Hai Guan: Started. Alright, cool. I’m gonna share my screen here. Let’s see… Can you guys see my screen? Should be a deck that says… Data Storytelling 101? Yeah? Okay. Cool. Alright, great, awesome. Cool. So, why don’t we get started? So, hey everyone, good morning again, welcome. My name is Hai, and Typically, we have, our colleagues here as well, Shane and Travia, but this time, for this lesson, I think that’s just me, running the show. So, Yeah, so my name is Hai.
I’ve been in the data science and analytics space for roughly 20 years, now, and I’m a co-founder at… the AI Analyst Lab, where we provide educational materials around Agentic analytics and how AI, sort of, like, practical workflows for using AI in our day-to-day. Specifically, we focus a lot on the analytics side of the house, data science stuff. But I think the, sort of, like, the things that we teach, could be pretty general purpose, and for those of you who have followed us. Previously, in other lessons, we run these on a weekly basis, and we have a lot of fun sharing what we learn and all the resources and stuff with you all.
As I mentioned, we have two other colleagues that are not able to join us today. One is Shane Butler, the other colleague is Sravya Madapali. We rotate in terms of running these, these different workshops, so next week, we have, we have another lesson where Shane’s gonna walk through semantic layers on how to guide AI to, give Trustworthy data results. And you can follow us on Maven or on our website as well, for any resources. A ton of free resources around workshops, past recordings, stuff like that. We also have paid courses, which we’ll go into at the end of the… at the end of the hour. Okay, cool. Well, let me, let me, let me make sure to, also, sort of, like, set the context here.
Today, we’re talking about data storytelling, we call it 101. Mostly, it’s focused on building an executive readout, with AI, and for this for this workshop, we’re going to leverage Claude code on the desktop app. So, for those of you who are familiar with that, You can feel free to follow along, but we’ll have all the resources, afterwards as a follow-up email, plus as well as the recording for this session. Cool, alright. So, let’s do a quick poll in the room here. So, what… So, let’s say if someone is senior at your company, at your workplace, and they’re asking for, you know, something very generic, like, hey, how’s the number… how does the number look? Or how the members look.
What would you… what would you send? In terms of your response. One being, hey, just a dashboard link that has the data. Two is a deck with a lot of slides, three, an email with numbers, or four, you don’t do anything, you postpone it. Let’s see, a lot of threes, a lot of email with numbers… Some twos, some 1s, nice. Cool. Diverse set of, diverse set of experience, or a diverse set of, techniques here. Okay, cool. Yeah, none of these are… you know, like, none of these are wrong, per se, but there are certainly ways that are, that would be more effective in terms of kind of, like, getting, what you’re after, and also addressing the question behind the question.
And that’s everything that we’re going to be talking about in the next, in the next 50 minutes or so. we’re gonna go through maybe, like, 10-15 minutes of content, through kind of, like, slide deck, here, and then we’ll get pretty… and we’ll do a demo, for maybe, like, 15-20 minutes, depending on how quickly, Claude runs, runs the, runs the… runs the command, and and we’ll go into Q&A afterwards. Cool. Okay, so before we get into the rest of the content here, I want to flip through a deck with you all, and feel free to drop in chat what you think of it. Let’s see… okay, cool.
So I think Attendee said, a deck with 10 slides, So, let’s… let’s say I have a deck here that, that talks about, kind of, like, a fictional company. Where, we did a churn analysis, and this is sort of like the deck that was presented, or that was given to, an executive. I’m just gonna flip through them, don’t worry about the numbers or anything, just look at the form and the shape of it, and, you know, drop it in chat, like, kind of, like, what, what your first reactions are. Cool. So, churn analysis… There’s an executive summary. There is a methodology. Followed by charts around different cuts. Followed by more cuts. And perhaps some descriptions of the numbers?
There’s another chart around monthly trend. More descriptions around some of the different dimensions that, that we’re looking at. health score distribution, so this is a B2B, company that, Again, don’t worry about the actual numbers or anything, but just the content. And I’m seeing in the chat, TMI, lots of text, hard to digest. And then we’ve got some scatter plots here. And then a cohort chart, cohort heat map. some comparisons… More comparisons… more charts… And then we have key findings here, summarizing what all those, all the data is suggesting. And then, recommendations. And then, closing out with questions. Alright, cool. Let me read the chat here.
Trying to show impact through data viz, impact too in-depth for executives. Attendee, that’s great. Too many slides. Cool. So it’s… or I guess, maybe I’ll ask, one question here. Who here thinks this is great? Like, this is… This is a good deck for executives. All right, Attendee says, will an Exec read all that? Just the summary slides? Yeah, so precisely. So, you know, like, I think this is… you know, like, everyone… everyone… everyone gets the gist of it. I think this is definitely an outlier here, like, we’re… we’re illustrating something extreme.
But it’s not uncommon for some of the decks that I’ve seen as a… as a data professional in the past, where, you know, it’s just a lot of information packed into one giant deck that, at the end of it, kind of, like, the messages gets lost, and, like, what the focus is, is unclear. So we’re gonna, you know, we’re gonna try to fix that, in the next, 10, 15 minutes. And, and, and see how we can do it with, with Claude, as well. Alright. So, jumping back to the slides, so one fundamental thing that we’ve seen, Just now, with that deck, is that it’s fundamentally answering the wrong question.
So, if we reframe it a little bit, kind of like stepping back a little bit, the original question was, hey, an executive was asking, hey, how are the numbers looking, right? Like, that’s… that’s on surface, that’s what the question was. And so that, 17-slide, deck there is basically answering the question of what data do we have? like, literally attacking the level of the question on service, trying to sort of, like, just address it head on. And so… you know, got label titles, you’ve got the executive summary, that’s pretty long, right? Followed by methodologies, followed by a bunch of metrics, bunch of cuts, a lot of different charts, and stuff like that.
But at the end of it, like, generally speaking, right, like, an executive asks a question is less about how the numbers look, like, not quite that literally, but more like, hey, what, you know, like, I’m curious about this thing. what should I do with it? And, kind of like, you know, the question behind the question generally runs a little deeper than what the actual question was. So… What, you know, like, if you, if you kind of, like. go along that route of thinking, then it’s really a different question. It’s really more about, like, hey, what should I do, based on what you observed?
Like, I raised a question, but then I’m not here to sort of, like, you know, like, determine the answers myself. So it’s really a different deck that, that the person is, or a different artifact, doesn’t have to be a deck, but we’re gonna do a deck, in this, in this, in this session. It’s really a different artifact altogether, so the analysis underneath that, deck that we just flipped through, that is fine. Like, assume all the data, everything is correct. Like, there’s no, you know, like, there’s no, there’s no errors, in them. Their work is fine.
It’s just that… Sort of, like, where things are and how they’re positioned is, is what’s impeding, sort of, like, the impact that it could ultimately have, if we just reorganize it a little bit… little differently. So, there are four things, I think, that could be pretty helpful as we think about how to surface the, a very effective structure to, on, you know, any given readout, or any given deck, for kind of, like, an executive audience, or somebody that is less Into the detail, if you will, and is trying to get at kind of like the, the original premise that we just talked through, which is, hey, what should I do with this information? I raised the question, what should I do with it?
So, number one here is, leading with the insights. So, all the analysis that’s been done, at the end of the day, it goes back to kind of like, hey, what were the insights? It’s not what the data is, it’s really around, kind of, like, what’s your observation and what’s your interpretation of it? So, over here on this slide, we’ve got two columns where The left is really the label, like, just describing what, for example, what the charts are, and we saw a bunch of those, right? Like, when we flipped through the… when we flipped through that deck, where it was literally just, you know, turned by segments, or revenue overview, or key findings.
kind of, like, wasted space, where you can actually do quite a lot more than just describing what this is, when in fact, like, you know, people reading that is, you know, like, they probably already understand that, hey, you’re showing me the, you know, the churn chart by segment. So. Some of the really effective ways to get at kind of like the, the… the main thing, the meat of it, is really just… just put the claim… put the claim in. Like, leading with the insights, and cut out any, anything, anything that’s… that… that’s not adding value. So… For example, here, instead of churned by segment, you can have some sort of churned account skipped onboarding, at a ratio of 4 to 1.
Instead of Q3 revenue overview, you can have, we lost $10 million in ARR, and most of it was preventable. Instead of key findings, you get straight into the finding. Like, so, we should fund the onboarding push, or keep paying this four times, churn, then we… that’s… that could… that’s probably unnecessary.
So, the idea really is, when you go to that, when you go to that slide, or when you go to that title, it should just very clearly State what your… what the, what the insights and the finding is, and that’s one way to sort of, like, keep the interest to people’s interest, especially for someone who is generally pretty, you know, short attention span, because they have a ton of stuff on their plate and a lot of decisions to make. Cool. Alright, and the second one that’s… pretty helpful, is kind of like, we call it three slides here, but I would treat it a little differently. Perhaps think of it as, like, three sections. Where the first one is the problem.
So… Stating… stating very clearly what the problem is and why the… why the audience should care about it. So, it would be some sort of describing, hey, what’s at stake, for example. but not, like, the fluffy things, or, like, the filler words, or whatever. It could just be, like, you know, like, it could be the meat of it, where, you get straight into kind of, like, you know, hey, if we, you know, like, hey, this is the thing that we’re tackling, this is why we want your This is why we want your attention to this problem. And then slide two, or the second component of it, would be the insights.
So, leading with the finding that we just saw, using the title, stated as claims, and you can back it up with the most salient evidence that you have, and generally that comes from data, obviously, but not, like, overwhelming people with, like, hey, I’ve done, you know, like, 10 cuts, or, you know, 12 cuts that we saw. Generally, all those 12 cuts would lead you into that one thing, or that two things at most, where sort of like, you know, hey, these are the really important artifacts that would support the claim, or that would support the finding that you should know about, and then just, just, just isolate it to that.
I think a lot of folks really like to kind of, like, show their work, if you will. Like, hey, I’ve done all the work for all the other stuff that I’ve… that I’ve looked at, so I need to show it. I think maybe it was Steve Jobs or somebody, it’s something about, like, you know, simplicity is the ultimate sophistication. That applies here as well. Cool. And then the third one is the ask. So, the third component, or slide three, would be the decision that you want this executive to make, like, the action that you want this person to take on.
So, you know, a lot of it could be in the form of, hey, if you don’t do this, then this is what it’s gonna cost you, or that if you do this, this is what the upside is, stuff like that. So… You know, like, so, kind of like a closure, like, the entire workflow is really about the… what the problem is. why you have this, finding, like, what the insight is, that would tackle this problem, and then what you can do about it. So, kind of, like, closing… closing the loop in three steps, in three components, and then everything else is appendix. So, you know, like, you have 12 different cuts, or, like, maybe, like, the third chart, the fourth chart, they’re still useful, they’re still relevant.
you can put in an appendix, for example, or, like, a detail section, where it’s, like, below the fold. When questions come up, you can always refer to it instead of, oh, I didn’t put any here, for example, but you don’t have to show it. Cool. So, those were the first two. The third one that’s really helpful is to take a position. So, what does that mean? maybe the example on this slide would be helpful to sort of, like, digest to illustrate that. So, taking a position means, what, you know, from… in your opinion, based on your business knowledge, based on your expertise. based on what you know from looking into the data yourself, what would you do about it?
Like, what… what is the specific action that you would recommend? And, generally, I think kind of like a, you know, like a non-action, that, that… masquerades as an action, typically would look something like the left box here, which is, oh, continue to monitor the trends, or, you know, some sort of, like, continue to do, you know, like, continue to look at the growth trajectory, or whatever it is. That is fine, in that, you know, hey, this might be important, but then it really doesn’t drive any action. It’s like, oh, yeah, like, business as usual. Then, at the end of the day, it really is, like, then, you know, like, what are we doing here, then, if it doesn’t lead to any changes?
So… the thing that I like to talk about for at least my team at work is that impact comes from, you know, a few forms, but generally, it’s It’s, what can you change because of the work that you’ve done? And typically, something like, hey, let’s just continue to monitor the trends, is not something that changes at all, and most people would just forget about it. Whereas on the right side, which is, hey, we should fund the onboarding push here, or accept a four times higher churn, on almost a third of the base.
That is very specific, where, like, you know, hey, you have a conviction that this is worth Experimenting with, or testing on, because you’ve seen the data, and it seems to be pretty… a pretty compelling case for you.
And even if you’re wrong, which is… which is fine, totally fine, you can… you get the dialogue started, you get the discussion and the debates going, and typically speaking, you don’t have to be right, you just have to be the person that sort of, like, gets people to talk about and care about this topic, and then maybe, at the end, the recommendation does not get adopted, but at least something is being done with it, and you’re being credited with the, kind of like, you know, like, the initial push. for it.
And at the end, if there is a really good outcome off of some of the tests that you guys have decided to run, or, you know, the decision maker approves, then it all goes back to you for For starting the, the conversation in the first place. All right, cool. And then, let’s see, the last one here is, is to anticipate the pushback. So, there’s probably, like, I’ve never seen a… a, a presentation, or, like, a readout, or, like, a, like a meeting with an executive where there is no pushback, where there’s no question, there’s no follow-up, and everything is just like, oh, yeah, great, like, that sounds, that sounds awesome.
The idea is not to, you know, like, get the artifact, or the readout, or the deck to be in a place where people have zero question. it’s really just about anticipating the pushbacks. They could be the obvious ones, or they could be, you know, like, you know, one level deeper, but just kind of, like, anticipate it as much as you can up front, and prepare for some sort of, like, you know, responses to that. And the more you can do it, the more confident, actually, you know, in the meeting, or in in the exchange, where the executive would be like, oh yeah, you know what you’re doing.
Like, you absolutely know All of that, because all my questions, you didn’t even… like, it seems like you’re well prepared for it. So, you know, some, some, and we’ll, we’ll get into the demo here, where we’re gonna, we’re gonna see some of these, in this fictional dataset that we just, we just flashed. You know, hey, is this… like, these are really common ones? Is this causation, or just correlation? Why should I believe in this number? this is what happens if we do nothing, or who else has looked at this, and am I the first one?
So, you know, always sort of like, you know, healthy exercise to do is, write down the hardest questions that an executive could ask, and then answer them in advance. And, you know, like, the appendix is a really great way to arm yourself with answers to those. Whenever they come up, you can always just flip to, oh yeah, by the way, we have the answer already, and here’s the evidence. Alright, Attendee. Okay, cool. So those are some of the, kind of, like, the high-level concepts that I wanted to go through in terms of, in terms of the, structurally some of the components that are really helpful in putting together an effective, Exec Readout.
So now, for the next, maybe, like, 15 minutes or so, I wanted to do a little bit of a demo, where, we, we’re gonna work with Claude to, to, to, to do all of these. Alright, let me swap my screen here a little bit. Alright, cool. So… Okay, so… hoping you guys can see this. Do you see this folder here? That’s called DS101 Demo? Just want to make sure I’m sharing the right screen. Okay, Attendee says yes, great. Cool. So, I have this folder here that you can think of as… it’s got some, It’s got a data set. This is a fictional company, a B2B company that has, You know, like, an account information, where, we know that, you know, where… who the… who the account, is. Flash it to you very quickly.
Who the account is, what segment they come from, what industry region, and then what contracts they’re… what subscription plans they’re on. Did they onboard? Did they go through onboarding with the customer success team? Did they churn, what’s their ARR, and stuff like that. So, pretty straightforward, dataset here about different customers for a B2B, made-up company. And so this folder here, think of it as it’s got finished analysis, where it has charts and stuff that’s been computed, it’s got this, obviously, the data sets, some of the insights, stuff like that. So we’re not building, like, an analysis here from scratch.
We’re… what we’re gonna do is, turn this into… or, I guess, use the principles that we just walked through and turn it into an executive readout such that it abides to some of the things that we walked through. So we’re gonna use Claude on desktop for it. So this is Claude. You guys might… You guys might have seen this, already. And so, I am going to use Claude Code for this. This would work. Also, on co-work, if you’re familiar with that, but I’m just gonna go with Cloud Code, because I’m more familiar with that myself. So, over here, this is Clock Code. And what I’m gonna do is spin up a new chat, and I’m gonna point plot code to that specific folder.
So this is the folder that we just saw, DS101 demo. So, this is… where I’m at, DS101 demo. So Cloud Code already has context on the, the specific folder that we’re working on, and all the underlying data and stuff like that. So, what I’m gonna do is I’m gonna paste in a series of prompts, and we’ll go through them together as well. Alright, so… This is prompt number one. Which is… You are helping me build an executive Readout. Read the analysis in this folder, the data, the charts, the notes. My audience is our CFO, and the decision on the table is where next quarter’s customer success budget should go.
And I want to rank my findings by how much they should change, what the audience does next, ignore how interesting a finding is, rank by decision rates. Then give me two or three candidates. Candidate leads for each one, the findings stated as a one-sentence slide title, the single number that carries it, the decision it points to. So tell me which one you would lead with and why, then stop and let me pick. I’m gonna add in a… another sentence here, because already… because I made a skill for this, so I don’t want it to use it. Do not use the Exec… Readout skill.
So what I’m doing here is I’m just using Claude, with the context of the folder to just reason, at the For itself, step by step, such that by the end of this, we’ll get the readout that has the principles that we talked about. I just fired it off. It’s probably gonna take a couple minutes. I’m hoping it’s… faster than that. But, right now, for folks who are… I don’t know how much you guys use, Spot, for example, they just came out with, Fable 5.1 yesterday, so a really capable model, but again, like, pretty expensive at it. So, I’m having Opus 4.8 do this for me, because, you know, there’s different models where the latest Opus is Opus 5, but for folks who are using this pretty heavily.
you might have either come across the discourse on LinkedIn or whatever, or you might have experienced it yourself, where, it’s, kind of like the responses is really hard to understand from, from, from Claude with Opus 5. So I default back to a bit, you know, like an older model, like Opus 4.8, and that’s totally fine, where I understand a little bit better, on what, you know, like, what the responses are. All right, cool. So, what, what Claude is doing right now, you know, you can… you can take a look at this, and if it runs long, I’ll… I’ll just show you a pre-run thread here.
But the idea is really, like, oh, we’ll read the notes, the data analysis, and then they’ll look at the data directly, and then verify key numbers, and then doing all that to inform what should be the two or three candidate leads to surface in the readout itself. Alright, cool. So, looks like it’s done pretty well, pretty quickly. All right, so, so here’s the… here are some of the… the findings that, Quad has come back with. Which is, hey, there is, again, like, you don’t, you don’t have to know about the data sets, like, the actual data itself, but just really the, kind of, like, the form factor, and assume these numbers are correct.
Okay, finding number one is onboarding completion and the churn engine, so there is opportunity. Basically, it’s, there is opportunity to get people to get onboarded. Like, for folks who haven’t been onboarded in this fictional company, their churn rate is much higher, and it’s also quantified how much opportunity there is, based on that, little kind of, like, finding here. And then there’s some other findings where, account churn nearly twice as hard for uncovered, so, like. Accounts that don’t have a customer success manager, they typically… they tend to turn, much more so than the ones that do.
And then… and then some sort of, the account behaviors, right before they churn, and then so on and so forth. So, different sort of cuts where, Claude has already looked through the data, and then, using its own reasoning, kind of, like, judged based on our prompt. that these are most likely the most impactful ones that we should lead for the CFO, and obviously it’s got Different, reasoning, like… Reasoning why, why it is. So, the recommendation is the onboarding piece, and and I agree with that. So, let’s go for the next prompt here. I’m gonna fire it off first, and we’ll read it together. Cool.
So, the recommendation is really the onboarding claim here, which is, we have an onboarding problem, onboarded accounts, turn at 4X higher rates, and there are $32 million of live ARR sitting in that bucket. So, different accounts, like, accounts that account for, $32 million have not been onboarded yet. So, you know, taking Taking the, the reasoning on that one, like, you know, like, maybe those are the, the really good ones to target. So, this follow-up prompt is really, hey, I’m leading with this, never finished onboarding, trying to 4 times the rate, and almost a third of our base never finished it.
So, structure the readout to three slides, kind of like what we just went through, the problem, the insight, the ask. And some detail around what those are. The problem is why this audience should care? With the stake as a number, the insight is, the… this… this lead here as the slide title, and then backed by two or three strongest pieces of evidence, similar, very, almost the exact same… same thing that you’ve seen, just a… maybe, like, 10, 15 minutes ago. And then what the ask is, so the specific decision I’m requesting, with cost, timeline, and the risk if we don’t do anything.
And every slide title must be a full sentence that carries the story, give me the outline, and then list what gets cut to the appendix. So what I’m asking for is just an outline for now, like, what that looks like, before it builds the whole thing. And then I am here to sort of, like, approve it or, or ask it for revision. Okay, so the response is, you’re leading with the onboarding opportunity, here’s the three-slide outline, the problem, which is churn, they raised $10 million in AR this period, and the losses are concentrated where we can still act. And, and so on and so forth. So this points to, you know, like, why, why we should care, and… what the opportunities are.
So this is framed as, like, you know, hey, this is the, this is the thing that we can intervene. And then, next slide is the insights, which is, the account that never finishes onboarding, churn at four times the rate. We’ve seen that already. And, and it’s gonna… it’s gonna have the data to back it up, in terms of, kind of, like, the different evidence here, cited here, and we’ll see it in the final build. What that looks like visually. And then the ask, fund a $350,000 per quarter onboarding pod, to protect the $32 million ARR that is still onboarded. So… Here’s the decision. So again, like, going back to the take a position. principle.
You know, like, even if you’re wrong on this claim, it’s fine. It is really the first two. That is carrying the load around, this is a problem worth solving for, here’s the finding that we should head towards this direction, and here’s one way that we can act that can, that can prevent this you know, like, that we, that we can arrest this, this, this, this, this, churn phenomenon. It could be that the executive does not agree with you, because he or she would have more business context, or, you know, like, whatever it is, or, you know, insider information, or something else, and that’s totally fine. You’re there to spark the conversation. Cool.
All right, and then it’s got, you know, cutting to appendix, these would be covered there, different, the different dimensions, different cuts, different slices of data, and, and telling, telling us, why it decides that these are not worth keeping as the most salient, Mostly on Insights. Alright, cool. And I will paste in a third prompt here, which is. build a readout from the outline we agreed, 3 slides plus an appendix, render it to a PDF, and then… The rules for how to build this, one idea per slide, title carries the finding, numbers come from the data only. If a number is not in the analysis, don’t use it. A chart earns its place only if it defends the insights.
So, one per slide at most. The… Appendix, should have the evidence, cut material, backup detail, nothing decorative. So it’s gonna go ahead and, this is basically approving this outline that we’ve, that we’ve laid out, and, it’s gonna go out and build its own, build the thing, itself. This might take a bit, even though the previous… The previous few… Prompts have been pretty fast. And in this, folder here, there’s also some style guides, where, if you ask Claude to just build like a deck or whatever, it’s gonna look kind of generic, or like, you know, AI sloppish, where it’s, you know, it’s just… kind of, like, that’s the default, right? Like, the average of everything that it’s trained on.
And… on… in this folder, the DS101 demo here, there’s some style guides to make it a little bit better, and, you know, like, incorporates the, some of the different best practices in data visualizations, and also color palettes and stuff like that. So it’s gonna, you know, go through a little bit, but it’s gonna follow that instructions as well. Okay, cool. Let’s see, got a question from Attendee. Have you found that Claude is better at making and editing presentation content than other apps like Gamma? Yes, absolutely. before, I would say, Claude Opus 4.6, I was a heavy user, and we’ve been a heavy user at the lab.
with Gamma, and ever since then, we’ve never touched anything other than, Claude or Codex, you know, just the… just the LLMs themselves. They’re really, really good at, at building decks. All assets are clear. Let me compute the exact backup figure, so it’s… Trying to do some of the… getting together the evidence and, and putting it all together. Let’s see… Will you share these prompts at the end of the session, please? Yes, you will get, everyone’s gonna get the prompt pack, the folder itself. like, the practice demo itself, so you can run it on your own. And then there’s actually a skill that does this in that folder as well, and so, that would be something cool to check out as well.
I’m just gonna say something as well, maybe he’ll come back to me. Come on, Claude. Okay, numbers are confirmed. Now, write the deck. Alright, hopefully, hopefully it’ll give us… Soon. Attendee, your question is, I’m from a small Caribbean island, basically looking for free tools to do something similar. Is that even possible? You can use the base you know, just Quad or, or ChatGPT, or Gemini. They’re pretty good at, at doing these, these, kind of, like, general purpose… they’re basically general purpose agents that you can, That you can just use, and, they’re 20 bucks a month at the… at the pro plan, so, relatively pretty, pretty affordable for what they deliver for you.
Alright, so it’s, okay, cool, so it’s done… Coming… it’s done… so it’s finished with the deck, and we’ll take a look. Quickly. We’ll flip through them quickly, and we’ll move on. Okay, actually, before we flip through them, this is the last prompt that I wanted to show you guys, where… You know, like, if you remember, the fourth principle here, aside from building the artifact itself, is come up with the pushbacks, like, prepare for the pushback, and so this is a prompt where we’re gonna… we’re gonna do that. This is the red teaming, the… the analysis itself. So this is the prompt. I’m just gonna fire at it, and we’ll read it. Now be my toughest reader. You are the CFO.
You have seen a thousand readouts, and you protect us budgets. Attack the readout. What are the five hardest questions you would ask the room? Where my evidence is the thinnest? which number you would challenge first, strongest case for doing nothing, and then switch back to my side for each objection, draft the answer I should have already, and then using only what is actually in my analysis. Like, don’t invent… don’t make up anything. And if I cannot back something up, flag it, and so I can soften it, blah blah blah blah. So, basically, this is… Doing an adversarial review of the thing that we come up with, so it helps you to be prepared for a discussion with that executive. Okay.
Let’s see, so here’s the readout. It is. to read. Let me just bring it over. Okay, so this is the readout. And this is not the readout. This is the Readout. Okay. Sorry, let me… It’s here. Cool. Okay, so slide one, churn erased $10 million in ARR this period, and the losses are concentrated where we can still act. That’s the exact thing that we… that we came up with, right, on the, in the outline. So, $10 million ARR loss to churn, 229 accounts, 110 million total booking at stake. So on and so forth. So again, just focus on the form factor, not the actual numbers and stuff.
And this is the insights slide, where accounts that never finish onboarding churn at 4 times the rate, and a third of our base never finished it. So complete onboarding churns at 4.8%, never complete onboarding churns at almost 19%. We have roughly two-thirds of churns traced here, and together, there’s $32 million in exposed ARRs. And then the third one is the ask, which is fund a $350,000 per quarter onboarding pot to protect the $32 million in live ARR that is still sitting, and here’s the decision that makes it easier… makes it easy for… to kind of, like, get an… almost like, you know, put it on a platter for the executive to say yes or no, or, you know, ask for more questions.
Here’s the cost of doing nothing. And here’s the… here’s the timeline. And here’s the risk, or I guess here’s the risk if we do nothing, and here’s what the cost of doing this is, but how much of a return we’re looking at. So that’s it. And then everything else is appendix, a ton of these, different cuts, and… And whatever else, so… index C, A, B, C, D, and E, these are different methodologies, different cuts that we looked at, coverage, and stuff like that. Cool.
So hopefully that feels a lot more coherent, a lot more, straight to the point, and a lot more opinionated, which is generally what execs are looking for when they are interacting with, you know, some sort of, like, decisions or some sort of, kind of like, you know, like, data interactions. Cool. Okay, so the, the final prompt here, the response to that from Claude, is that, okay, CFO is in the room, here are the five very tough questions that would be really good if I, can kind of, like, prepare for it. First one, causation. you’re selling me a fix for a cause you haven’t proven, because we’re looking at, you know, like, just, just simple slice and dice of the data.
So, obviously, that’s observational, that is correlation-based. So, there is no causation being claimed, but then it may have sounded like 1, so be prepared to speak to that, for example. Protect $32 million is a sleight of hand, protect does not recover, and… okay, whatever it is, I’m not going to read through it, but it’s basically questioning, like, hey, is $32 million really the ceiling here? Or, like, the actual thing that we can actually, you know, like, impact? What exactly is this period? So, clarifying what that is. And, How any evidence that the pod that you’re suggesting would work. And then, and don’t inflate the numbers on me.
And, if you remember the prompt, we also asked it for, you know, okay, what evidence is the thinnest? And so it’s come up with these, these guys. And it also has the backing story, kind of like drafted answers for how we can address some of these questions, here for us to read already. Now, I would absolutely caution that you exercise judgments Like, you know, like, AI, LLMs, Claude, Codex, ChatGPT, They all would come up with anything that might sound very sensible. But it may not actually be appropriate for your business domain and your context.
So in the lab, we are very cognizant of all that kind of stuff, so we always advise people, and we have framework and stuff to really check your numbers, and then also make sure you don’t outsource your understanding to, kind of, like, to AI, because that’s the fastest way to, sort of, like, get burned. So, you know, different, different answers. I would treat it as drafts, and make sure you understand and make sure they make sense to you. But, like, these are kind of, like, assistance, if you will, if you treat it like that. Alright, cool. That is all that I wanted to show you, and there is one more thing here, which is there is an executive Readout skill in this folder here, too.
So you can see that, you know, a slash, and then followed by the skill name is, something that is a skill in Cloud that you can, almost like a recipe, if you will, that would do the things that you wanted it to do, pre-coded. So this is turn a folder, finish analysis into a three-slide executive readout. ranks the findings, asks the user to pick the lead, and so on and so forth. Basically, all the prompts that we have turn into a skill that would be very easy to replicate, so you don’t have to, like, you know, paste in prompts and stuff. That’s yours to keep. Alongside with all the different prompts as well. Alright, so we have 9 minutes. I’m gonna flip through this very quickly.
The last of these guys, hang on one sec, sorry. Alright, cool. So, let’s see, so we talked a little bit about this, what did the AI… what did the AI do, and what stayed human, the drafting, so coming up with the different The different arguments, the insights, the flow, the outline, all that. AI is really, really good at. You know, any of the LLMs are amazing at coming up with that. Now, the judgment cannot, again, cannot emphasize that enough, like, understanding cannot be outsourced. Your judgment still is you. Like, you’re still ultimately accountable for both, you know, do the numbers look right? Are the arguments coherent?
Do they… do they make sense based on the business context and domain knowledge you have about your business, about your company? So all of that is yours. So, just make sure you don’t outsource any of that to AI and just blindly be like, oh, yeah, that’s what… that’s what Claude says. That’s not gonna apply, obviously. And, let’s see, what are these? Yeah, so you’ll get all these things, you’ll get everything here, in your, in the follow-up email as well. And I will paste in some links where, I talked a little bit briefly about, kind of like, we run these lessons, every single week. Next week’s lesson is around building semantic layer, so AI defines your metrics.
and you don’t have drifts or, you know, like, definitions that can be defined, you know, like, many multiple ways in your company. I think that would be a really good one to attend for folks who are interested in something like that. So here’s the link. You can obviously scan the QR code, or I just dropped it in chat, you can sign up directly from that as well. And at the beginning of the session, I also talked about We also run paid courses, and so two that are coming up. One is on building Agentic analytics, building an AI analyst, so almost like, you know, what I just showed you is, like, a very tiny portion of what’s possible in terms of building out an, an Agentic analyst.
That would, that would be specifically tailored to your company’s context, understanding how to do Kind of, like, how to… how to do the analysis itself, setting up the environment, building out the skills agents. And, and, also sort of like, you know, understanding how to check its work, how do you know things are correct, how do you have guardrails, evals, and how do you leverage different models to, to help you, with any sort of data analysis or analytics workflow that you might, that you might do in your company. So, we have that starting next week. The next cohort starts next week. It’s a 5-week course. We’re doing this for the first time. Previously, it was a weekend bootcamp.
Now, because there’s so much content, and there’s so much latest things that we wanted to incorporate, we’re making it a 5-week, sort of like a 5-week course. 20% off with, with the code here, or, if you just scan the the thing, or I will just also put it in chat. Here… And then, finally, we have, another course around the analytics thinking, so asking better questions, getting… making faster decisions. It’s really all about the fundamentals of how to understand, kind of, like, what, What questions are worth asking for, and then delegate the execution piece to AI.
To… to help you do the… do the, the actual analysis, for example, and then how to, you know, towards the back end of it, how do you then wrap those up into something similar to what we just did here, but at a much more involved scale, to convince, influence, and drive decisions that would actually change the outcome for the business? So that is in October. Again, 20% off on that one as well. It’s called AI Analytics for Everyone. A lot of folks who have followed us have done either or, these, these courses. And I just dropped, the… The link for that one as well. Cool, and I think we’re doing pretty well on time. That is all that I wanted to go over today.
Hopefully that was helpful, and I can stay over a little bit if folks have questions. I’ll go through some of the chats and answer those. Otherwise, you can feel free to unmute and ask anything top of mind. So, Attendee, you have a question. What do they mean, semantic layer, specifically for AI? Yeah, so, Think of it as, how do you codify, for example, a common… a commonly aligned, what’s called, like, metric definition? Like, let’s say if someone says, hey, what is our churn rate? Like, you know, you have a billion different ways to define that, right?
Like, you know, a churn could mean a user who no longer does something, you know, X days later, or it could mean, you know, like, the percentage of sessions that didn’t do X, Y, or Z in some period of time. So, like, you know, like, all those could be anything. And unless you have a codified version of a commonly agreed upon definition of that somewhere. AI itself is also going to get confused. It’s just going to make something up that sounds good to it, but then may not be the right thing for your business. So we’re gonna kind of, like, go through what that looks like, generally speaking.
So it’s not just like, you know, a lot of the AI problems or things that benefit AI also are the fundamental things that we as humans don’t you know, like, could have done much better, but, like, just because now, with AI’s, help, it gets exposed a lot more. So, at my work, for example, we have 4 or 5 different versions of the same, let’s call it, like, same metric, that is used by 4 or 5 different departments across the board. So think of it as You know, like, the marketing folks would have, you know, like, a churn definition differently than the finance department that would have churn definition differently than the product. the product board.
You know, like, a lot of it is, like, nobody’s wrong in any of that. Like, they have their version that is sensible, but it’s just, you know, like. when you come together, everybody gets confused because, hey, the numbers don’t match, but, you know, they sound the same. So a lot of it is actually human problems that is now exposed and amplified by having AI be able to, sort of, like, do a lot of the executions.
Attendee: So, so, quick question on that, follow-up question. So are you then saying you would build a… A wording to define the metric universally.
Hai Guan: Yeah, so at least kind of, like, writing it on paper, like, in a systematic, like, a structured way to do that, that’s kind of, like, where I’m getting at.
Attendee: Okay. So, so at least within your, department or group, at least everyone’s using it the same way, even if The other groups are still using their bonafide semantic layer.
Hai Guan: Yeah, so it’s, it’s more a process thing, so what we’re doing is we are force… We’re forcing the issue of… align… like, grabbing all the decision makers, or the leads at each of these departments together. So think of it as, you know, like, right now, whenever there is… we wanted to make sure a metric gets certified. Certified meaning, like, there’s a commonly agreed upon definition. And so, we get everybody who uses this metric somewhere in their thing, in their day-to-day, together. We identify owner. Like, you know, who should be the ultimate person that owns this metric? And then we get them to agree on one definition, like, very important.
If it’s not one, then you either relabel this thing or fight it out why you should have multiple. And then when everyone signs off that, hey, this is the right way to define this guy, and this person is the owner for it, then my team would help codify that in the semantic layer, and that becomes the golden source of truth for the company. Does that make sense?
Attendee: Okay, okay, so this is sort of the next step, after you would, in… just a couple years ago, or even now, argue over the definition for a… for a dashboard, right? Yeah. And then you would take that Agreed upon definition for the dashboard that everyone is now using, and put it into this semantic layer.
Hai Guan: Yeah, that’s right.
Attendee: Okay, got it. Thank you.
Hai Guan: Alright.
Attendee: your session, I don’t… I don’t have any questions. I think it’s very clear and very useful.
Hai Guan: Cool, alright, thank you. Thanks for… thanks for that. Let’s see, Attendee, you have a question here, and maybe we can drop after this. Isn’t it a losing proposition that different interpretations of the same metric evolve? because of a reason, Attendee, do you want to speak to, sort of, like, what… what you… what you were referring to?
Attendee: Well, just expanding on the previous comment, I was just wondering, the same metrics got interpreted in different ways by different departments, for example, was due to some reason, right? And I’m thinking just codifying it into one agreed-upon, definition for the company may not serve the purpose to different departments, so it might be a losing proposition, is what.
Hai Guan: Got it. Yeah, I think that’s a good point. So, I will tell you… let me tell you the most common root cause that I’ve come across, now that we’re doing this for like, forcefully, is that, maybe 85% of it is because departments just don’t talk to each other. That’s it. Like, they… the reason is they don’t talk to each other, and then we’re being the middleman to be like, hey, you guys have to talk to each other, because… Otherwise, it comes back to us. We’re trying to build, you know, like… like an AI analyst for the company, for example, and the… if humans get confused, the AI is going to get very confused.
And so, the root cause, really, majority of it, is, just departments you know, like, they fixate on, oh, this is how we’ve defined it, and that’s how we define it, too. Again, like, I work for a smallish company, so, like, you know, the processes have not been, you know, like, very mature, for example. But, like, you know, at bigger companies that I’ve been a part of, like, processes and stuff like that, gets a lot more, get a lot more robust from that perspective. But that was kind of, like, the eye-opening thing that, that, that, that I’ve come across, where, you know, it’s really just Getting people to talk.
Attendee: Yeah, I think from one interpretation, in my mind, it’s similar to, say, dialect, right? Or, right, the same word might mean something a little different, depending on the dialect, or… Or, you know, region you’re from, right? So… but when you’re watching TV, everyone uses it the same way. Yeah.
Hai Guan: And then, you know, again, like, some of these things… some of these is just to make sure people have the right… the same understanding of it, and if they disagree that this dialect is not servicing their need, then they talk it out, and we facilitate that, and we’ve… accommodate it. Things like, you know, like, obviously you have a really good reason for introducing a variance of, let’s say, a churn rate, because maybe the coverage for this other supposedly better way of defining this does not have, let’s say, like, historical numbers, and you need it for your purpose. And so we carve out variants for that, specifically, but we don’t call it the same thing.
We make sure there is guardrails around, you know, like, it’s not just churn rate, it could be, you know, churn rate. underscore something, like, like CS customer success, or something like that.
Attendee: Interesting. Oh.
Hai Guan: Cool. But, you know, again, a lot of these is, human and process problems. Nothing technical about any of these, but they’re so, so important because Data Foundation, I’ve never seen a company that solved Data Foundation holistically, and, again, right now, with the use of AI, it just exposes the fact that this becomes even more important, and so… You know, like, in our courses, we talk a lot about, kind of, like, how to How to unify things like that, how to have the proper evals, and how to, you know, like, what are the important steps to, to be able to get to a place where it’s, where, where it’s, there, there’s kind of, like, a common, dictionary, if you will, for… For all these.
Cool, alright. Great. If no other question, you know, like, again, I’ll follow up with emails around all the resources, everything that we’ve touched upon in this, in this, in this session here, and, looking forward for, looking forward to, you know. Seeing you again in our next, sessions, in future sessions, and perhaps in one of our courses.
Attendee: Thank you.
Hai Guan: Alright, no worries. See you guys.