Build an AI Data Analyst in Claude Cowork (No Code)

0:01 So we'll we'll go through basically today how today's going to work. we're going to go through some slides. we'll also do a bit of a demo in co-work. A lot of our sessions in the past are are in cloud code. We'll be in cloud desktop today leveraging co-work primarily. how we'll kind of run through it. So, we're going to talk a little bit about what cloud co-work even is. how it compares to cloud chat and cloud code.

0:37 Then we're going to talk about how we build and use it for data analysis. specifically how we built a AI analyst plug-in for co-work that's derived from our kind of agentic analytics system. we have for cloud code and codex and some of our open source models. I'll walk through the components of those. we'll run some analysis like a quick kind of demo analysis probably just like nine or 10 minutes.

1:13 Then we'll walk through some of the practicalities of applying that at your own job. And then yeah, I'll obviously share the plugin. I'll I'll show you how to set the plugin up as well. It's really easy. but I'll share the plugin with all of you and then I'll give you like a little kind of like homework for you know, tonight or later this week or next week and I I'll give you some guided steps in a follow-up email. so you can kind of try to run this at your own job because we're going to try on some synthetic data today.

1:47 if you have questions as you go along, feel free to drop them in chat. and I or Stravia will will answer them. oh we should do intros actually. I'm Sean. I'm one of the co-randers of AI analyst lab. We do a bunch of these workshops around objective analytics. Been working on data science for about 10 years. And then two colleagues that are the other co-founders that work with me. Hi who is currently on vacation in Europe.

2:20 And then Stravia if you want to give a quick intro. Hello everyone. I'm Ravia. I'm a co-founder along with Sean here. I've been in the field of data science for almost 14 years now. very excited to share to be along with you on this lesson today on cloud co-work. >> Cool. Thanks Ravia. And then also for questions if you want we have a slack community. so invite out. I can get the invite out.

2:55 >> Perfect. Thanks, Raia. if you have if you want to ask questions there later on following the session like after we give out some of like the guided homework I guess you could call it. we're available there anytime. cool. All right. So, couple questions just to get a read of the room before we get into some of the material. first thing I want to know is just kind of like where is everyone at with this stuff? So maybe if you want to type a number in the chat here, type a zero if you've never used cloud code or co-work before. One if you've used one or the other, but not for data analysis. And two, if you've if you use them for data analysis.

3:43 Nice. We got a good spread. Lots of ones and twos, couple zeros. Wherever you are, it doesn't matter. just kind of helps me figure out how we want to tailor the presentation. But yeah, we got a pretty good spread here. What's data analysis? Yeah, that's a good question. You know, I think we have some I think we have some YouTube videos on that. I think YouTube's filled with videos, all right, so next question.

4:16 for those of you who are kind of analyzing data with AI and it doesn't have to be cloud it could be anything right I don't know feel free to drop in chat like how are you how are you doing that you know are you pasting data like directly into chat like Gemini or chat GPT or cloud have you or your team kind of built some systems out at your work. Are you using some enterprise solutions? There's a lot of enterprise solutions that have AI built into them. now nice co-pilot premium with cloud in Excel connected to superbase using co-work to build masterwork. Nice.

5:04 Yeah, paste directly and asking questions. Yep, I think that's a great place to start. building systems using team solutions. Nice. We got a good good spread. Lots of people using AI for analyzing data, which is pretty cool. I think like when we asked this question six months ago, it was like people weren't even really trying to analyze analyze their data at all with with AI. So, this is cool seeing the spread of how people are using it.

5:39 one thing I'll say with like I see a lot of people who are like, "Oh, I'm pasting data directly into a chat and kind of working with it there. this is like I think the best place to start to kind of get a feel with how it works whether it's Claude or Chatbt or Gemini or whatever." what I have found when you're working directly in chat, which is why we're doing a session on co-work today, is a lot of it kind of will read the output I usually get when I use chat is it reads like a summary of the data.

6:10 It doesn't necessarily answer the sort of like decision driven questions that I actually had for it. So, I have to ask a lot of follow-ups and then another one. And after a while I feel like I'm kind of basically doing the analysis myself but like via reprompting it many times. So that brings a lot of time that brings a lot of tokens. it always also carries this risk of like not really having the visibility of what's going on underneath or easily being kind of like steered down a wrong path. So that's why we're kind of working with co-work today, why we work a lot with cloud code.

6:49 So, in the next slide here, I'm going to talk a little bit about like what the difference is between chat, co-work, and code. but before we get into that, just something for you to kind of like noodle on as we go through, you can, you know, write this down or think about it or it's going to come back later on when we talk about kind of like the homework. Feel free to drop it in chat, too, if you want. but you know, think about some real questions about your own data, your own actual work or personal use cases that you'd want answered this week, like which customers kind of stopped using as much last month or turned last month? What drove a drop in acquisition like new signups? Where are support tickets piling up? Like where are people kind of complaining about our product or our business? write that down. Drop it in chat if you want. You don't have to, but definitely start thinking as we go through this hour about what a question might be around your own business product or data as the the end of this. We're going to kind of talk about how you can actually answer that with co-workin.

7:58 Yeah, looking for for trends and and sees that survey. It's a really good one. doing that sort of like qualitative research analysis stuff. it's like so much easier to digest that sort of survey information than it ever was before. Yeah, customer segmentation is another really good one. so these kind of question cloud gives you three ways to kind of work on questions like these and they're sort of built for different workflows I would say. chat I think of as like thinking with claude. So you ask you draft, you brainstorm, it responds, but you kind of drive every turn by reprompting like you know you're chatting with it, right? co-work is delegating now moving from like chatting with it, working with it to sort of delegating to Claude. So it's like before your colleagues working together or maybe you it's like you have like an intern and outcome to have.

9:30 Hey Sean. >> and code is unable to hear. There was a lag and you your voice was breaking. >> Did I lag? >> Okay, I got a I had a little signal that said my internet was breaking up. so where was I? I think I I was talking about co-working to Claude. So you point at your folders, your tools. it's going to just you're going to describe an outcome. It's going to plan the work and run it and hand you back finished files, but it's like kind of doing those multi-turns in a single workflow rather than having to go back and forth reprompting like you would with chat.

10:08 and then code is is kind of building the systems with claude. the same sort of idea as co-work but you're typically in a terminal you're working in a repo with your whole has access to your whole machine behind it. It can take over your entire machine if you want it and you give it permission to everything on your computer every tool you have installed. So I think a lot of this if I think about like chat versus say like co-work a lot of it comes down to a primary difference in that chat cloud can read what you upload but it can't you know go search on its own it can't save anything back to your computer and co-work you know it reads your files and writes real files back. So that's the difference. That's the difference is like why co-work can hand you a finished deliverable instead of this kind of like back and forth wall of text. we're not going to go super deep into cloud code today. We have a bunch of sessions about that in the past if you check out our our YouTube workshops. We have many more coming as well. but I like co-work is like a nice I'd say kind of middle ground to getting into working with like building systems, delegating to systems in a much more sort of like userfriendly approachable way. not that cloud code is like terribly unapproachable, but can be a little intimidating if you haven't worked in in terminal before. I say some some primary difference between co-work and code just to highlight them.

11:48 before we get into it though cloud code can do a lot of stuff that co-work can't. so co-work runs and then kind of isolated workspace a sandbox. cloud code runs on your actual computer with all your installed tools your Python your credentials your whole environment. it gets you like version control get branches whereas like co-work projects are are more or less just folders. co-work in terms of data connections like co-work connects to can connect to cloud tools and you know data warehouses through connectors that are developed and and provided by anthropic. So you can hook co-work up to like your companies like snowflake or bigquery data warehouse via those connectors. but there are limits to like what connectors exist if those companies have created connectors. Whereas with cloud code you can basically connect to any sort of database or data warehouse even if it's like behind some VPN of yours like you would previously just writing Python. and then code can be like very script scripted and automated. you can have like a fully functional kind of endto-end pipeline with code co-work you can run scheduled tasks like through their UI but it's just not as as flexible as cloud code but today we'll talk about most about co-work if you want to learn more about cloud code yeah you can check out our other other workshops or our course okay so just to give just to get an idea so we're all kind of on the same page of kind of what does what, I've got a few scenarios for you here. You can guess at these. I'm you know I'll give the answers pretty fast, but I'd love to know what you think A, B, and or C.

13:53 which tool which you know version of of claude I guess chat co-worker code you would reach for you would try to leverage in terms of each of these scenarios and you can actually use more than one of these in most cases. So, first one, say you're talking to Claude. You say, "Pull together a one-page brief on our Quran results. Numbers are in a spreadsheet on my desktop. I want the briefs saved in a reports folder." Got a lot of B's and oh, we got a BC. We got a couple B and C's. Nice. Yeah, the B is probably the this you guys are correct. B is the I would say the lightest way most approachable way to do it. You could do this with co-work but you can also do it with cloud code obviously. So B and C are the correct. This is not something you would do with chat though.

14:48 I would say these these questions are more what can you not which one can you not do them with and which ones you can't because there's multiple answers. second one. So, let's say, you know, we are talking to Claude and we say, "Look at the customer feedback in our Slack channel this week. What themes should I pay attention to?" I love this kind of use case of having it come through Slack because it's just like doing it yourself is so information overload.

15:16 A B and or C. Let's see. We got a lot of C's. We got a B. Oh, see Andrea, we got an A. So, you can do this with any of them actually. You can use A, B, or C. you can use the connectors from cloud to connect directly to Slack via chat and via co-work and you can also use MCPS APIs to connect in cloud code too. So, you can actually use use any of these to to answer question like this, which is really nice.

15:52 Obviously, if which is better. so like if you're if you're do if you have like very specific kind of workflow you have around like the type of themes to pull out or how you want it to address like what you should pay attention to like who's important to you, what topics are important to you, then B or C is going to be like a better way to do it because you can set up kind of customized workflows whereas chat is going to be kind of like picking on its own a bit.

16:26 >> >> B is probably the best like most approachable way to do it. yeah, exactly, Brandon. More context and structure. but anything, you know, it's kind of like I feel like start with the most approachable way and then you can go more regress and robust as you proceed. and then last one, say we have something that's rebuild our nightly metrics job. So, it reruns the numbers and commits an updated report to our team's repo.

17:02 We got some C's. Grab your B's. You might be able to rig something up with codework in this. I would probably use cloud code for this. anything where I really want to have something kind of that is truly on some like automated job that is like committing and working on its own updating like a repo or a shared resource. I usually will leverage cloud code for that because it's kind of like writing scripts, right?

17:37 Cool. Well, now you guys know a little bit about how you would use one versus the other. Oh, great question. Devitra, would all these work only on a pro account? So, yeah, co-work. You're going to need a pro account. you can use this stuff with chat without a pro account, but you'll want a pro account for for using co-orker code. Okay, so we talked a little bit about the difference between chat co-work. Now I want to talk some more about like how we leverage this for data analysis.

18:18 So what is an AI data analyst and we have a lot of sessions on this in our past YouTube workshops that you can check out as well but at a high level I'll go through a couple slides. So it's not the model itself. it's model agnostic the whole setup. So you can run it with whichever cloud model you want, right? You can run it with sonnet or opus or fable. You're probably not going to run it with sonnet. You're probably run it with opus and you can run it with any kind of model version of opus. we run a lot of our analyses and we run it today with opus 4.6. I'll show you how to switch models in cloud co-work. This is the model that came out in midFebruary.

19:03 and we actually find for data analysis, this is probably this one meets the threshold for most data analysis. Basically models beyond this we find are more expensive in terms of tokens, more expensive in terms of time they take to process things and they also kind of go off the rails a bit more. So they are improved for I would say coding tasks. if I were to like build a system of analysis, I'd probably use Fable or Opus 5. But to run the actual analysis that that system has been built for, I usually use like Opus 4.6. Yeah. So, yeah. So, yeah, Brandon, I use Fable for a lot of like the building itself.

19:48 so quick kind of version of this if I were to think of like a very high level parts that comprise AI analyst. you have there's a couple versions of this. We have this plugin that we're going to share today from co-work. we have a a fuller repo that's also open source that you can use with cloud code. so we'll talk about kind of simplified version. Part one of your kind of AI analyst plugin is going to be basically like instructions. so this is standing context that applies every session you open cloud with this plugin that says like you know who you are, what your company does, how you like your output delivered. Without it, every session would have to start from zero. So, you want to set up that sort of memory in a way of of who you are. Part two is skills. These are basically playbooks for a type of work, one kind of work. So, a file that says when you get a task like this, here's the process, here's the format, here's what good looks like. So, it's how the analysis runs your way instead of a very generic way without you having to reexlain it every session. And that claude 101 session we have that I shared in the chat in the beginning of session goes really deep into like what skills are and how to build them. And then part three, what's there like connectors? You know, this is all kind of like a a mini version of of what's in there, but that's like the data connection. today we're just going to look through a folder of CSVs, but in your world you could be you could be connecting it to like even a Google Drive or or one of your data warehouses like Snowflake or BigQuery or whatever.

21:47 And then the thing that kind of wraps these all together for co-work is a plugin. So plugin is basically you can just think of it as a package. So it has skills plus connectors bundled. So it all installs really easily into cloud co-work. So you have to like add some one one thing at a time. our full system has a lot more of this has a bunch more kind of agents, pipeline workflows, helper function scripts, hooks and like context and knowledge layers. we've put a lot of that into the plugin itself. but we'll also talk about those later that you're going to leverage cloud code if you wanted to use the full system.

22:28 So how it works if you were to use this kind of in practice again kind of simplified version here basically you have your your question comes in right your input before anything runs we want the analyst to load two things the instructions so the standing context about who you are what the company is and then a context store so a folder of files holding your like metric definitions, notes about your data sets, any corrections or mistakes from past runs. Then after it loads that, skills fire and they fire in like not a not a a totally predefined order. Claude will reason through which skills to fire when. You can also tell it directly when to fire one one over another, but some of them will always come up in in certain orders. So, for instance, it will always first try to frame your question you're asking into an actual decision that's trying to be made. It'll then profile the data. It's really important for it to do like a little bit get a little bit of understanding of the context of the data and then it'll get into the analysis itself. That's where it can kind of start going off many different directions based off the question it framed in the data profile and then it'll fire off some kind of validation checks on its own and then finally after all that only after all that will provide you your output whether that's charts or a document or a deck.

24:05 so all of that happens in co-works sandbox reading the folder you pointed at or the data you pointed at and out the end comes like some files refer a chart save back to your folder and then you can you can set it up in such a way where if it makes mistakes as it inevitably will it can log those and so the next run those corrections will be baked into the So, I'm going to send out a I'll actually jump over to Claude Co-work right now. I'm going to send out a kind of step-by-step guide of how to do this, but let me just show you how to add a plugin to co-work rather than talking through it in this slide.

24:57 So right here I'm in cloud desktop. you know you can see I can switch between chat and co-work here. Notice once I switch over to co-work it gives me the option to selecting folders that I can work in. So, to add a plugin, and I'll actually go ahead. I'm going to share this in the email with the full step by step later on, but I'll go ahead and add the plug-in link, here in the chat.

25:40 It's pretty easy. You just basically go to settings and then you're going to scroll down on this left bar to customize. And then you'll go to plugins. I already have the plugin installed, the AI analyst plus plugin, but you won't have this here. So you'll go to add. You go to add a marketplace and then add from a repository.

26:14 And then you're going to just paste in that that URL I just put in the chat. And then you can have it sync automatically if you want. You'll have to give cloud permission to use your GitHub user in order to do this. You could also turn it off if you don't want to give it that permission. And then it's going to tell me I've already got this uploaded since the marketplace is already added. But in your case, you'll add that and then it'll pop up at the marketplace and you'll just click a button that says install.

26:55 Once you've done that, yeah, the AI Analyst Plus has been installed and you can kind of look in this if you want. You can also see this in the GitHub repo itself, but you'll see all the skills that are available in the plugin. as well as all of the agents. So, pretty easy. Close out of this.

27:27 And then I'll say a couple other things for using it. One other thing I do sometime, what I personally do, you don't have to do this. So, one thing with co-work that's a bit different than say if you're using cloud code in like a VS code terminal or some other terminal, Anthropic really really likes to force their their their their own skills and plugins on you and they do have some data analysis skills that are okay.

28:02 so what I I do is I actually go into co-work in settings and I go to the bottom and this kind of works though kind of skips it too and I edit these global instructions and I say hey for data analysis tasks use AI analysis plus skills and always start from this decision framing step. the reason I do this is because I find that Anthropics builtin analysis thing, it just like goes off on its own and starts analyzing data and trying to answer questions that you didn't ask about. So, it'll give you some analysis back, but it can be pretty irrelevant.

28:43 Doesn't always work. I'll show you later. actually when I when I leverage the plugin here what I'm actually going to do is I'm going to call it by doing slashanalyst. Come on. And you can see this skill pops up that's within the plugin itself that says start analysis with AI analyst plus way framing the decision first and running the method end to end. So, this basically forces the use of the plug-in and kind of overrides Coowwork's baked in analysis skills and yeah, DT will send out the recording.

29:25 Okay. So, what is in the plugin? We'll do a pretty quick kind of walkthrough of this. a lot of stuff that's in the plug-in is the same stuff that's in our AI analyst plus like full-on repo which we've done a bunch of sessions about. Like almost all of our workshops are usually about one aspect, small aspect or one skill or one agent workflow. so obviously way too much to go through in just an hour. but at a high level I'll point out a few things. So framing the question, we have a bunch of skills and workflows around this baked into the plug-in.

30:04 so we go through like a question, it'll go bring you through a question layer of like and when you try and do analysis, what is your goal? What's the decision you're trying to make? What's the metric you're trying to look at? some forming hypotheses around it. You'll see this working in a few minutes here as we do it. top right's knowing your data. So everything from connecting data via different setup wizards we've developed for different warehouses which is probably what you'll run later on as you won't necessarily be running this on CSVs but connecting to your actual company's data warehouse to doing like deep statistical profiles of like understanding like what are the tables what's the data quality look like you know basically these are things that if you were an analyst yourself doing data analysis.

30:57 They're like the first two things you would want to do to run through an analysis before you actually analyze the data. Then we have a huge a huge catalog of skills around the analysis itself that are basically dictated by the question framing and data profiling. it even gets into a lot of correlative stuff, a lot of kind of root cause segmentation stuff, but also gets into like experimentation.

31:28 a lot of things around designing AP tests there. And then bottom right is trusting the output. So one of the kind of most important parts of this is not just getting numbers back, but knowing how reliable they are, if you can trust them. So we have built in a few skills that are kind of like a very simplified version of evaluation suite. So you can understand like hey what is this number trace back to in terms of a source. So it's not just like a random number I'm seeing. I know exactly where it was pulled from and why. And then how reliable is the system in deriving this answer. So if I ran this question five times, am I going to get five different answers or is this answer going to hold across the analysis?

32:21 And then similarly we have some skills around like memory so logging corrections so the same mistake isn't made twice every run. and reusing like a known SQL that works so it doesn't have to invent it each time. and then output standards and delivery. These are pretty similar. So having customized standards for how the story of the analysis is told, how it is visualized, how it is, kind of voiced and documented in a slide or a chart or a doc depending on the stakeholder that's reading it.

33:07 And then there's some there's some kind of extra skills and agents in here too that specialize around causal inference too. Those are a bit more experimental. Okay. So let's get into the demo. I'll say something I've noticed with co-work is and I say cloud of code this this happens. feel like co-work is a bit more likely to kind of like chat like just like have a mind of its own and go off in a direction that I don't necessarily want it to. It'll if you ask it ask it a question, it's going to form its own opinions around like why you're asking that question and it'll answer that with probably a correct number but not necessarily a useful number. And so when I'm analyzing data in co-work, I really try to be as detailed as possible. And we actually have added in gates into the plugin. So if you're not detailed and you're using our plugin rather than the built-in kind of anthropic data analysis, it's going to stop you and it's going to start basically interviewing you to understand like why are you actually asking this question. In this case, we're just going to give it the answers kind of up front and be as detailed as possible. If you do that, it'll read through it and it'll know that you can skip the gate. But, just good good way to frame any sort of prompt here, as you're working with it. And honestly, this is good even if you're not using AI for analysis.

34:53 This is how I would like approach analysis or or approach a really good question framing is you know name the deliverable you want. name the inputs. So like what data right so like f which folder files or what date range you want to leverage add any nuance the details like that you know about it that aren't already baked into the context of the system. and then name the decision that the answer is going to support. So the decision sentence is really important because it anchors the analyst on what type of analysis to do. Without it, it has to guess what you're going to use the answer for and you'll get a very kind like generic version with it. The analysis can actually like point somewhere and be something that's actionable.

35:46 All right, so let's pop over to co-work and it's gonna come on exit full screen. it's going to take a little while to run. So, while it does that, we can kind of answer some questions. All right. So, I'm back in co-work here. quick kind of show of how this works again. So you can switch from chat to co-work. Once I've switched to co-work, I can now tell it a folder to use. So I'm going to point to this folder I have on my desktop called Novart demo. And it just has a few CSVs of data like the products, our order items, and like orders. So you can think of this as like a e-commerce Amazon kind of synthetic data set. So I'm going to open this. It's going to ask me if it's allowed to change files.

36:49 I'm going to say allow or I'll say always allow. This will allow it to write to the folder. If you don't want to have to say allow or deni or deny like approval over time, you can change to automatically or skip all approvals or you can keep it on manually approve. You probably want to stay on manually approved in the beginning just so you get a feel for what it's asking you to do. And then you can switch your model over here. So you can switch the family of model, right? Fable is like the most sophisticated. Haiku is the simplest.

37:25 You're going to be wanting to use like opus or higher. it's going to default you to opus 5. we I usually if I go to more models, I go back to opus 4.6. And then I change the effort to high. So this I think is is quite sufficient for doing analysis tasks. And then I'm just going to pop in a prompt I have saved over here. Once we get it running, I can read it out to you guys and tell you why we wrote this. All right. So, the first thing I'm going to do is I'm going to use the analyst skill. I can call a skill by doing slash and a skill. And you'll see that list pop up. This is going to ensure like I said earlier, I'm going to allow it to access this folder again. this is going to ensure that it is using the AI analyst plus skills. It might still not use a miss.

38:32 It might still try and pull some of the built-in co-work ones. So, we'll see. But the question I'm asking is look at the Novaart sales data in this folder. I'm telling it right if we think about th those things we want to tell it, we want to tell it deliverable inputs nuance decision. So, I'm selling it the inputs right now. Look at the Novart sales data using the orders, order items, and product files.

39:05 I'm going to give it some nuance. I ran merchandising. I need to decide which product categories get more marketing budget for next quarter. There's my decision, what it serves for. I'm going to give it like what I want back. Compare revenue for the most recent full quarters in the data. Break it down by category. Tell me which categories drove it. and then the output, right? give me a onepage brief and just give me one chart I can put I can give to my team and then if any of the data looks off, flag it.

39:39 Don't just kind of like breeze through it. So, this is a pretty specific you can kind of try different levels of specificity. You don't necessarily have to be this detailed. I was very detailed in telling it what analysis to actually do. You can let it be a little more open-ended. If you're too open-ended, what will happen is basically it will stop running and it'll give you like kind of like a popup in the chat and it'll ask you some interview questions and it'll have options predefined options that you could answer as well as like a open kind of chat box.

40:25 yeah so I'm using local folders. you could operate this in Google drive or one drive directly. no real issues there. I I end up using cloud code a lot when I do that but you could connect you could connect go to your Google drive and operate there as well. So, it is going to run through this. I think I don't know how long it take like 8 to 10 minutes. but I don't know like you guys kind of see the analysis here. How long do you think this would usually take if you were running it yourself?

41:06 like how many minutes or hours or days if you're trying to do this analysis on your own versus trying to just have this output by Claude right now. Feel free to drop that in the chat. I think this would take me probably longer than 10 minutes. so you can also follow along with what it's doing. So basically what co did it created a plan first off would not do it at all. Yeah. There's so much stuff you would just not do at all.

41:45 but that now you can kind of undo which is like also a little dangerous because it's like when some of the I feel like the friction of how long analysis took before added like a point for us to like actually have to be thoughtful as humans of like hey is doing this worth it in terms of is this going to be something useful for the company whereas now when a lot of that friction is removed I do think that people are doing more analysis that's not necessarily valuable for the company. So, you kind of have to add this like filtering mechanism in yourself. But there's also a lot of stuff that would be helpful to customers that you would just not do before because you didn't have time that you can now. but you can see it has created like basically a plan of what it's going to do. It's going to profile the data files. It's going to identify the two most recent quarters. It already did these first two things. Then it's going to build a revenue chart. it's gonna write brief and then it's going to do some validation. So, just that workflow that I kind of mentioned before in the slides. it actually went ahead and didn't use our skills, which is annoying. So, you can see what skills it's pulling from when it does everything. So, you know, why it's doing the workflow it's doing. So if I look at analyst, this is that first skill I said that said please use the analyst plus stuff says where it came from. Same thing with analyst core. This is basically the core workflow it works through. but then it also used data viz. Datav is anthropics data visualization skill which I would say is like okay. It's pretty like it's kind of like getting your basic mattplot lab plots. You can do a lot more customization around data viz to follow kind of like best practices of storytelling with data. I will probably ask it after this runs to rerun the chart with our skills. But you can kind of see even you can go even like a layer deep deeper and you can see like exactly what it's doing. so it's hiding this in the background. But this is one thing I do think gets presented better than code cloud code presents it. You can you can kind of look at this in terminal in cloud code if you hit command O. but I do like how co-work presents it. So you can see it's thinking to itself around like what I'm asking. you can see at where it's pulling the data from and any things it's running against. So like it doesn't for instance the like knowledge folder doesn't exist yet because we don't I haven't added context about our our company or business to this yet. So it's going to have to infer some of that itself. It's finding some insights on its own.

44:53 yeah, I do like how you can kind of go through all this and then at the end it has delivered us our chart and our readout as we asked. So, I'll go through the readout, but here's the chart it created. Actually, it's it's okay. Actually, it's not bad. So, category revenue Q3 2024 versus Q4 2024. but it's not really like clear around like what the insight is from this, which is what we've kind of built our data viz skills around. So, I'm going to say like, "Hey, you used the built-in data viz skill.

45:35 I want you to use the data viz practices from AI analyst plus. And let's get it to rebuild the chart for us. Yeah, Frank, you can ask you can ask it for kind of recommendations on model and effort choice. we have done a little bit of this in cloud code where we've set up like routing for certain parts of the workflow to use simpler, faster models. for instance like labeling I think someone mentioned like going through survey data earlier like labeling survey data you could probably use something like sonnet for that it's a lot faster and cheaper compared to like something here where you saw how to reason through a bunch of different steps you want to be like on opus for that for knowledge I you can do it either way. I think like at the project level is probably best because at some point you're going to have like a huge like context bloat if you have like some massive global things, but you could have some stuff that's global, some stuff that's project based.

47:01 So let's read just I want to I know we only have 10 minutes left. I'll stay over a little bit for questions, but you can see it's giving us our brief here. It has kind of the headline. So, says like the overall revenue grew 41%. every category grew but not equally. There are some that grew 51%, some that grow 36%. it highlighted the standout categories, clothing, books, electronics. I'm not going to read through all this, but you're welcome to run this yourself later on. As you saw, it's pretty quick. And then it's going to give us some suggestions for the budget. So, in this case, it's telling us to invest behind kind of the faster growing categories like books and clothing that are growing above a average. but it also is going to like ask us to diagnose some of the some of the slower growing ones like electronics is growing relatively slow to other categories but before you know it's not just going to tell us to increase or cut that budget. It's going to recommend follow-up analysis to understand why that's happening. So in most cases if it's not just going to like tell you do this thing if it doesn't know why it's doing it and then also pulled in some data flags. So one is the re revenue definition could be incorrect. So we don't have any sort of context layer built into this yet.

48:44 and then it's currently excluding returns and cancellations, though we may not want to do that. So, it's telling us some of the assumptions it's made. And then we can see here it did upgrade our chart. very different than the data viz one that I made. A lot of the charts that we try to use with our best practices is like sharing like the insight in the chart exactly. So you don't have to like kind of analyze the chart yourself. The chart's going to tell the story for you. So in this case, it's just like we saw and the readout shows us which categories are outpacing the rest.

49:31 All right. So yeah, I mean like any anything's going to be able to do some sort of analysis like this. Gemini notebook can do this. You can do this in a chat. You could do this without adding a plugin. co-work allows you to have create your own plugins. So you can like customize the skills and workflows specific to your expertise as a analyst or your expertise in the business. Gemini notebook may allow you to do that as well.

50:06 so yeah, not saying you use one or the other, but this is just how how to use co-work. if you use that, I honestly don't use co-work that much. I I typically use cloud code or codeex for for my work. Sometimes it sucks staring at that terminal so much. It's like hurts your head. I do like the chat on your face a lot more. okay, four minutes left. I will stay over for questions though. we are basically done.

50:37 okay. All that being said, it gave us a brief, it gave us some, visualizations. You don't necessarily just want to trust that output right off the bat. these are a few checks that are kind of really easy checks that you don't need to write any code to do that I kind of will do to like validate numbers. this the same kind of checks I would do if I wasn't using AI. But now when you're doing AI and all this like execution layer has been done for you.

51:09 I think before you're kind of probably doing these checks as you go through the analysis itself that you're running it yourself. And now it's just like this is a way to make sure that the number is like somewhat valid. If it fails any of these, you know that you've got an issue going on. But there's much more robust checks you can you can run on this stuff. So before I share out any results, first thing I really do is I want to trace trace the numbers back to their sources. You can ask Claude ask Claude to do this.

51:44 We've built in like this provenence skills that'll help you trace and record where it's being pulled from and how. so it actually it actually flagged like for instance like hey I I had an assumption that revenue is calculated in this way that might not be correct. I think the the first if you just wanted to like go right off the bat, take like the kind of like biggest number, not in terms of like the magnitude of the number, but like the kind of most important number that the kind of most of the decision is kind of leaning on and ask like where did this come from? Show me should point you back at the files, the data set, the query. two, yeah, you could ask another LM to check. We do that in our fiveweek course. We have like codecs check a bunch of clouds work. We have open source models check bunch of clouds work. So doing that sort of like triangulation. There's like a couple methods to this actually. not to get too far out of on a tangent from co-working to cloud code. but you can have other models check single models work or you can have multiple models run the same analysis in parallel and see if they triangulate to the same number within some sub some interval. So that's a really nice way to validate Frank.

53:06 two sum the parts like this is like a very basic check that any analyst is going to do regardless of AI but make sure all the category numbers add up to the total. If they don't, something's obviously incorrect there. maybe some like weird join that fanned things out or filtered things out. and then three, tie the number back to something you already know. so this is more like does does this logically look right? If I know these numbers from last quarter from some presentation I did or someone else did, are these still within like the same kind of realm of magnitude? So those are very manual, easy checks to do. realistically you probably want to build out a whole eval suite for your analyst.

53:54 so you can know like you know like how do I know this is right? what happens if say like definitions change that goes pretty beyond co-work co-work's kind of like leveraging I would say a tool to do analysis whereas that goes into like building a system of like an analyst that like checks its own work and and systematically and automatically updates as changes are made in the business. So if someone updates like their definition in a meeting for a certain metric, co-work's not going to catch that. But if you build a system where you have a shared context layer that's hosted in GitHub or some sort of like metric dictionary software you can have code pull from that directly.

54:43 We teach that in our fiveweek course. So we have a fiveweek course coming up on September 8th. it's called agentic analysts build your AI analyst. we teach it end to end. We we work primarily in cloud code but we also go in codeex. We also use open source models to build we'll go basically take this from tool to to building your own system for your for your company. So week one is around building AI analyst you know your first skills learning how to design your analyst system. Week two is around expanding so connecting to your data developing a production system kind of like the way you'd build with a real team. Week three is around eval. So validating your output and scoring them against like known questions and answers so you can understand the accuracy of your analytic system over time. week four is around context. So it's like your metric definition, your business knowledge, your semantic layers encoded so corrections will will stick to the system and get better every week. And then week five is around kind of like some of the the questions we got here like running this with different models not just with with the cloud models but with like codeex models or with open source models.

56:07 So if today's interesting that's kind of like the deeper end of it. so for tonight, or this week or next week, what I'll do, I'm going to send out, kind of step-by-step tutorial a bit about what we went through here today, but how you can do it for, your own work. but I would say like, you know, if you don't have cloud desktop installed, you'll want to install that. You can add the analyst from the link that we'll send out. Point it at your folder or your real with a real export of your data or point at your data warehouse.

56:45 There's some setup wizards in there that we'll talk about in the guide. And then try and use that same shape I used to answer that question you thought about earlier in the session today. You know, think about deliverable, inputs, nuance, and decision. we'll we'll give you the template for that. kind of that same kind of exact prompt that I learned earlier and you'll fill in the blanks with with your use case. So, everyone signed up here, we'll we'll send that email out to you with the guide and the plugin, and if you have any questions, you can you can drop them in our our Slack our Slack channel, which we'll add in that email as well.

57:28 Okay. Oh, yeah. If you want to join our fiveweek course for building, you get 20% off. You use code build 20 or you can hit this QR code and that'll al that'll get you the that'll get you the the signup page with the promo already built in. But I can stay sorry I went a little bit over today. There's just like so much stuff to go through whenever we have these. And I know I missed a lot of questions, but maybe I can stay over 10 to 15 minutes to like 11:15 Pacific.

58:00 and answer some questions. Sarabia, any questions in here that you saw come up? >> I think you try to answer most of them through while you're looking through, but it'd be great to take some live questions, too. Yeah. >> Yeah. If anyone has some live ones, feel free to drop them in the chat again if I lost them earlier. Have you connected it to revenue outcomes of the recommendations? Yeah.

58:32 So, one of the things you know this is a fake synthetic data set. So, the recommendations in this analysis I walked to it's going to have that. Yeah. One thing we do is we whether you're working with like a AI analyst or just doing analysis yourself any sort of recommendation that is made you want to see what was the actual kind of result of executing and applying it at the business a couple months later. So I mean it's it's it's not really like AI specific that's just like data science and analyst specific. We do it is nice with the cloud code. So we have an our we have another course called AI analytics for everyone or week five we kind of talk about how do you follow up with the business with with what actually came in terms of the outcomes a couple months later. How do you kind of automate u reminders of of looking into that? How do you measure them causally and remove any of the other kind of like confounding variables that could have affected the business during that time? But I think it is really important to to follow up later.

59:54 What times did the course take place? Yeah. So let me pull up the page. If you go to this link, there should be a sy there has the times. Let me confirm is mornings. Basically, how it's is it's it is I think two live sessions each week. got check which days they're on.

60:31 in the morning. So, should be fine in ter so be late afternoon evening if you're in the UK. the sessions the live sessions are two hours. So everything in this course the Gent analytics one is live because here we go. Let me try and get this shared be So I think you can go down to >> You froze for a second, Sean.

61:06 Oh, I was just saying with our agentic analytics course everything is live not async content because it is kind of like developing so fast that after every single session we change the course and update it with anything that's happened new in industry or any feedback from our students. We take it really seriously. that being said, everything's also recorded. And so the recorded the recordings are available usually two hours after the session that day runs. It'll be in the student portal, so you can watch it on your own time. if you go to that course page, you can see the dates and times.

61:48 so yeah, they're running from 7 to 9:00 a.m. Tuesdays and Fridays is when we typically do the kind of live lecture sections and then we will have at least one if not two optional office hours on Wednesdays or Tuesdays and Wednesdays each week from 8:00 to 9:00 Pacific. So after each lecture, the lectures are very hands-on. We do a lot of like building in in the actual lecture sessions themselves. It's not just like looking at slides or demos like these like the workshop today was. So it's a lot more interactive. And then we also have like kind of optional exercises homework after each lecture that we can talk about in the office hours. Or if you go and want to just watch the recording of the lecture on your own time and ask questions in the office hours, that works too.

62:51 In terms of weekend course options, we used to run weekend boot camps. we are testing out weekday ones right now. So, we don't have a weekend option for this at the moment. Might bring those back in like November, December based on how this goes, but you could, you could obviously watch the recordings on the weekends if you want. There's just so much content to go through. It's I don't know, it's like four to five hours a week of stuff we want to get through when you include like the hands-on exercise stuff. So, it would have to be multiple weekends in a row, which is hard to get people to to commit to, but we might bring that back in the future.

63:47 structured and unstructured data. Yeah, I think it's really good at turning unstructured data into structured data. So like unstructured data is so hard to analyze, right? I think you can work directly with unstructured data. So like someone talked about surveys. I've done a lot of analysis on like customer calls. and so you can it can analyze that. But at some point like depending on how much unstructured data you have it can get like a bit bit of like context bloat and start anchoring to a few pieces of the unstructured data or a few kind of insights it has. What I find works better is if I have Claude help me develop like a structured format of that unstructured data. So basically turning that into rows and columns and analyzing it that way, which would have taken would have been really hard before, right? You would have had to have humans going through and labeling stuff.

64:54 You do want to have humans label some of that stuff because you want to like basically assess the accuracy of turning that unstructured data in structured data. But it's pretty good at that. Ja in terms of how would I pick between the two courses AI analysts for everyone versus this this other one aentic analytics. Yeah good question. So this course here aentic analytics building an AI analyst this is about building an agentic analytics system building it on your own from scratch as well as developing on ours learning how to evaluate and validate the output. So it's basically also like a crash course and AI evals talks a lot about understanding how to build engineer and manage context as well both for you individually as well as share across organization and then it talks a bit about like different types of models you can leverage and and how you like kind of switch the model between each of your analysis and then evaluate whether that model change or or even like the context management engineering change actually affects your system. So it's all about building agentic systems to do analysis for you where we think a lot of the data science analytics profession is headed over the next several years.

66:16 a lot of the learnings here can be also just transferred over to like you're not necessarily building an agentic an analyst but maybe you're building some other sort of agentic workflow. So I used to work for legal tech where we were building like agentic lawyer workflows and so a lot of the learnings here you can kind of cookie cutter over to that. our other course AI analytics for everyone that's more about analytical thinking that is more around you know there's this just because we can now do analysis with these tools we don't necessarily know that we're leveraging them in a way where they're actually actionable where they're providing value for the company we don't always know what they're doing under the hood like what type of analysis they're doing so that course is more like a crash course in data science and analytics. So, think of it as like a very practical applied real world version of like a data science master's degree. but all of the execution layers of like that analytical thinking and different types of analysis and statistical analysis and also just like kind of regular coral analysis is executed in cloud code or codeex rather than writing a python and SQL yourself. So different kind of courses one is more around like hey I want to build an agentic analytic system. The other one is I want to do analysis.

67:48 I want to learn how to do think analytically and know what's going on under the hood but execute it with AI. Man, we don't have any courses for FPNA professionals, but we have had several FPNA people come through both of our courses. both at the kind of like financial analyst level to like we've had multiple CFOs come and take our course too. so we actually want to create a course for this, but I don't know enough about finance to do that. So like I'm kind of scared because the we need to get we need to get an expert in and finance basically on our team before we develop that course. like you know the the whole point about the course of building the analyst about how do you build context in how do you build ewells in that's literally the domain expertise that you get from the FPNA space and once you learn that J I think sorry Gino you would basically be you will have all the tools to take whatever we are trying to share with you to build that for your use cases >> yeah I think these I think you would have the skill exactly what MRI has so I think the jetic analytics of course is still is great for FPNA professionals but it's not necessarily going to be in that domain. So the use case we use is like a ecommerce more like product analyst view but we can talk about FPNA use cases in any of the office hours and talk about how we would totally change this for for like the kind of tools or workflows you use as well.

69:30 That's what honestly our office hours usually work end up being like the office hours end up being less about like hey let's review this content and more around like hey in my job or in my personal project I'm doing XYZ can you basically give me guidance on how I would do that and it's not just us giving guidance it's like all of the all of the other kind of builders who are taking the course everyone's that's like what we really like about the live sessions is especially in something like this that's very new right there's not necessarily a textbook on how to build genic analyics I'm sure someone has written a textbook and it's probably already out of date but it's nice to have everyone like talk about what they're doing in their actual job all right let's see if there's any other questions in here.

70:29 I think I got most of these. >> I think there's a question now from Ruben. Is there anything for product management? looking for ways to do complex competitive analysis and benchmarking. >> Yeah, I'd say probably anox for everyone would be a good course on this. we do I mean there's a there's basically yeah I would say every single week of that's pretty relevant to primary for everyone is like our course of like how do you how do we teach people how to do like your cost analysis comparative analysis causal analysis experiments using AI tools primary clot code it's going to be leveraging it for the analysis itself rather than building a system of analysis. So that might be a big a good course for you.

71:26 we did build it with like kind of like product managers in mind though although eventually we're like actually anyone could leverage this course. What else? That sounds like a good lightning lesson too. Maybe we should do something like that in the future. Also like a competitive analysis lightning lesson would be cool. Always looking for ideas for other free workshops we can do.

72:02 nice. Okay, Ruben, maybe like if we're not already connected on LinkedIn if you want to ping me. that'd be fun. We right now we have our workshops burst through I think end of October but yeah we could fit that in give us some time to work on it. All right I think I think we've answered most of the questions here and a lot of people have left although we still have 30 of you hanging around. Thanks for staying over 18 minutes. thank you for all the questions and the engagement and answering all of my annoying polls along the way. I will send out that kind of step-by-step guide of how to get the plug-in set up and try on your own data later today after I walk my dog. And then the recording should automatically send out within 48 hours from Maven.

73:06 But I might have it on YouTube up later today as well before that. If I do have it up earlier, I'll drop that in I'll drop that in the Slack channel. If you have any questions, reach out to us on Slack, reach out to us at hello analyst Lab.ai. reach out to us on LinkedIn. We're always here to answer your questions. And then like I said, every Wednesday we're doing one of these sessions. So hopefully we'll see some of you next Wednesday. Thanks everyone.