":27,"_39":156,"_13":157,"_34":158,"_36":159,"_24":27,"_144":160,"_28":161},{"_44":146,"_33":147,"_31":148,"_139":-5,"_42":43,"_41":149,"_141":27,"_39":150,"_13":64,"_34":35,"_36":151,"_24":27,"_144":-5,"_28":152},{"_44":136,"_33":137,"_31":138,"_139":-5,"_42":43,"_41":140,"_141":27,"_39":142,"_13":64,"_34":20,"_36":143,"_24":27,"_144":-5,"_28":145},"1a599b54-79e0-11f0-9605-0224a6c84d84","transcript_d2ac22cf-8805-4aba-b9f4-1db5ba21ad07","1a599bf4-79e0-11f0-9605-0224a6c84d84","parentId",21,"latest",["D",1755266386998],"{\"id\":\"d2ac22cf-8805-4aba-b9f4-1db5ba21ad07\",\"name\":\"Episode [Editing].txt\",\"content\":\"Episode [Editing]\\n===\\n\\n[00:00:00] \\n\\nThis episode of Building with AI is brought to you by Autoblox, the company I founded to solve one of the biggest challenges in AI today. Making LLMs safe, reliable, compliant, especially in high stakes industries like healthcare. If you're a company building AI powered products, whether that's a virtual health assistant, AI driven documentation, or any LLM handling sensitive information, you know the risks.\\n\\nLLMs hallucinate, they leak sensitive data, and in regulated industries like healthcare, finance, and insurance, compliance isn't just important, it's a law. So what's the challenge? Most companies rely on manual spot checking and vague governance policies. That's not scalable. What you really need is a way to proactively catch AI failures before they happen.\\n\\nBefore an LLM suggests incorrect medical advice or violates HIPAA compliance. That's where Autoblox comes in. We're an automated QA and compliance platform for LLM products, purpose built for regulated industries. With Autoblox, your team can run structured AI evals, detect risks before deployment, and enforce real time guardrails in [00:01:00] production, all while maintaining auditability and compliance.\\n\\nWe're already helping companies like Hinge Health, Pelago Health, and others scale AI safely. If you're building AI for healthcare, finance, or any industry where safety and compliance matters, you need to be using Autoblox. Go to autoblox. ai to learn more or book a demo today. Again, that's autoblox. ai, the trust layer for AI and regulated industries.\\n\\nNow back to the episode.\\n\\nHaroon: Hey Brandon, welcome to the show.\\n\\nBrandon: Hey, good to be here. Har.\\n\\nHaroon: Do you mind starting off with a quick introduction?\\n\\nBrandon: I'm Brandon Ute. I am the Chief People Officer at Zapier, which is an AI orchestration platform. Makes it easy to connect the tools and LLMs that you already know and love, uh, to get a ton of work done using automation and ai. And of course, uh, here at Zapier leading our people team, I get to see firsthand.\\n\\nUh, variety of ways, uh, to do that for talent and HR use cases. Other than that, uh, I grew up, uh, near Detroit. I come from a family of auto workers, so I've [00:02:00] always been really interested in, you know, how you build things and, uh, you know, build entire economies and communities around just, you know, important work that does good in the world.\\n\\nHaroon: So you have a pattern of being in the center of automation, ~uh,~ both in the auto industry and now what's happening in across tech, ~uh,~ which, which is super cool. I did not know that.\\n\\nBrandon: That's a good way to think about it.\\n\\nHaroon: All right, so before we get tactical, Brandon, there's a lot I wanna cover with you. ~Uh,~ and ~you know, we, we,~ we'll cover this during the introduction to the episode, but, ~uh,~ upskilling 700 employees across ~dozens of con, uh,~ dozens of countries is no. Easy matter. And so, uh, before we get, ~uh,~ too in the weeds, can you gimme a two minute story of how Zapier's AI transformation began?\\n\\n~Uh,~ what was a moment that the company decided that this was an all company priority and it was gonna go all in on making sure that AI literacy was as high as possible?\\n\\nBrandon: Well, you know, the, the triggering event for Zapier was the release of chat GPT-3 0.5 in the fall of 2022. By the time we got into the spring of [00:03:00] 2023. ~So~ about two and a quarter years ago, we had been using, ~uh,~ 3.5 enough to know that it was gonna meaningfully influence how we achieve our mission, which is to make automation work for everyone.\\n\\nAnd one of the first use cases for, you know, an LLM, like 3.5 that we were seeing is that it could, use API documentation to help build APIs.\\n\\nthat's was really important for an organization like Zapier. You know, folks come to, ~you know,~ Zapier to connect all of their organization's applications and then automate work across 'EM, and APIs are the primary conduit for doing that.\\n\\nAnd here was a new way to like massively scale the rate of API building and maintaining too. And so that was really interesting for us. And so in March of 2023, Wade, our CEO wrote a memo to our team and he called it AI Code Red, the company's first ever code [00:04:00] red. It was, ~uh,~ literally and figuratively kind of alarming and it.\\n\\nWas meant to be that way. because what we were seeing in March of 2023, and now I think so many of us are seeing the same thing, was a wave, right? A huge wave of innovation, of opportunity. But the other thing, and I really admire how, uh, I think Transparent Wave was about this, you know, waves can do two things, right?\\n\\nYou can ride the wave to greater success as an organization. Or the wave can crush you, over time. And that's kind of exactly what we were seeing and thinking, ~uh,~ back in early 2023. And so the AI code Red memo was really a call in. It was an invitation in all in together as a company to start really building and puzzling ~with,~ with even more intention that we had, ~uh,~ been doing in the last, ~uh,~ the prior six months to really understand how to scale our mission and how to scale this company with ai.\\n\\nat the middle. Now, happy to talk more about this, [00:05:00] but I will, I will leave you with a teaser that, you know, we published this early enough in like the LLM, you know, wave. ~That~ that memo or that call to action was not universally popular at Zapier \\n\\nHaroon: Okay. ~Uh,~ a lot to unpack there. And the first thing that comes to mind is, ~uh,~ code red. I mean, it, it sounds alarming almost. And, ~uh, it's,~ it's jarring. And, ~uh,~ it seems like that was somewhat intentional. ~Uh,~ do you think that's an important part of getting in entire organization, especially one that's, ~uh,~ the size ~of,~ of Zapier, ~uh,~ which correct me if I'm wrong, 700 plus employees for,~ um. For, ~for getting an organization of that size to mobilize, ~uh, in,~ in unison is, is that a key part of just getting that message across? Is making the message strong?\\n\\nBrandon: Y short answer is yes, I think so. Especially when we're calling the team into, ~um, you know, ~an area we are, we have not been before, right? It's uncharted territory. And the other thing is, it's really important. So it's new, it's relatively ambiguous. It's a little intimidating and scary, and [00:06:00] it's really important for us to figure this out.\\n\\nSo you take all of those ingredients together and what I have learned, in every organization I have, I have led, um, is, ~uh,~ however clear ~we are,~ we think we are being as leaders. The two ticks clearer than that. ' cause to your point, ~you know, you know~ Zapier, it's 750 people ~in,~ in one sense, that's a big organization.\\n\\nIn another sense, it's not a big organization. Right? ~We,~ we have customers who have, ~you know,~ 8,000, 12,000, 20,000 employees. those are big organizations. ~Uh,~ but I would say to your point, like you start getting into even just a a hundred, 200, 300 people. And the amount of ~like,~ clarity and, ~uh,~ focus that leaders need to, to help provide to the team, ~um,~ because becomes really meaningful.\\n\\nAnd being again, almost, ~um,~ painfully clear and repetitive I think really helps. And I did say all the more so when the team is distributed, like so many of our organizations have.\\n\\nHaroon: And what was the internal reaction to this code? Red.\\n\\nBrandon: It was mixed at first. We had folks that had already been, ~you know,~ [00:07:00] using LLMs every day, ~who,~ who understood exactly what Wade and our leadership team were pointing towards and were encouraged. By the clarity and the focus. ~Uh,~ we had folks in the, in the middle, and then we had folks who, ~you know,~ were at least initially kind of like, I don't know, like this is, this is literally alarmist.\\n\\nLike, what do you, what's going on over here? Like, I don't, I don't yet see what you're pointing at. And I also, because this is so new for all of us, not just at Zapier, I'm not exactly sure what to do yet. How is this gonna influence my work? How's this gonna influence my job? We had some folks, and again, like Zapier's a pretty open culture, you know, ask like.\\n\\nWill I still have a job? And the list goes on and on. And so again, ~like I, you know, I think in a,~ I think I would grade an organization's initial call to action on AI in the, in the response of the team, not by whether everyone has consensus or everyone's immediately supportive, but rather like the real culture test is, ~are people,~ do people have enough like psychological safety,~ you know,~ to ask questions and [00:08:00] share concerns.\\n\\nWith a lot of these like transformations, it's the rate of learning that really matters, not whether, you know, you have all the initial answers or where the entire team is initially on board.\\n\\nHaroon: Yeah, and I imagine that's a really delicate. Balance and sharing this call to arms, ~uh,~ so to speak. And we've seen a few other leaders, ~uh,~ follow that playbook Recently, I think at Toby, at Shopify, ~uh, there's been, I,~ I think the CEO at Fiber had done something similar as well. Those are a couple of notable public examples, but, ~um, sh~ sharing the magnitude of the shift that's happening, ~um,~ and then balancing that with the psychological safety as you hint at, and what are some.\\n\\nTactics that you might employ in order to strike that balance in a way that is ultimately productive for the organization and not necessarily creating more anxiety and fear.\\n\\nBrandon: Are a couple things. One, ~uh.~ A lot of the rubber hits the road on these big calls to action in team specific settings. So ~what,~ what that code read ultimately meant for engineering [00:09:00] team practically, ~you know,~ looked a bit different than it did for our marketing team. And so, ~you know,~ like what, ~uh,~ Toby did at Shopify, what Wade did here at Zapier, that's really important as that kind.\\n\\nTip of the iceberg call to action or invitation in for the team, but it's just the tip of the iceberg, ~you know,~ specific tactics that you can scaffold off of ~that~ that really help do some team specific ask me anythings, Hey, ~you know, we heard,~ we heard this from Wade. We're all figuring out together what this means.\\n\\nAnd actually on our team, ~we're gonna have to, we get to, ~we get to hold the pen and right power of that story ourselves. And so that's a place where re-anchoring the team on, ~um,~ the usually pretty meaningful degree of agency. We have, you know, some, sometimes we can get really stressed out by the, all the additional ambiguity and responsibility that comes with being called into something new.\\n\\nBut in so many of our organizations, ~the,~ the other side of that,~ you know,~ new responsibility or ambiguity is. ~Um,~ opportunity and agency and in fact, organizations,~ you know,~ if you're Wade or you know, anyone on our leadership team, we're counting on our teams to like hold the [00:10:00] pen and write some of the story, ~uh, you know,~ as part of the broader company.\\n\\nSo, team specific, ask me anything that really helpful. The other thing, ~you know,~ there's some very simple company wide ingredients. We've seen be very helpful for making a kind of big, big, scary, ambitious call to action really practical, really quickly. And some examples of that would be as, and this first one's gonna be non-intuitive potentially, but the, ~uh,~ company's AI use guidelines. you look at that and it's like, really, that's like the big inspiration that helps a team like get its head in, in arms around, ~you know,~ a big AI call to action. In, in, in practice, both at Zapier and in so many of ~our,~ our customers organizations, the answer is yes. IU guidelines can be a really helpful ingredient because great, IU guidelines don't, ~um,~ they don't constrain people's thinking.\\n\\nThey provide clarity on like golden paths. What tools can we use at our organization at this time? What data, customer data, employee data can I [00:11:00] feed, ~you know,~ into these systems and why? ~Um,~ and that clarity, like with a lot of the research on this topic, you provide ~the,~ the boundaries of that sandbox to build within and it actually boosts the team's creativity.\\n\\nIt's actually the total absence ~of,~ of clarity or guidelines. That can constrain the team's creativity because we appreciate the power of AI and also the risks associated with it. And so in the absence of, ~you know, ~guardrails or guidance, ~um,~ teams will be pretty hesitant to start building. And we really wanted folks to get hands on as quickly as possible.\\n\\nSo AI use guidelines were great company specific, ~uh,~ action. And we did that within like a week and a half of that, ~uh,~ code group.\\n\\nHaroon: roughly if you had a breakdown when the initial code red was shared across your employees. What was the rough breakdown? If you had to put a number on it? ~Uh,~ how much of the response was fearful and anxious versus excited? Optimistic, and, ~uh,~ yeah, more on the, ~uh,~ more of a positive response in that sense.\\n\\nBrandon: You know, ~I have, I would be,~ I'd be hard pressed to break the [00:12:00] ratio down person by person, but for, for a reason that I think is really instructive, which is, I think for most of us, we were all feeling a mix of those sentiments. ~You know, I'll,~ I can speak for myself, for example, you know, March of 2023, when that code red comes out, I am, I'm excited, I see an opportunity.\\n\\n~Um,~ I'm also nervous. Because that same opportunity can also be a threat for so many of our organizations. Like we have to turn it into an opportunity. And so,~ uh, you know,~ it just introduces ~a,~ a whole other layer of uncertainty on the business. And then practically speaking, I, I knew that I would also have to help my team, ~you know,~ hold the pen and write our page and the story of Zapier's AI transformation.\\n\\nAnd I wasn't exactly sure what that was gonna involve. ~Uh, you know,~ I didn't know what I didn't know yet. And as a leader, ~you know,~ that can be. ~Uh,~ that can be a little stressful.\\n\\nHaroon: Yeah. So in individually everyone's feeling these mix of emotions, and as leaders it's important to try and get folks leaning more towards those positive emotions, those, ~uh, ~feelings of optimism ~and,~ and excitement. And you [00:13:00] described some of the tactics that you used there, and who were some of the leaders who were involved in this transformation process?\\n\\nSo you're the chief. People, officer. And, ~um,~ I'm curious to learn why it's important for the people function, ~uh,~ to be central in leading this transformation and what are, what some of the other functions are that, um, you would suggest should be involved in this process as well.\\n\\nBrandon: I, I draw a lot of inspiration from what we've learned at Zapier, honestly over the last two and a half years, and I really admire how Wade our CEO approached this, which, you know, said, Hey, this is. Uh, this is existential for our organization and ideally it's gonna be a source of rocket fuel for, for the organization, for the customer, and that's great.\\n\\n~Um,~ at the end of the day, if you asked who's accountable for Zapier's AI transformation, Wade would say it's himself, right? Just like it is, like any like massive thing that we're trying to do as an organization. Importantly, Wade holds each of his execs. [00:14:00] Accountable for answering to their recipe for their organization.\\n\\nSo as it relates to like how we build, maintain, and improve our products, ~you know,~ Wade is looking at uh, Q, our CPTO. He's looking at Moaba, our head of engineering. He's looking at geo our head of product and saying, Hey, tell me the recipe. And by the way, if there are pieces of infrastructure that are missing or if you just have questions, ~like,~ let's talk about that.\\n\\nBut at the end of the day, ~like, you know.~ Just like anything else related to the company's product development engine, he's looking to the same three leaders, ~uh,~ for a point of view on how AI can influence how we do this. Same thing goes for everything related to talent and hr. He's looking to me to answer to how AI can meaningfully improve how we do that work.\\n\\nHe's looking to Lauren, our customer support exec for everything related to support and so on now. Within all of that, and to your, to your point, ~um,~ the people team plays a really important role in creating a lot of the enabling ingredients for individual teams to make good progress and produce results with ai.\\n\\nJust like [00:15:00] with any other organizational transformation. And that includes, for example, making sure that we're helping teams form a point of view on the extent to which they have like the skills and the capabilities that are needed to,~ you know,~ kind of cook their recipe. \\n\\nHaroon: ~I, I, I like this,~ I like this language ~of,~ of recipe and having the leader share their recipe for their respective functions. ~Um, and, and,~ and being held responsible for those. Now, if we're thinking about that recipe and the. ~Uh,~ byproduct of that recipe as being AI transformation within that specific function.\\n\\nWhat are some of the key ingredients that you're taking a look at, or what are some of the key ingredients that Wade is taking a look at in order to assess the, I guess, feasibility and rigor ~for that specific set of, of, of, uh, ~for that specific recipe, for that specific function?\\n\\nBrandon: you bet. You know, at the end of the day, AI is a tool. It is not an outcome. And I think it's really helpful when we're thinking about AI adoption, ~uh,~ to reground in that, ~you know,~ all of organizations have missions products and [00:16:00] customers. We have goals, things we're trying to do in the world. ~Um. ~AI might influence the strategy around that, especially if your company has like an AI play that faces the product and the customer.\\n\\nSo, you know, absolutely there. When it comes to the company's internal AI transformation, though, like AI is a, it's a medium or a tool to like meaningfully improve our ability to do the things that we already say matter most in the organization. And so ~the,~ the first ingredient. Is clarity on what excellence looks like and how you measure it.\\n\\nWhat are our goals? And you know what's so interesting about this arena? So many company transformations, like ~the, uh,~ there are certain cultural or operational components that are enabling forces. So to your point about ingredients, so you know, organizations that have a lot of clarity around what great looks like and are able to measure that in a way that people can see, ~like,~ my performance against this particular goal, or KPI.\\n\\nThat's hugely helpful. And so it can be ~a,~ a tailwind or a helper. The opposite is also true. Organizations are pieces of organizations that lack clarity on what excellence look like or how to go to it or what [00:17:00] have you. You think about like, what, what do we do with ai? It's like, I don't know, like what, what matters most and how do you know how you're doing?\\n\\nSo again, it sounds fundamental and I think it is in the literal sense, ~you know,~ as an ingredient for AI transformation. We talked about psychological safety earlier as a second one. It's incredibly important. It was incredibly important for other transformations that came before, before it. And the same thing is true with ai.\\n\\nNow, folks need to be operating in an environment where they can share a wild unhinged idea, right? Ask a hard question, say, I don't know how to do this yet, ~can I?~ Can you help And then last but not least, a third ingredient I think is trust in leadership. Folks need to be able to trust what their leaders are saying.\\n\\nBoth where leaders have some confidence or a vision for where the organization needs to go. ~Um,~ and also need to trust that, ~you know,~ management will be open in saying that we don't yet know the answer to something, ~you know, ~where that's the case. ~Uh,~ but also be really clear about, ~uh,~ whether and how we're gonna go get that answer.\\n\\nHaroon: Love [00:18:00] that. ~I, I, I, I, that's, ~that's a great breakdown. And ~the, the part that the,~ the part that I wanna double click into here is measuring excellence. ~Uh,~ so you mentioned that as being a key ingredient makes a ton of sense. And a lot of company leaders who I'm sure listening to this episode, they're gonna be wondering how do I establish a baseline of how literate my organization or function is in terms of their ability ~to,~ to understand the different use cases and workflows that ~auto~ AI can be used to augment ~or, or,~ or improve the operating leverage of and establish that baseline so that they can incrementally improve and have a. Objective metric ~to~ to measure as they invest in some of these initiatives. How did Zapier start measuring the excellence early on and how has that evolved over time? I.\\n\\nBrandon: You know, back in spring of 2023, we started measuring one thing in particular, very intentionally, and that was, ~uh,~ both self-reported use of AI to, ~uh,~ improve our daily. Work and performance. [00:19:00] So we would ask,~ we started,~ we added a question to our engagement surveys on exactly that topic. Effectively, ~like, you know, like,~ I use AI to basically supercharge my work.\\n\\nAnd it was on a scale of one to five, where one is, ~uh, I,~ I def that definitely is not true for me today. And a five was, that is absolutely true for me today. And ~you know, ~just based on Zapier's mission and what have you, by the time we first measured this in March of 2023, this was like two weeks after.\\n\\nThat AI code read, 75% of the organization already said that the answer to that question for them was yes. That they were using ai, ~you know, in their,~ in their core work. However, when we measured it again, ~uh,~ a few months later, it went down, I think 11 points, I think from 75 to 64. And you might wonder ~like,~ well what happened there?\\n\\nLike, you know, you in that time you developed your I use guidelines and, you know, provided, you know, we got chat GT Enterprise and, and you name it. And here here's my, as I and I talked to a bunch of folks about this, and here's what I came to understand. The thing that happened in the intervening 90 days is that we raised the bar for what?\\n\\n[00:20:00] AI like fluency or AI adoption really looked like we just got a lot clearer as an organization that, whereas maybe in March of 2023, ~you know,~ just like dabbling with prompting and, ~you know,~ using it to summarize a basic, ~you know,~ thing here or there ~was~ was enough by the time we measured again later that summer.\\n\\n~Uh,~ I think we got pretty clear that we're talking about like meaningful improvements in everyday performance using ai. And so we saw that dip down to, I think around 64 ish and then climbed steadily over the last two years, whereas, ~uh,~ in April of this year, 89% of folks, ~you know,~ with this higher bar for what adoption really looks like or saying, you know what, this, I am, I'm genuinely using AI to, to boost performance in my core work.\\n\\nSo that's one. We also, on the backend, ~you know,~ like a lot of organizations are paying attention to, like actual, like product, ~uh,~ engagement statistics. So, ~you know,~ daily active users or weekly active users with, ~you know, ~chat GPT or some of our other core AI tools would be examples of that. So I'll pause there.\\n\\nThat's, ~you know,~ I'd say on the adoption side of the equation, I can definitely talk [00:21:00] about how we're thinking about, ~uh,~ like how to measure, ~you know, ~impact now and in the future.\\n\\nHaroon: Please. Well, I think maybe a good framing for this is the company-wide AI fluency framework that, ~uh,~ is being used at Zapier. I know that, ~uh,~ Lenny Roski in his newsletter had covered this, and I've seen it a few times on LinkedIn over the past couple of weeks as well. I would love to learn more about that fluency framework and dig into that some more.\\n\\nBrandon: Yes, that'd be a great way to bring us up to the present. 'cause that's relatively new. That's about a hundred days old. So, ~you know,~ we had been doing all of this upskilling and practice sharing and use case development with AI inside the organization for, again, ~you know,~ the better part of two and a half years now, but about.\\n\\n90 days ago we started thinking about hiring and ~you know,~ there are definitely like very explicit AI type roles that we were hiring in the organization for a couple years now. And those had AI relevant interviews and skills tests and that's excellent. But based on what we were calling ourselves [00:22:00] into, it's like, hey, at this point.\\n\\nAI fluency is effectively ~a,~ a baseline expectation for every role in the company. So how can we make sure that we're assessing for that for a hundred percent of folks that we hire ~into the company?~ And that's where the inspiration for the AI fluency framework came together. And so I can describe it in a minute, but we, ~uh,~ use this framework to assess a hundred percent of applicants for a hundred percent of jobs at pu.\\n\\nAnd the framework has four tiers. I'll really focus on the, on three of the four. 'cause ~the,~ the kind of quote unquote first tier is literally called unacceptable. Like it is the unacceptable tier because ~we,~ we can't hire folks into the organization who, ~um,~ are operating at that level. Now, the good news is ~that~ that unacceptable tier is largely reserved for folks who just are ideologically or practically opposed to using AI in their daily work.\\n\\nMaybe they have climate concerns, maybe they have ethics concerns that they haven't, ~you know,~ kind of reconciled or sorted through. And by the way, like I have a lot of empathy for all of that. [00:23:00] Like, you know, it's ~the Zapier's~ Zapier's ability to achieve its mission and serve its customers. ~Um, the,~ the path to that looking forward like absolutely involves, ~you know.~\\n\\nThe use of ai and again, thoughtful human-centered use of ai. Absolutely. And, ~you know,~ we don't avoid all of those questions in terms of how we use it and what it's for ~and,~ and what have you. I think it's ~really,~ really important. But we're looking when we hire, ~uh,~ to bring folks onto the team who want to puzzle through those questions with us together.\\n\\nAnd so, ~uh,~ again, but I think being really clear about. What our organization is trying to do in the world and why we think that matters that we are successful is really helpful, and I think candidates deserve that type of transparency. I think it's at least as important to be clear about who your organization is, not for as it is to do like kind of the splashy typical marketing around ~who,~ who you think you're for or why it's great to work here.\\n\\nSo that's the unacceptable tier. The other three in order are called capable. [00:24:00] Adoptive and transformative capable would be for, ~uh,~ for folks that are starting to use ai, they're building some degree of how to fluency, but they haven't necessarily translated the how to into actual impact in their daily work.\\n\\nLike they, they probably don't have an example yet of how AI is influencing their core kind of productivity or performance ~metrics. Uh,~ and adoptive is, is actually exactly a step in that direction. These are folks who do have tangible examples of using AI to improve their productivity and performance in the work as we typically understand it, transformative, we encounter very few people who operate at this level today, but I think that's what's so helpful.\\n\\n~I think a free,~ I think a good framework like this or a rubric related to talent, like the top tier or a level should be something that's really stretchy. Because ~you want to,~ you want to create a scale where like ~those, those like~ that true excellence can pop in a way that ~you,~ you know it exactly when you see it.\\n\\nAnd so we reserve transformational for folks who, ~um.~ Have examples of reimagining the core work itself with [00:25:00] AI at the core, really from the ground up. And again, very few people have examples of that because very few organizations have examples of that yet. But that's what we're applying for and we want to be able to ~kind of~ see that excellence in ~that, um,~ those skills when we see that.\\n\\nHaroon: How do you describe the transformative candidate? ~So what does the transformative candidate look like here? And~ it totally makes sense that the bar here is high because you also wanna ~set up a, a high bar to just~ set a standard for folks that is not necessarily commonplace. ~Um,~ but can you maybe describe what a transformative candidate that comes through your pipeline might look like and the behaviors and characteristics of them.\\n\\nBrandon: Yes. ~Yeah. It, it's a, ~it's a pretty comprehensive, ~uh,~ package of skills and ~it's not, uh, not, ~and they're not all obvious. So let me start with the obvious. The obvious would be folks who ~ha are,~ are like fairly skillful builders with ai. They can take. A vision for a meaningfully different way, a more effective way of getting a certain job done, whether it's product development or sales development or recruiting, whatever the case may be, and [00:26:00] build those new workflows, build those new ways of working with AI at the core.\\n\\nAnd again, that's really important and I think it's the most obvious skillset that comes to mind when folks are thinking about transformative level impact. That's only one half of the coin though. 'cause the other half of the coin. And we see this as ~like, you know,~ really an under understated, but equally important ingredient for folks that really change the game with AI is what I would call, ~uh,~ craft or taste.\\n\\nAnd what I mean by that are folks who deeply understand. What greatness looks like in that area. Again, think about product development, sales development, recruiting, and so on, who really understand the end user. 'cause they've spent a lot of time with them. ~They understand. They understand.~ They have enough experience to understand, like the trade-off, at least the traditional trade-offs, right?\\n\\nWhether that's speed versus cost versus quality, or there's the classic, ~like, you know, you can have,~ you can have two, but not three. Pick two, right? It's like, well, with ai, I'm not convinced that we still have to just pick two. Maybe we can have all three. So ~they,~ they have this experience with the end user, some of these core [00:27:00] frameworks for understanding what excellence could look like.\\n\\nThey understand the traditional barriers, ~you know, to,~ to excellence and they're able to use that to reimagine the core work. ~You know,~ for some folks who think about, well, does that knowledge like constrain you? 'cause they're like trapped in the old ways of thinking. And I'm like, well, no, not enough for someone that's operating at this level.\\n\\nThey are aware of traditional ways of working. They are not constrained by them. So you think about AI builder skills as one set of capabilities. You can think about the second ~as~ as craft and taste like really understanding what excellence looks like. And there's a third piece, which is leadership. It's can this individual migrate not just systems and processes, but people right from older ways of working to newer ways of working.\\n\\nSo those are really the three core capabilities that I would think about. Just to sum it up again. AI builder skills kind of subject matter expertise or craft. And then the third is just change leadership.\\n\\nHaroon: so one of the conceptions that I imagine a [00:28:00] lot of the folks in the audience and might be listening to this. Might have in mind is that somebody needs to be quite technical in order to be a transformative, ~uh,~ candidate. Is that true? Do they need to be, do they need to be competent with building AI workflows in code or is that not necessarily a prerequisite here?\\n\\nAnd are there ways for folks who are quote unquote non-technical to be the transformative candidates? And if so. What does that look like?\\n\\nBrandon: you're hitting on what I think is like one of the most like, exciting things, ~uh,~ about ai because my answer to your question, does someone need to be technical to be transformative is no, at least in the way we traditionally think about like a technical individual in software. I think you named one of those traditional characteristics, like a technical person is someone who, ~you know, can,~ can code, can ~like~ write software.\\n\\nLike that's what it means to be technical in software. [00:29:00] You know, the truth of matter is, uh, well, a couple things. One, uh, you know, I, I deeply respect the craft of software development and I believe, you know what I hear from some of our engineers that, ~you know, you can,~ you can vibe code your way. To like maybe 80% of a fully baked software solution.\\n\\nBut that other 20% is the difference between something that like makes for a good demo and something that is like sustainable, scalable, and really impactful from the customer's perspective. So call that maybe like the last mile in software development. And I can still see at least today that, ~um.~ You know, software developers and PMs and like, ~you know, that~ that domain expertise around software development is still incredibly important.\\n\\nSo, ~you know, we,~ we keep humans very much in the loop in the work that we do at Zapier. However, ~uh,~ there have never been more tools in ways of working where folks who do not know how to code can build solutions using their subject matter expertise, their craft, their understanding of the customer. ~Uh,~ to really move the ball forward on performance and productivity.\\n\\nAnd I can, ~you know,~ think of no better [00:30:00] example from Zapier's recent history than our own customer support team. ~You know, um,~ we do a quarterly hackathon and the first hackathon we did after that code red in 2023. Was, was, ~you know,~ no surprise, AI focused and every team in the company participated in this hack week, including our customer support team.\\n\\nAnd even back then, you think about how far the models have come right in over two years. But over, ~uh,~ over two years ago, four folks from our customer support team who would not be considered technical in the conventional sense. They're not coders or what have you. They used, ~uh,~ chat GPT to code a applet.\\n\\nFor Zendesk, which is our customer support tool. And why? Well, because they really understand the customer. They really understand the craft and they know enough about those two things to know that one of the biggest prizes in customer support is making it easier and faster for a customer support Replit to ~uh,~ digest a ticket or a set of tickets and really understand the crux of the customer's problem so we can.[00:31:00] \\n\\nResolve it quickly and accurately. And in fact, that's the mission of the support team at Zapier. It's fast, accurate support period. ~You know,~ so you have Lauren Franklin, our customer support leader, who's done a really nice job of creating that clarity with the team. What is our mission? What does excellence look like?\\n\\nAnd the great thing is she has two metrics, right? So the mission of the team is fast, accurate, support. There's a metric for fast that's called average ticket handle time. And there's a metric for accurate, which is related to csap, like from the customer's perspective, did we, how well did we solve the problem?\\n\\nAnd so it is just like really focused, it's really compact. And so with that in mind, without Lauren having to tell the team what to experiment with during this hack week. They code, ~uh,~ application for Zendesk at the time was called, ~uh,~ Zen GPT. It took a bunch of Zapier specific product and customer, ~uh,~ data as the knowledge base and uses it to summarize the ticket so that we can solve the customer's problem.\\n\\nMore quickly and more accurately. ~Um,~ today that, ~uh,~ hack week, ~you know,~ [00:32:00] project is now fully deployed across the whole customer support team. It's called Support Sidekick now, and it's one of the ways that the support team at Zapier has been able to reduce that average ticket handle time by almost half. In doing so, ~uh,~ maintain or even increase customer satisfaction with our response to \\n\\nHaroon: I wanna dig into drive up this AI literacy, AI fluency to that 89% mark that you had shared. What role do hackathons and hack weeks play, and what are some of the other tactics that Zapier has deployed in order to drive that metric up?\\n\\nBrandon: The hack weeks are an interesting place to start. 'cause part of our theory of the case on how to scale AI adoption in the organization was to integrate some of the learning and the building and the sharing ~of,~ of what we're building into as many of the organizations existing practices as possible.\\n\\nI mean, ~with,~ with [00:33:00] ai, there is plenty that is new to like take it. And so like instead of lighting up in addition to just the, the, the newness of ~the, of the,~ of AI itself, ~um,~ in creating a bunch of additional practices in the organization to support it, we took more of a integrated approach. So the hack weeks are a great example where like AI's at the center, like really,~ you know,~ motivating folks to, ~uh,~ solve a problem.\\n\\nWith ai. ~Um,~ so that's one. The others, ~uh,~ when it comes to ~like, what we,~ what we recognize or, or what we think good, good, ~uh,~ performance looks like. You know, like most organizations, we have a, ~uh,~ like a core set of expectations, and we call them impact behaviors. And, ~uh,~ we've incorporated like the kind of building and experimenting with AI into those core expectations.\\n\\nNow, what matters the most? When it comes to performance, ~it's,~ it's productivity, it's output, right? For support, it's fast, accurate support, average ticket handle time, customer satisfaction for recruiting. Time to fill. Candidate experience offer to close rates.~ Like, you know, ~[00:34:00] in, in big picture, you know, for our recruiting leadership, it's,~ you know,~ how faithfully are we delivering with the organization on the hiring plan.\\n\\nSo those are some examples. ~You know,~ the AI fluency framework that we talked about earlier, that's an integrated approach, right? It's a new framework, but it's for the purposes ~of,~ of hiring. Now two other things that really helped, ~which I, I think are like,~ just outside of the integrated approach. One is, uh, just setting up a really clear place where folks can ask questions.\\n\\nIt sounds so simple, right? Like, well, what do you mean? Like, just ask a question. It's like, well, well, how Right? These questions are big. They can be intimidating. And, but, and importantly, not only do you not wanna, not only do you want to feel, ~um. It's~ kind of confident in asking a question, but you also want to get a, actually, back to the customer support example, A fast, accurate answer, and that doesn't happen in most organizations by accident and it wasn't gonna happen by accident, is Zapier either.\\n\\nThe good news is the solve for that is pretty straightforward, and here's what we did. For what it's worth, we created a Slack channel [00:35:00] and ~that~ that was the place I was like, this is the place any and every AI related question can go. You're not sure which model to use, ask it here. You're not sure whether this data can be used as a source, use it here.\\n\\n~Um,~ you're trying to figure out how to interpret this part of our AI use guidelines. Same thing. Single channel. So that was one piece. The other side of the coin is that we asked a few of our most skillful AI kind of experts, ~uh,~ to basically get in that channel and be accountable for helping folks get answers to their questions. And so you even think about like, gosh, what would it look like if we just started, if we actually changed the job description of some of our most skillful, ~you know,~ AI folks so that, ~you know,~ call it 20% of their job is helping others in their journey, right? Their AI fluency journey. And just like anything else, like what gets measured gets done, and ~I would,~ I would be willing to bet that any organization that is really clear.\\n\\nAbout repurposing some of their, ~you know,~ experts, time in service of helping everyone else [00:36:00] come along in the journey. And is really clear about some of those like golden paths, ~you know,~ single channel for q and a, ~you know,~ single form for demos where people can get inspiration on use cases, that type of thing.\\n\\n~Um,~ that \\n\\ncan be rocket fuel. \\n\\nHaroon: ~Amazing.~ And ~if,~ if we were to have a histogram of the categories of questions that folks have asked since Code Red, what would you say are the top few categories of questions that, ~uh,~ folks are asking? ~What are,~ in other words, what are some of the biggest challenges and. Question sets that are top of mind for folks at Zapier.\\n\\nBrandon: Mm-hmm. You know, in the, in the first year, so call it like spring of 23 to spring of 24. There's a lot of questions like, let's say like fundamental, like prompting. Just like happy, happy paths for getting certain types of results. Which model is best for what, and a lot of like, am I allowed to? and part of that you solve for, ~uh, now we have some, you know,~ our onboarding experience includes a bunch of that information upfront such that everyone who joins the company from that point forward has clarity on [00:37:00] some of those fundamentals.\\n\\nSo we're using, ~you know, to your point, by~ like histogram, like most common. Kind of themes or topics from the channel to inform what we should prioritize when we're onboarding folks into the company. ~Um,~ so that they just ~kind of~ have that context or that understanding, ~uh,~ from the get go outside of some of those effective, you know, kind of fundamental how-tos, ~um.~\\n\\nWe'll get questions ~about~ like, has someone solved for something before? So this is like, Hey, I'm about to, I'm about to go like, try to figure this out and before I do, like, just in case, you know, I missed something like, has anyone else solved this or solved an adjacent problem? So maybe, maybe it's two different things, but the underlying, ~um,~ kind of thing you would build in order to solve it is roughly the same. So we hear that too. And by the way, I love to see that because~ like. You know, we,~ building with AI is, is a means. It's not an end. And so I like seeing folks like trying to figure out if someone has, ~uh,~ solved for an adjacent problem where they can ~like~ pick up and maybe get 80% of the way they're using someone else's \\n\\nHaroon: That's one of the questions that comes top of mind for me when I'm thinking about managing AI [00:38:00] transformation at a, ~uh,~ larger company is, let's say you empower everybody to be able to build AI workflows using the AI tools that ~they're able to, they~ they have access to. How do you prevent people from not necessarily stepping on each other's toes, but creating those redundancies and maybe.\\n\\n~You know, in in,~ in the engineering world you have this idea of spaghetti code, but almost this like, ~um,~ spaghetti workflows situation where people may be creating similar workflows using completely different tools and there's nuances to each of those. ~Is that,~ is that an issue at all? ~Like, is that actually an, an issue?~\\n\\nAnd if it is, I'm curious how Zapier is cleaning that up and creating a. More efficient system for managing all of the different things that folks are building across teams. \\n\\nBrandon: Yeah. ~Uh,~ this is an active, ~uh,~ question for Zapier. So ~the,~ the assurance is we haven't solved for that yet, but it is [00:39:00] partly by design. ~You know,~ I think ~when,~ when leaders are calling their teams into their version of an AI transformation or any transformation, ~you know,~ we have to decide like how much. ~Um,~ like tops down control.\\n\\nWe want to exercise over the effort. And when it comes to Zapier's AI journey, there were a couple of things where ~we,~ we did feel like we needed to provide that, ~you know,~ like tops down, ~like~ architecture. So the AI use guidelines would be a great example of that. ~Um,~ the AI fluency framework would be a good example of that.\\n\\n~Um,~ the integration of the relevant AI knowledge or skill building into, ~you know,~ our, ~uh,~ new hire experience, our onboarding. Into our impact behaviors, right? Our core expectations of ourselves. 100%. Super helpful. And I think that's a place where like uniform or tops down clarity actually helps with speed and impact.\\n\\nWhen it comes to things like use cases and who's building what and why, what have you, gosh, we do not have a great solve for that yet. And ~you know, ~in some ways ~it's like, it's a, uh,~ it's a problem of our own making. Because folks have really picked up these ways of working and are experimenting, like we're doing so many of the things that we [00:40:00] aspired to when we wrote that code red over two years ago.\\n\\n~Uh,~ so ~I would call it a,~ it's a good news problem, but it's still a problem. We're interested in, ~uh,~ figuring out. ~We,~ we have very intentionally, ~uh,~ accepted degrees of, ~uh,~ redundancy and, ~uh.~ Chaos, for lack of a better term, right? In service of speed. ~Uh,~ it really just like boosting our rate of learning effectively.\\n\\n~So, you know,~ the things that are, I think are going particularly well today are folks sharing their use cases, asking for help, like the, ~you know,~ degree of psychological safety that you can observe is really helpful. So I like to see that. But if you were to ask me today, ~you know,~ Brandon, where is your library?\\n\\nOf like most everything anyone has built, ~you know,~ with AI in your organization. And, and do you have like a taxonomy around it where you can search? By the way, the reason I'm getting specific is 'cause like, this is what we're working on right now. Uh, you know, where you can search by like, problem to be solved or like business metric, you're trying to influence, you know, stage.\\n\\nLike is it, is it, [00:41:00] you know, prototype? Is it,~ you know,~ in production, is it scaled and so on. You know, where can you read all the documentation? That kind of thing. We don't have an answer to that yet, but it's a question that we're really interested in solving, ~uh, over ~\\n\\n~the~\\n\\n~next few~\\n\\nHaroon: Awesome. Watch this space. ~Yeah,~ it's, ~uh,~ excited to see what comes out of, uh, Zapier in the next few months ~on, on,~ on that thread. And, ~you know,~ sort of along those lines, I'm curious about. The ROI of, ~uh,~ AI workflows and the importance that you place on making sure that the workflows that folks are building are actually moving the needle.\\n\\nAnd ~I know this might not be necessary. I'm,~ I'm sure it's not a secondary goal, but there is the benefit of. Improving AI literacy from building any AI workflow, whether it's improving ROI or not, you're getting more familiar with the tools that you're using. You're getting more used to getting hands on, but at a certain point, it's also easy to spend a lot of time on architecting AI tools that may not necessarily move the needle, and [00:42:00] maybe using AI in situations where. There's a high level of human of the loop required. ~Um,~ and you may end up in a place where the AI workflow and maintaining it is, ~uh,~ such a high level of effort it may not necessarily be justified in continuing to spend time on that because it's quite a manual process. Is that also a challenge that you've experienced during the AI transformation process?\\n\\n~Maybe it, maybe it's not, uh, but I'm curious, is that something that it has been a challenge,~ and if so, how have you managed that challenge?\\n\\nBrandon: You know, I think you're right to talk about this notion of ROI, because every, everything I'm about to say is still anchored on that like fundamental concept. When we think about ROI, there, there are a couple things that have been helpful. One is, ~are we,~ are we measuring ROI on a use case by use case approach or, ~uh,~ using a more of a portfolio approach?\\n\\nFor example, if I tell my recruiting team. [00:43:00] Or more accurately if Tracy, our brilliant global head of talent, tells her recruiting team, ~uh,~ you know that, ~uh,~ you should only build AI use cases. ~Uh,~ if you kind of know, know upfront, they're gonna have ~like~ a really meaningful ROI like, ~don't,~ don't build stuff that on a use case by use case basis is not gonna have ~like~ a strong OI.\\n\\nWell folks are gonna get pretty careful real fast about what they're building and ~you know,~ what we're seeing is it can be difficult to anticipate way upfront, ~you know,~ which of these, 'cause ~these are~ ultimately until they're validated. All of this is experimentation. Its experimentation for several reasons.\\n\\nThe most obvious one is from a technical perspective. We're not always sure or often sure upfront whether, ~uh,~ ai, ~uh,~ can solve the problem. So now we're not even talking about cost. We're just talking about ~the, ~the, the numerator of ROI. Like, will it have material incremental impact on this metric compared to, ~you know,~ conventional levels of performance?\\n\\nAnd the [00:44:00] answer upfront from almost all this is, we're not sure yet. That's why we're experimenting with it. The second, ~uh, you know, ~question we, we don't know upfront is, well, how much does it cost to build and run? And does that make for an attractive kind of baseline? ROI? It's very hard to know That upfront.\\n\\nDepends how you build it and how much compute is involved or how much upfront r and d was involved. ~Um,~ and then thirdly, and this gets under discussed, ~uh,~ Harun. Will people accept it, right? There's this human, ~you know,~ element of like, you know, we, let's, let's say you can build, ~um, you know,~ an ai, ~uh,~ approach to, ~uh,~ interviewing.\\n\\nActually, this is a live example. It's having, ~like,~ right now in the market, right? So it's like,~ you know,~ we're gonna, we're gonna AI our interviews, no more humans doing interviews, and we can show that the AI is, ~you know,~ 20% more accurate at assessing these skills than a human. And turns out it's not that costly.\\n\\nLet's say all of that's been proven out. So like you've got that ROI, it's like incremental impact [00:45:00] on skills assessment over relatively low cost. And you're like, we think,~ you know,~ that's great. But again, the under discussed portion is, well, what's it gonna take for, ~you know,~ all the folks involved in like being interviewed, for example, candidates to like, is do they want, do they wanna be interviewed that way?\\n\\n~Um.~ How will that influence our, ~uh,~ like the rate of applications from highly qualified candidates, will our competitors use, ~uh,~ keep a human in the loop to the extent that that's really valuable to, to, ~you know,~ the most promising candidates, ~um,~ as a competitive advantage ~to,~ to Zapier If we were, by the way, we're not yet.\\n\\nAnd maybe won't for a long time, but anyway. But you know,~ like, you know,~ if we take a human entirely out of the loop, ~like~ are we ~kind of~ giving away part of our game, ~you know,~ to our talent competitors? And so there's that piece. It's the, it's the more like anthropological, how does human behavior show up within this that isn't always, ~um,~ accounted for in an ROI calculation, but ought to be.\\n\\nWhen we're figuring out what to build. And so not only, ~uh,~ do we not tell, ~you know,~ tell folks up front, like you have to prove the ROI is gonna be [00:46:00] positive on a specific experiment. ~Uh, where we, ~where we can, like, we take more of a portfolio approach, right? Which is, here's a problem that we wanna invest in solving.\\n\\nHere's the, ~uh,~ KPI that we are obsessed with, by the way, ~it's one we're, you know,~ it should always be something the company already cares about. And we're gonna run three experiments, right? There are like three ways. ~And,~ and then ~maybe,~ maybe one of them like is the winner, maybe two of the three together, ~like, you know,~ have the impact that we intend.\\n\\nMaybe none of them work. And we go back to the drawing board and come up with three more, but taking more of a portfolio approach, but anchored on like specific problem spaces that, ~uh, where,~ where material improvements would make a big difference. Like they have a huge impact for the business and for customers. \\n\\nHaroon: Awesome. And Brandon,~ just as an aside, I totally lost track of time. I know we only have four minutes left and I'm assuming you hard stop here, uh, in, in four minutes. uh, that's right. Uh, that's correct. Right.~\\n\\nBrandon: ~Correct.~\\n\\nHaroon: ~Uh, so let me just, uh,~ just a couple of, ~uh,~ questions to round things out. ~Um,~ so my first question for you is~ we, so ~we surveyed our subscribers and the two biggest challenges that they were facing internally at their organizations was one, lack of integrations with, ~uh,~ data sources.\\n\\nAnd then number two is dealing with sensitive. Data from [00:47:00] their users. ~Um,~ I'm curious what the biggest challenges, other than the ones that you've already mentioned have been for the Zapier organization in enabling this AI transformation journey. If you had to list off maybe one or two.\\n\\nBrandon: Yep. Well, I would co-sign what you've heard from, ~you know,~ a lot of your stakeholders. ~Uh,~ and I think organizations are coming to this understanding that like ~our,~ our data is our differentiator. And so all the traditional like data and data science questions that our data teams are already working on, like they just become more, I.\\n\\nAnd the, but the good news is none of those questions are new, right? It's ~like, what,~ what data do we need to be collecting and for what purposes as a source of competitive advantage for our organization and to better serve our customers? ~Um,~ how do we store synthesize and sanitize that data? How do we connect data sources from disparate sources?\\n\\nWhat like, you know, query layer, application layer, do we put on top of that to actually answer questions and, and solve problems ~for,~ for the business? So the good news is those are timeless kind of data team questions, [00:48:00] and they're just more important now~ is the,~ is the short answer. \\n\\nHaroon: Then~ I, I know, Brandon, you have to jump in a few minutes here.~ Last question for you is, what has been the most surprising insight that you've derived during this journey?\\n\\nBrandon: Harun. I would tell you that the thing that you're making me think about, it may not be the biggest surprise, but I think it's the thing that I am paying attention to the most and it's giving me a lot of motivation and inspiration for the years ahead, is seeing how thoughtful human centered integration of AI into folks' jobs like the business impact is amazing, or it is really exciting to see that.\\n\\nThe impact on the individuals themselves is really neat, right? We're talking about abstracting a lot of tedium out of typical ways of working, such that humans do more of the things that humans do uniquely well. And like I could have told you that three years ago, right? Cognitively, like intellectually, ~that~ that might be the case.\\n\\nBut actually seeing how that plays through and the people that I get to work through every [00:49:00] day, like what,~ you know, ~what it means to someone to know. That they're being invested ~in,~ in this way and they're becoming truly elite in a future facing way. ~Um,~ not to mention how it impacts or improves, just like the everyday experience of doing the thing that they do best.\\n\\nIt's incredible. And, you know, I can't make any claims as to, at the macro level how AI's gonna influence the job market and the ins and outs of that. What I can see today are. Everyday examples of folks doing jobs that really matter to them in ways that I think are gonna serve them really well and their\\n\\ncareers. \\n\\nHaroon: It's incredible. ~it~ it sounds like one of the byproducts of ~this,~ this journey has been the folks at Zapier are becoming more confident and less scared by all of the different. Fears that surround AI transformations, ~uh,~ because they're getting hands on, because they're addressing the fear, ~uh, direct, uh,~ direct on and learning about where AI is maybe successful, maybe where it's not, ~um, rather than just, uh, theorizing, which, uh, you know,~ we could still say we we're, we're at the very beginning stages of AI adoption, ~uh,~ across enterprise.\\n\\nSo, ~um,~ I would say most folks are still [00:50:00] at that stage. \\n\\nBrandon: I think maybe one of the biggest, biggest takeaways for me from the last few years has been, ~you know,~ getting really specific, brings clarity and clarity, really helps with confidence. ~Uh,~ and ~you know, the,~ the type of clarity or specificity that I think really serves all of us with a transformation like this is like staying grounded in our existing mission stakeholders, whether that's students, customers, employees, whatever the case may be.\\n\\nLike that's the thing that's not changing, right? Like we all wake up to do a job for a specific reason and for specific people. And what AI's really about, at least as I think about it, is how can we take a huge step forward and how we do that, ~uh,~ together. \\n\\nHaroon: Brendan, thank you so much for your time.\\n\\nThis episode of Building with AI is brought to you by Autoblox, the company I founded to solve one of the biggest challenges in AI today. Making LLMs safe, reliable, compliant, especially in high stakes industries like healthcare. If you're a company building AI powered products, whether that's a virtual health assistant, AI driven documentation, [00:51:00] or any LLM handling sensitive information, you know the risks.\\n\\nLLMs hallucinate, they leak sensitive data, and in regulated industries like healthcare, finance, and insurance, compliance isn't just important, it's a law. So what's the challenge? Most companies rely on manual spot checking and vague governance policies. That's not scalable. What you really need is a way to proactively catch AI failures before they happen.\\n\\nBefore an LLM suggests incorrect medical advice or violates HIPAA compliance. That's where Autoblox comes in. We're an automated QA and compliance platform for LLM products, purpose built for regulated industries. With Autoblox, your team can run structured AI evals, detect risks before deployment, and enforce real time guardrails in production, all while maintaining auditability and compliance.\\n\\nWe're already helping companies like Hinge Health, Pelago Health, and others scale AI safely. If you're building AI for healthcare, finance, or any industry where safety and compliance matters, you need to be using Autoblox. Go to autoblox. ai to learn more or book a demo today. Again, that's [00:52:00] autoblox. ai, the trust layer for AI and regulated industries.\\n\\n\",\"createdAt\":\"2025-08-15T13:58:00.197Z\"}","label",[137],"606c5fd4-79de-11f0-8af7-0224a6c84d84","openai.ts","606c609c-79de-11f0-8af7-0224a6c84d84",7,["D",1755265645265],"import { OpenAI } from \"https://esm.town/v/std/openai\";\n\nconst openai = new OpenAI();\n\nexport interface BlogPost {\n title: string;\n content: string;\n}\n\nexport interface Newsletter {\n subject: string;\n content: string;\n}\n\nexport async function generateBlogPost(transcript: string): Promise {\n const completion = await openai.chat.completions.create({\n messages: [{\n role: \"user\", \n content: `Transform this transcript into a compelling blog post. Include a catchy title and well-structured content with headers, key insights, and actionable takeaways. Make it engaging and SEO-friendly:\n\n${transcript}`\n }],\n model: \"gpt-4o-mini\",\n max_tokens: 2000,\n });\n\n const result = completion.choices[0].message.content;\n const lines = result.split('\\n').filter(line => line.trim());\n \n return {\n title: lines[0].replace(/^#+\\s*/, ''),\n content: result\n };\n}\n\nexport async function generateNewsletter(transcript: string): Promise {\n const completion = await openai.chat.completions.create({\n messages: [{\n role: \"user\",\n content: `Create a newsletter from this transcript. Make it punchy, engaging, and optimized for email. Include a compelling subject line and content with bullet points, key quotes, and a strong call-to-action:\n\n${transcript}`\n }],\n model: \"gpt-4o-mini\", \n max_tokens: 1500,\n });\n\n const result = completion.choices[0].message.content;\n const lines = result.split('\\n').filter(line => line.trim());\n \n return {\n subject: lines[0].replace(/^Subject:\\s*/i, ''),\n content: result\n };\n}",[147],"a154b156-79e1-11f0-a69a-0224a6c84d84","main.ts",25,["D",1755267042931],4,"http","import { Hono } from \"https://esm.sh/hono@3.11.7\";\nimport { serveFile } from \"https://esm.town/v/std/utils@85-main/index.ts\";\nimport { uploadTranscript, getTranscripts, deleteTranscript, clearAllTranscripts } from \"./database.ts\";\nimport { generateBlogPost, generateNewsletter } from \"./openai.ts\";\nimport { createBeehiivDraft } from \"./beehiiv.ts\";\n\nconst app = new Hono();\n\napp.onError((err) => Promise.reject(err));\n\napp.get(\"/\", c => serveFile(\"/index.html\", import.meta.url));\n\napp.post(\"/upload\", async c => {\n const form = await c.req.formData();\n const file = form.get(\"transcript\") as File;\n const content = await file.text();\n \n const id = await uploadTranscript(file.name, content);\n return c.json({ success: true, id });\n});\n\napp.get(\"/transcripts\", async c => {\n const transcripts = await getTranscripts();\n console.log(\"Returning transcripts:\", transcripts);\n return c.json(transcripts);\n});\n\napp.delete(\"/transcript/:id\", async c => {\n const id = c.req.param(\"id\");\n await deleteTranscript(id);\n return c.json({ success: true });\n});\n\napp.delete(\"/transcripts/clear\", async c => {\n await clearAllTranscripts();\n return c.json({ success: true });\n});\n\napp.post(\"/generate/blog\", async c => {\n const { transcriptId } = await c.req.json();\n const transcripts = await getTranscripts();\n const transcript = transcripts.find(t => t.id === transcriptId);\n \n const blogPost = await generateBlogPost(transcript.content);\n \n return c.json({ success: true, content: blogPost });\n});\n\napp.post(\"/generate/newsletter\", async c => {\n const { transcriptId } = await c.req.json();\n const transcripts = await getTranscripts();\n const transcript = transcripts.find(t => t.id === transcriptId);\n \n const newsletter = await generateNewsletter(transcript.content);\n const beehiivResult = await createBeehiivDraft(\"newsletter\", newsletter.subject, newsletter.content);\n \n return c.json({ \n success: true, \n content: newsletter,\n beehiiv: beehiivResult\n });\n});\n\nexport default app.fetch;","transcript-processor",[154],"a14efd1a-79e1-11f0-8685-0224a6c84d84","index.html","603bb0a0-79de-11f0-b10f-0224a6c84d84",24,["D",1755267042931],"\n\n\n \n Transcript Processor\n \n\n\n

Transcript Processor

\n \n

Upload Transcript

\n \n \n

\n\n

Stored Transcripts

\n\n \n\n",[163],"609d8b36-79de-11f0-be4e-0224a6c84d84","deno.json","609d8bfe-79de-11f0-be4e-0224a6c84d84",8,["D",1755265645266],"{\n \"$schema\": \"https://raw.githubusercontent.com/denoland/deno/348900b8b79f4a434cab4c74b3bc8d4d2fa8ee74/cli/schemas/config-file.v1.json\",\n \"lock\": false,\n \"compilerOptions\": {\n \"noImplicitAny\": false,\n \"strict\": false,\n \"types\": [\n \"https://www.val.town/types/valtown.d.ts\"\n ],\n \"lib\": [\n \"dom\",\n \"dom.iterable\",\n \"dom.asynciterable\",\n \"deno.ns\",\n \"deno.unstable\"\n ]\n },\n \"lint\": {\n \"files\": {\n \"include\": [\n \"deno:/https/esm.town/**/*\"\n ]\n },\n \"rules\": {\n \"exclude\": [\n \"no-explicit-any\"\n ]\n }\n },\n \"node_modules_dir\": false,\n \"experimental\": {\n \"unstable-node-globals\": true,\n \"unstable-temporal\": true,\n \"unstable-worker-options\": true,\n \"unstable-sloppy-imports\": true\n }\n}",[170],"4852fa14-79e0-11f0-9a75-0224a6c84d84","database.ts","6046a3fc-79de-11f0-8034-0224a6c84d84",22,["D",1755266464136],"import { blob } from \"https://esm.town/v/std/blob\";\n\nexport interface Transcript {\n id: string;\n name: string;\n content: string;\n createdAt: Date;\n}\n\nexport async function uploadTranscript(name: string, content: string): Promise {\n const id = crypto.randomUUID();\n const transcript: Transcript = {\n id,\n name,\n content,\n createdAt: new Date()\n };\n \n await blob.setJSON(`transcript:${id}`, transcript);\n return id;\n}\n\nexport async function getTranscripts(): Promise {\n const keyObjects = await blob.list(\"transcript:\");\n console.log(\"Found key objects:\", keyObjects);\n \n const transcripts = await Promise.all(\n keyObjects.map(keyObj => blob.getJSON(keyObj.key) as Promise)\n );\n console.log(\"Raw transcripts:\", transcripts);\n \n const validTranscripts = transcripts.filter(t => t && t.id && t.name && t.content);\n console.log(\"Valid transcripts:\", validTranscripts);\n \n return validTranscripts.sort((a, b) => \n new Date(b.createdAt).getTime() - new Date(a.createdAt).getTime()\n );\n}\n\nexport async function deleteTranscript(id: string): Promise {\n await blob.delete(`transcript:${id}`);\n}\n\nexport async function clearAllTranscripts(): Promise {\n const keyObjects = await blob.list(\"transcript:\");\n await Promise.all(keyObjects.map(keyObj => blob.delete(keyObj.key)));\n}",[177],"a15c45ec-79e1-11f0-b25a-0224a6c84d84",["D",1755267043013],2,[29],"60517ab6-79de-11f0-a88f-0224a6c84d84",".vtignore","60517b74-79de-11f0-a88f-0224a6c84d84",6,["D",1755265645396],".git\n.vscode\n.cursorrules\n.DS_Store\nnode_modules\nvendor",[188],"b5b3f686-79df-11f0-b1f9-0224a6c84d84",".claude","b5b3f74e-79df-11f0-b1f9-0224a6c84d84",16,["D",1755266218155],"directory",[195],"pulls","branches",1,"forks","envVars",{"_209":64,"_211":64},"connections",{"_209":64,"_210":64},"count","unfilledCount","emptyCount",{"_44":194,"_33":195,"_31":196,"_139":-5,"_42":43,"_41":197,"_141":27,"_39":255,"_13":64,"_34":199,"_36":-5,"_24":27,"_144":-5,"_223":256,"_28":257},{"_44":247,"_33":248,"_31":249,"_139":196,"_42":43,"_41":250,"_141":27,"_39":251,"_13":64,"_34":20,"_36":252,"_24":27,"_144":-5,"_223":253,"_28":254},{"_44":187,"_33":188,"_31":189,"_139":-5,"_42":43,"_41":190,"_141":27,"_39":244,"_13":64,"_34":20,"_36":192,"_24":27,"_144":-5,"_223":245,"_28":246},{"_44":183,"_33":29,"_31":32,"_139":-5,"_42":43,"_41":14,"_141":27,"_39":241,"_13":185,"_34":35,"_36":38,"_24":27,"_144":-5,"_223":242,"_28":243},{"_44":176,"_33":177,"_31":178,"_139":-5,"_42":43,"_41":179,"_141":25,"_39":238,"_13":157,"_34":35,"_36":181,"_24":27,"_144":-5,"_223":239,"_28":240},{"_44":169,"_33":170,"_31":171,"_139":-5,"_42":43,"_41":172,"_141":27,"_39":235,"_13":64,"_34":20,"_36":174,"_24":27,"_144":-5,"_223":236,"_28":237},{"_44":162,"_33":163,"_31":164,"_139":-5,"_42":43,"_41":165,"_141":27,"_39":232,"_13":157,"_34":20,"_36":167,"_24":27,"_144":-5,"_223":233,"_28":234},{"_44":153,"_33":154,"_31":59,"_139":-5,"_42":43,"_41":155,"_141":27,"_39":229,"_13":157,"_34":158,"_36":159,"_24":27,"_144":160,"_223":230,"_28":231},{"_44":146,"_33":147,"_31":148,"_139":-5,"_42":43,"_41":149,"_141":27,"_39":226,"_13":64,"_34":35,"_36":151,"_24":27,"_144":-5,"_223":227,"_28":228},{"_44":136,"_33":137,"_31":138,"_139":-5,"_42":43,"_41":140,"_141":27,"_39":222,"_13":64,"_34":20,"_36":143,"_24":27,"_144":-5,"_223":224,"_28":225},["D",1755266386998],"children",[],[137],["D",1755265645265],[],[147],["D",1755267042931],[],[154],["D",1755267042931],[],[163],["D",1755265645266],[],[170],["D",1755266464136],[],[177],["D",1755267043013],[],[29],["D",1755265645396],[],[188],"b5c63cce-79df-11f0-ae09-0224a6c84d84","settings.local.json","b5c63d8c-79df-11f0-ae09-0224a6c84d84",18,["D",1755266218238],"{\n \"permissions\": {\n \"defaultMode\": \"acceptEdits\"\n }\n}",[],[195,248],["D",1755266218155],[213],[195],["D",1755194257445],{"_98":97,"_105":101},{"_98":97,"_99":-5,"_100":101,"_18":57,"_102":103,"_104":27},{"_13":87,"_33":88,"_44":43,"_89":45,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":46,"_96":97},"ctx",{"_80":-5,"_124":295,"_296":297,"_298":299,"_300":27,"_301":27,"_302":27,"_303":25,"_304":-5,"_305":-7,"_306":-7,"_307":64,"_308":27,"_309":27,"_310":-5,"_311":27,"_312":27,"_313":27,"_314":-5},"ENV",{"_266":267,"_268":269,"_270":88,"_271":272,"_273":274,"_275":276,"_277":278,"_279":280,"_281":282,"_283":284,"_285":286,"_287":288,"_289":290,"_291":292,"_293":294},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","fdb67791413f4386b4dbf18f354029a617f6160f","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

Diff · v26