Testing Claude Cowork on a Data Model

0:02Okay, we're back.

0:03Bookshop.

0:05Bookshop is a demo dataset built by Tableau.

0:10In fact, let's go to this page here and just uh zoom back in.

0:14It's a demo dataset built by Tableau and it's designed to help you learn how to use the data model.

0:21It's a specifically set out data model.

0:25And it's actually a really good standard data set because as the data model has evolved, you've actually been able to do more and more with the Bookshop data model in order to demonstrate the data model.

0:35But

0:36For all intents and purposes, the only reason I'm using it is because I know it fairly well and it's already built for me as well.

0:42I'm lazy.

0:44So there's already a TDSX file.

0:46If you don't know, a TDSX file is a Tableau data source file.

0:50The X stands for packaged.

0:53What that means is that the Excel file actually exists alongside the TDS file.

0:57The TDS file is the data source file.

0:59Tdsx file has the data source file and the Excel file together in a package zip.

1:04That means when you open it, everything is good to go.

1:06So that's essentially what that means.

1:09And so what I want to do is use this as a baseline.

1:12So in the last video, what I did is I used a superstore

1:15I did some analysis, we drilled into sort of one subcategory.

1:18And what I said in that video is that listen, the challenge with that data source is that it was flat.

1:23And I was using Claude cowork to sort of go through it.

1:25So it's actually quite

1:27fairly straightforward for it to do that because it's it's one table, it's one data source.

1:31You can pretty much just you know navigate through the data source fairly easily.

1:35What I wanted to do is see if

1:37Claude Cowork can overcome this challenge of a data model, being able to bring in multiple tables from different places, make the right connections between them.

1:47and arrive at a perspective of analysis if that makes sense.

1:50And what I did in the last video is I didn't I didn't have any questions in my mind.

1:54I just asked it to sort of start looking at it and I just checked the homework as we went along.

1:59So in essence I have

2:02the data model setup.

2:03I have a blank workbook.

2:05We'll do the analysis inside of Cloud Core work and we'll validate it here in Tableau.

2:08So you'll get to see the side-by-side comparison.

2:11And yeah, we'll just see how it goes.

2:12This is sort of the diagram if you're not familiar with the data set.

2:15This is how the

2:17dataset comes together just just for everyone's benefit.

2:20This signal here is a join.

2:21So the suggestion here is you create a logical view which comprises of two physical tables so book and info

2:28with this sort of join function here with so book ID one plus book ID two this should throw the AI off um we'll find out if it does or not

2:36Um, but I won't tell it any of that.

2:38I'll just say do this analysis and I'll ask it for information from this table, uh one of these other tables, and just see if it kind of figures it out.

2:46Okay.

2:46We've got cells which are a union over here.

2:48That's another physical

2:50uh table that's brought together into logical view that has our sales publisher goes to edition via publisher id so to get publisher information for any book we have to go via edition

3:00And for all these other tables, they're just basically one relationship away.

3:03So these are these are fairly straightforward.

3:05In all you know, honest terms, you could join these as well.

3:09one by one and get sort of um the answers you want, but that would make the data source much larger.

3:13Hence a relationship works in this case because you don't have to join 13 tables on to get

3:18you know, this mega table as it were, which is also quite query-intensive as well.

3:22Okay.

3:23Yeah, if you don't know about the data model, I should I should call out that me and Kirk have something called the Data Model Masterclass.

3:28We re nearly about hit 10,000 views

3:31For something as dry as the data model, which means I think that video is a good thing.

3:35You should go and check it out.

3:36We tried to create it as like a one-stop shop to teach you everything you need to know about the Tableau data model.

3:41So go ahead and check it out.

3:42But here's the problem

3:43If AI can take everything that's taken me and Kirk in that video a really long time to learn and compress it into minutes, then I think we do start to look at

3:52the inflection point of this technology actually starting to be able to be useful.

3:56I've not scripted this, I've not done this like the last one, it's completely unedited, so this is pretty rough.

4:01I have tried to record using slightly better setups, slightly better, but I've

4:04Got my camera or nothing.

4:06Just got a better mic just so it doesn't sound rough, which the last one did.

4:09So let's have a look at this.

4:11So if you see my last video, you'll saw what I did with the bookshop dataset.

4:14I just opened this up and actually I

4:17uh pointed this folder to the superstore direction over here.

4:20So this is what you can see is the previous one.

4:22All I've done now is pointed to the bookshop data set in here.

4:25And in that folder, let me just try and bring it up on screen.

4:29In that folder I have

4:31these CSV files.

4:32Well you might ask, well how did I get bookshop into separate CSV files when in fact they have one Excel file on the tablet website

4:40I built a tool and I used AI to do that.

4:42I literally wrote one prompt and I said, hey, um I'd like you to build a Python tool that takes any Excel file and turns it into separate CSVs for each sheet

4:51That was the prompt.

4:52It did it in about two minutes.

4:54I ran the Python script.

4:55It worked first time.

4:56It's worked ever since.

4:57And I've used that ongoing basis.

4:59It just sits on my desktop as a tool and I just call it from the command line and that's pretty much it.

5:04So that's what I've used in this particular case.

5:06Um pretty straightforward and that's what we're gonna do.

5:09Now you'll see that it actually you know labeled everything correctly.

5:12So everything's called Bookshop and then it's got the name of the table.

5:15So the AI should be able to sort of figure its way through this.

5:18So I'll just start with like a

5:20Initial prompt.

5:21I don't always start with a question.

5:22I start with an initial prompt just to make sure it understands what's in the folder.

5:26So um what's in this folder?

5:28Okay

5:30Now it goes up, it will check the folder, it will ask, hey, can I look at this folder?

5:34I'll say yes.

5:36It will send a request.

5:37I'll check what's in the folder

5:39And when you when you see something called working, you can actually view the steps in a lot more detail, right?

5:45So when you say list contents of the selected folder, it's basically running a commanded behind the scenes to read this folder.

5:51So this could be

5:52a command line prompt, it could be a Python prompt, it could be a bunch of different prompts.

5:57But let's have a look at this.

5:59Your folder contains bookshop data spread across 15 CSV files.

6:03Here's what's then reference data

6:05Authors, Bookshop, Edition, Publisher Series Award and Info.

6:09Okay.

6:10The book catalogue is actually there, which is interesting for that reference.

6:13Transactional data, sales, Q1 through Q4, quarterly sales data, these are the largest files totaling 3.

6:192 meg.

6:20Good context.

6:21Checkout, library slash rental, checkout records.

6:24Rating CSVs customer rating.

6:26This looks like a complete bookshop database, correct?

6:28Separate tables.

6:29Would you like me to do something with this data like analyze self-sense create a report or explore any particular aspect?

6:36So

6:37That's gonna we're gonna analyze sales trends, but what I want to do is just see where it takes me

6:45And I you know, if I was being an analyst inside of a business, I would probably have a specific question in mind.

6:52But I don't want to I don't want to ask something too specific.

6:55What I really want to do is just

6:57see what kind of analysis the tool can do on its own volition.

7:01So me not asking anything.

7:03And then try and see if I can sort of

7:06show you what it's doing as much as I can, as much as this tool allows me to, and then go and check that and validate that.

7:12And then we'll get into sort of the drilling into the questions.

7:14We'll try and throw some some questions to throw it off track.

7:17So

7:18I am actually going to use analyze cell strand.

7:20Now if you watch my previous video, you'll know that I use a tool called handy.

7:23So let me just make sure that's running.

7:25Handy, there you go.

7:26All good.

7:27Let's do this.

7:28I'd like to analyze the sales trends and pull out some personal information about books that I should be paying more attention to in this sales data.

7:37There you go, I said my prompt and it's just typed it out.

7:41So let's go ahead and do that.

7:42So I ask a very sort of bland question.

7:44Basically, I have no clue.

7:46Tell me where to look.

7:47I'll analyze the cell data and identify books that deserve more attention.

7:52Let me start by exploring the data structure and then digging into the trends.

7:56So that's already a really good step

7:57If you notice on the right hand side, it's gone and got the artifacts.

8:02The artifacts are the files that it's currently processing.

8:04So it's only got book, author, edition, sales key one, and ratings.

8:08csv

8:09Now, you'll see it's got the steps, it's gone and got those books, then it's run a command, count the sales records per quarter, find top selling ISBMs across all quarters, running command.

8:20Now this is interesting because

8:22The ISBNs are in a very sort of different place.

8:26Let's try and see what it's done.

8:27Count sales records, find top selling, IBM, comprehensive sale analysis.

8:31Excellent.

8:31Now let me identify specific books you should pay more attention to.

8:34with actionable insights.

8:36Okay.

8:36It's a bit of black box because I'm still not seeing the exact queries that it's doing here.

8:41In Claude chat, you would be able to dig into these and drill into these and see what it's doing.

8:48And so excellent.

8:49Now let me identify specific books.

8:51Let's view the steps.

8:52Identify books needing specific attentions with reasons.

8:55Now I have a comprehensive picture.

8:56Let me summarize this for you.

8:57Sales, trends, and books needing your attention.

9:00Key overall tens, total sales, fifty-six thousand units across all courses, seasonal pattern, strong Q3 surge, twenty-two thousand versus Q1.

9:10Q4 stabilization sales settled at 13,000 units.

9:14Top priority.

9:15Books you should focus on The Star Startling End of Mr.

9:19Heed Who by Jonathan Kotzwinkel

9:22Why?

9:23475% growth, 4.

9:254 rating, but only 76 sales.

9:28Action.

9:29This is your biggest hidden gem, premium priced, with explosive momentum.

9:34Consider featuring it prominently.

9:36Okay.

9:36Hashtag Quakasalfi by Longston Lipman.

9:41Okay, action.

9:43Something is driving sudden interest.

9:45Investigate why and capitalize on it.

9:47That's pretty good.

9:49Cal I can Kalakala Avenue by Elmer Komroff.

9:554.

9:563 with a thousand reviews but only 398 sales.

9:59Heavily reviewed but underselling.

10:01Okay

10:02That's interesting.

10:03If it's heavily reviewed but underselling, that would suggest obviously we don't sell everything that's been reviewed.

10:08Um so maybe there's a

10:10uh marketing of visibility just letting people know that we have the book even if if it's very popular.

10:15That's a good that's actually a good spot.

10:17Um Banana Slug and the Glass Halfful by Glory Green, Y4.

10:223

10:23What um 1000 reviews, strong Q3, and 1% great action children's book with seasonal appeals stock up for next Q

10:31quarter.

10:31I actually really like these very clear actions.

10:35They're very good.

10:36I have to check them, but they're not just like hey, some of them are like dig into this more, do this specific thing, hey there's a seasonal pattern here, you should do this.

10:45And your revenue powerhouse keep in stock.

10:48Okay.

10:49Okay.

10:49Waters to watch.

10:51The breadth of analysis here is is wild.

10:54Like there's a lot here.

10:55I could not have done this in the the same amount of time it took to do this, but okay.

11:00Okay, okay.

11:01So

11:02Here's the thing.

11:03In the last video, everyone said to me, Oh, you can't trust this stuff.

11:06Okay, you cannot trust this.

11:07You have to check everything.

11:08I was like, well, listen, what I'll say is your senior colleagues, your leadership in your organization

11:14are paying attention to the point when you don't have to check this stuff, right?

11:19And the minute and the minute that is even remotely possible, the inflection point will come.

11:25I will go and check the majority of this, but I won't check everything

11:28I will sort of try and get a cursory view of this data and probably this is plenty for this video, so I will ask more specific questions about each of these things.

11:36So I'll dig into them.

11:38a little bit more detail and we'll start to find out.

11:40So we'll just start very basic good.

11:42Total sales 56 56,000 units across all quarters.

11:46So let's go over to Tableau.

11:48Okay.

11:49And our sales table is this this one here.

11:52So we can have sales count and the sales order sale date is what we want.

11:58So let's just go ahead and put that across

12:01And let's use quarters.

12:03So we've only got four quarters of data, CSV, that's fine.

12:0656,350.

12:08I'm looking here at the bottom where it says sales count, 56,350.

12:13So that's accurate, I believe.

12:15That's um fifty-six thousand.

12:17It does use a a wavy bracket, so it's not being exact

12:31Seasonal patterns, strong Q3 surge

12:34Yep, we've seen that.

12:35It's right there.

12:36And um almost three times break mid year, key force stabilization, self-strained settle at roughly 13,000 units.

12:43Okay, so let's go ahead.

12:44Let's go ahead.

12:45Let me actually close this because we don't need this anymore.

12:47It's gonna confuse me every time.

12:49Yeah, 13,000 cells, just 93 sales over that.

12:52Okay, so all of that is correct so far.

12:54Okay.

12:55Top priority books you should focus on.

12:57Now this is interesting because it's skipped, it skipped a whole bunch of steps right into like specific books.

13:03So

13:03The Startling End of Mr.

13:05Heed Who by Jonathan Cottsbinkle.

13:08Okay?

13:09Biggest gem, premium priced, twenty-two dollars with explosive momentum.

13:14So the pricing is also going to be

13:18in the book info page.

13:20Alright, so I believe the book info page.

13:24The pricing is a good question actually because

13:29I think it's just going to be at a line item level, right?

13:32So if I just take a single order and we say adult members and I just do cell count just to see

13:40But it's the right thing, yep.

13:42And then you got the discount.

13:44But actually no, you don't have the price in here.

13:46So how would you know the price?

13:50You know the number of sales, sales amount, and let's just check this.

13:53It's been done for us.

13:54It's price, which comes from which table?

13:58Which table does price come from?

14:00I don't have to look for it, I can just type it

14:03And it comes from addition to addition and then you do minus the discount.

14:10So it's basically 100% minus whatever the whatever this would give you the actual price of the book.

14:15So

14:15If the discount is 0.

14:172, that's a 20% discount.

14:18So 1 minus 20% means it's 80% of the price.

14:21So it's 80% minus price, multiplied by price, which will give you the sales amount.

14:24Okay, so that's how they've done it

14:26And so we can actually bring this over here.

14:28Let's put it.

14:29Why did I do that?

14:30I didn't mean to do that.

14:32That was a really bad idea.

14:33I want to put it here for now.

14:34Okay, so we can see the price just underneath.

14:36So you can see each each line item price and then if I go get the book with the book title, the book book title.

14:45You can put that after the order ID.

14:47You'll see.

14:48There we go.

14:49Okay.

14:49So we get to see each individual book.

14:52Some of these have different versions, right?

14:55Or different orders to be fair, so that's fine.

14:57And then the ISBN

14:59And that comes from a different table, so we can go ahead and do that.

15:02And that picture gives us the nice list where we can check everything we need to check.

15:07Alright, so we will also go get the title.

15:12And in here we can say none, hit apply, and then we can go and get our title.

15:17So the startling end of Mr.

15:19Hudu by

15:21Jonathan Cotswinkle.

15:23Okay.

15:24So let's go back in here.

15:25Let's put this in.

15:27Uh just keep that only.

15:29Come on, I hit apply, hit okay.

15:32Good.

15:32Here are all the transactions.

15:3476 transactions.

15:35I'm looking at the bottom cell here, count of sales.

15:39How many does it say 76?

15:41That is accurate.

15:424.

15:434 star rating.

15:44So let's go ahead and get that

15:474.

15:474 star rating.

15:49So the rating is going to be man, this is tough because I cannot remember what anything used to be.

15:56Uh where is this?

15:57Where is the ratings and the rating table suppressed price?

16:00So rating is an interesting one.

16:02If I put it on text for now

16:04You'll see it says does a sum.

16:06It shouldn't be a sum, I believe.

16:08Let's just do average to get the lowest score, yeah.

16:11So what it's doing is it's adding it as many times as it's um

16:15Seeing the book, which is why you get that total, and I think it's because of the way the table is related.

16:19So if we go into edit the data source, ratings is if we hover over this line, let's just do this.

16:27is connected by book ID.

16:29So for every book you'll have a rating.

16:31So that's why that number was a bit high because the relationship is somewhat uh a little more tricky.

16:364.

16:373 is a rating I believe.

16:39Four point

16:404, 4.

16:41370.

16:42And here it's going to be the same for every book, so that's fine.

16:44What we could do is we could put the rating here for that.

16:49uh make a discrete and then you could bring it forward.

16:53So you could do this.

16:55Bring the title forward.

16:57There you go.

16:58See there's your rating.

16:59Those are all the orders that relate to that book.

17:02Then we don't get repeating lines.

17:03So it's a little bit more

17:05It's a bit easier to see what's going on.

17:08So the sales amount some of these have reduced price actually.

17:11You can see that they like some of the transactions have a different price.

17:17That's probably because there was a discount, right?

17:20Let's go find that out.

17:21Let's put that on there.

17:22I should really put it here and say this is discrete, just so we can see that.

17:27So

17:28Yeah, some of them have discounts, some of them don't have discounts.

17:31You can see the reason the price changes is because it's different.

17:34So so far, we've just built out a table, we've validated some of the core information.

17:38And hopefully you've seen how this analysis would work.

17:41But

17:41This is more exploratory analysis rather than like visualization.

17:44So yeah, premium book price at $22.

17:48Is that actually correct?

17:49Yes, it is.

17:50Uh $20 here.

17:51What we'll do is we'll do

17:54Because it's um each line is a record, the sum is basically the same thing.

17:59So Yeah, we'll leave that there.

18:02So you can see it's $23

18:03twenty twenty one it's on discount.

18:05What it's suggesting is consider featuring it prominently.

18:10So this is your biggest hidden gem, premium priced, four hundred and seventy-five percent growth.

18:16Okay.

18:18So let's validate that.

18:20So what we can do is we can take out the order ID detail here.

18:23Okay.

18:2576 transaction.

18:27What I'll probably do is actually can I clear the shaft?

18:30Yep.

18:30So we just get our seventy-six.

18:31We filter to the one book.

18:33Go get our sales date up here on the top.

18:36and go do quarter and then we can put this on rows.

18:41So there you go you've had your big jump in sales percentage difference so 0%

18:49three hundred and eighteen percent and difference to the previous one, thirty-seven and a half percent on Q one to Q two

18:59This is an interesting one.

19:00So if I bring count of sales back, cells, let's put that next to each other there, so we get this.

19:06So we started at an 8 and then we've gone to 46

19:10This is 318%.

19:11So this is an interesting.

19:13Is this wrong?

19:13475% growth.

19:17Let's just take this.

19:18Can you explain the 475% growth?

19:22Because I see a different number.

19:24I wanna know if it's like relative to a specific point in time.

19:29Okay.

19:30So it's given the formula pretty quickly.

19:34What this means.

19:34It books sold five pen times and more coupled in Q four than it did in Q one.

19:38So it's versus Q one

19:39rather than month to month.

19:41So if you don't know, you can do this in inside of a tablets.

19:44Doing the previous, right?

19:45So we can go in here to this quick table calculation, select first.

19:49And now every one will be relative to the first one.

19:52So now our bottom chart obviously is 0% compared to Q1.

19:56Now it's 37% compared to Q1.

19:58Now that's 37% compared to Q1

20:01Here's a big number, 475%.

20:04Accurate.

20:05Spuss on.

20:06So that detail of explaining it would have been nice if it was there, but actually the analysis is bang on.

20:11Okay.

20:11So

20:12Yeah the first piece of analysis.

20:16Pretty good.

20:17Pretty good.

20:18And by the way, that took me five minutes to validate and

20:21probably go through the thing and by the way I also started with a completely open-ended question, you know?

20:27I just I just said, hey, you know, just tell me something interesting.

20:30And there's like here are very precise things.

20:33It'd be interesting if I found, let's say, 20 other things that were more interesting, more compelling, but then I I didn't ask it for that, right?

20:40So if I have an opinion, I should share it and let it do the analysis.

20:44But if I don't, these five things, well at least the first one, looks pretty good.

20:47Everything checks out.

20:49Okay, let's do s one thing.

20:52Let's do one thing.

20:54So

20:55We go to the dataset and so we had some book information, we had some ratings information, we had some sales information.

21:04So that means it must have also grabbed the addition.

21:08You can't get to sales without addition.

21:10Okay?

21:11So this is super interesting.

21:13Now

21:15If we I wish I hadn't closed that web page, I'm gonna have to open it again.

21:19Here we go.

21:19I didn't close it, I minimized it.

21:21Thank goodness.

21:21ISBN is only an addition, and addition is only in

21:29To book ID to this and then ISBN.

21:34Right.

21:34So the journey the journey to

21:39Get the sales data had to traverse through book, edition, and sales queue one, right?

21:44There's no other way it could have done it now

21:47I in earnest, I I can't see the Python that it did behind this to actually get to this.

21:52So I can't validate how it's doing this.

21:54It could have built the table separately and just basically

21:57did the analysis or stitched the analysis together.

21:59It's not using a data model like construct to do it, but it's just doing the steps and basically saying, right.

22:05I have sales, sales have ISBN, ISBN's addition, edition, and book IDs over here.

22:11So to get there, it's having to do this stitch, right?

22:13So

22:14The price information is also here, which is pretty accurate.

22:17I could ask it, hey, can you give me the page number or can you tell me the genre just to see how it's processed that?

22:23But that doesn't mean it's used this data model.

22:24It could have used a different way of getting to the same answer

22:27Doesn't have to be efficient, doesn't have to be fast because it's not in the construct of tableaus.

22:31So honestly you could have done it anyway.

22:33They all lead to the same answer and if that makes sense.

22:37So that's yeah, that's pretty I think that's pretty pretty interesting.

22:42Yeah, I'm still yet to see this thing get something wrong.

22:45Now, let's look uh here.

22:47I could explore these five questions, but I think this will just get very boring.

22:51So

22:51I've done this.

22:53Go investigate it yourself.

22:54Um get Claude Cowork.

22:56Try this yourself.

22:56You find something different, you find it wrong, let me know in the comments.

23:00I will check one more thing.

23:03The Spark and the Ashes by Ursula Corinne.

23:06Then we'll go into the rest of analysis.

23:07So the Spark and the Ashes.

23:09Let's go back to sheet two.

23:11And that's gonna be fairly easy one to just say none spark and the ashes hit apply

23:19So I've got this chart now.

23:20I can just very quickly change the the thing.

23:22So here you go, fifty-one, fifty seventy-five, one hundred thirty-five, sixty-nine.

23:27This is exactly what Tableau is very good at, exploratory analysis.

23:29By the way, like most people don't

23:31Use Tableau like this, right?

23:33Like most people build a dashboard and then send it out to their team so their team can do this kind of stuff.

23:38But like this to me is like the sole reason why Tableau Desktop is still extremely valuable.

23:44being able to do this.

23:45And why I think this is an inflection point is because it looks like clawed code is giving people a natural language way of doing it and giving reliably correct answers.

23:57Okay?

23:57That's a big deal.

24:004.

24:014 rating, 1,000 reviews, only 300 sales, um, 330 sales, high review engagement with a poor conversion.

24:08Is it a pricing issue?

24:10Okay.

24:11So this one's an interesting one.

24:13330 cells.

24:14Let's go check that.

24:15Is that right?

24:15Yep, 330 cells according to the bottom here.

24:18And um lots of volume.

24:21Okay.

24:22Okay.

24:23Thousand reviews.

24:24Let's ask the question about why poor conversion.

24:27So what does it mean poor conversion?

24:30Can you explain the poor conversion?

24:35for and I'll just do this and I'm still talk I'm still typing.

24:40I can't it's something when I'm focusing I just type all the time but when I'm actually

24:45like very like just starting a task.

24:47I use this tool up as opposed to I just forget sometimes.

24:50Okay.

24:52To understand why it was high engagement but low self

24:55conversion.

24:55So it did that analysis, but I'd love to know what it what's its reasoning.

25:01It's almost flipped the question, right?

25:03It's almost sounds like it

25:05I've asserted this and now it's going to figure out my assertion when actually I was just restating um this.

25:13And I think it's because of the way I wrote the prompt.

25:15I wrote the prompt in a way that asserts the facts.

25:18So

25:18I should have said, can you explain why you think the poor conversion, in fact, this is AR, we can stop it.

25:25I can tell it to stop.

25:26So I can take this and say

25:28It's it's still probably running the background, so I've probably messed this up.

25:32Can you explain why you think

25:38the poor there I should say there was poor conversion for the spark on the ashes.

25:45So this is a bit more appropriate.

25:46So I'm asking in, can you tell me

25:49Why you think this is, okay.

25:51So good question.

25:52Let me clarify what I meant.

25:53This is the prompt I should have done for that book.

25:56Okay

25:56Number of views is that total sales is that reviews to sales ratio.

26:00Okay, that's a metric.

26:02That's very cool.

26:03It's 3.

26:034 reviews per sale

26:06Typically, only a small percentage of buyers leave reviews.

26:09Industry norms suggests maybe 1 to 10% of purchases review your product, so you'd expect normal ratio of 0.

26:1701.

26:18Oh my gosh.

26:19Chef's kiss, isn't it?

26:21That's actually really good.

26:22One review for zero cell.

26:24This book has three point four reviews per cell.

26:26That's inverted.

26:27There are more reviews than sales, which is unusual.

26:30This assumes that the reviews dataset

26:33is our own data set, right?

26:36Uh we could have bought the reviews data from someone else.

26:39Oh, that's why am I even guessing it's given some explanations?

26:43Reviews from other sources.

26:44The ratings data might include reviews from library borrowers.

26:47Other retailers or aggregate sources, not just your shots.

26:50Historical sales not captured.

26:52Great shot.

26:52Having samples.

26:54Advanced reader copies.

26:55I didn't even know that was a term.

26:57You could generate reviews.

26:58Data quirker ratings and sales data may come from different time periods.

27:02Exactly.

27:03Great question.

27:04Yeah, good point.

27:05We only have Q well one year's worth of data, but the reviews could be historical

27:10I may have overstated pure conversion.

27:12What the data actually shows is, oh, good correction.

27:14This book has stronger reader engagement, but relatively modest sales compared to other books with similar review units.

27:21It's not necessarily a convergent problem.

27:22It might just be the book hasn't had enough of the ability to turn the positive word of math.

27:27Okay.

27:28Would you like me to dig into the raw data to compare review to sales ratios across other books to see if this pattern is consistent?

27:33Yes

27:34Be my guest.

27:36It could be innocuously explained by just being, hey, we have one year's worth of data, but our reviews data is a table and contains reviews.

27:48from almost the storic sort of nature of the book.

27:51So it could be the book's been around for three years, but we're just analyzing cells from one year

27:57Actually write at the average 3.

27:5928 for your bookshop.

28:00So I was wrong to single it out as having poor conversion that's performing normally

28:04So this is the first example of where I would say you've seen this issue of assertion where like AI sees one thing and it goes, well, this is wrong.

28:14without doing this bit, which is like going down that sort of trail of thought and actually doing the investigation.

28:21Now

28:21The review ratio, I'm not gonna go bother checking that because I I've uh there's not a time where I haven't done the math and it hasn't checked out, but it's the assertions, it's sort of the steps that it's going through to sort of make these sort of comparisons that are somewhat broken.

28:35So

28:36That is where you still need what I would call a data analyst, literally in the business sense rather than the data sense.

28:44to come in and ask these questions like as someone who's smart enough to know, but yeah, here is a business reason why that could be the case.

28:52Let's dig into that accession.

28:54And so someone like that will probably come into this with a much more precise sort of fine tip

29:00tool perspective to ask good questions about the data that this tool could then run with.

29:05So if I try and sort of explain the balance of where this tool so far would be good

29:10is if you're not relying on it for the, hey, tell me what to think of my data, and instead you have some specific opinions or a specific investigation you'd like to do with your data

29:22And you can come in here and get it to just raise a focus on that thing.

29:26Basically get it to do the math for you.

29:28And the whole time you've got like a language interface that you can use with it.

29:33But you're not having to do what requires the data analyst skill set, which is understand the data model, understand how to open up Tableau.

29:42In simple terms, imagine don't need to use a Tableau interface, but you've still got the ability to ask fantastic questions

29:50But all you need is a keyboard.

29:51That's essentially what I think this tool so far is showing that it can do very, very well.

29:57Okay

29:58So I'm still quite happy with this.

30:01Now I had a conversation with Kirk and we weren't sure if this would be able to do the strings of being able to

30:10Connect all the data sources.

30:11So what I want to pivot to now is it's clearly demonstrated that it was able to do that because the sales data that we ask it to analyze and all the information that it pulled about the price

30:23All came from all of these different tables.

30:26So, what I'm gonna ask it to do is something rather unusual

30:30I'm gonna ask it to explain the relationship between all the data sets, right?

30:35Basically, hey, I'm trying to understand this data

30:38Can you explain the relationship between these tables?

30:44I won't guess at it and we'll see if it doesn't.

30:47So I'll say, okay, I've been

30:50This is where I should type because it'll be much faster.

30:53Okay, I've been challenged by my data team to understand how you've built the relationship between each of these tables.

31:00Could you map out how you understand these tables come together and also map out how you've used them to respond to the questions I've asked so far?

31:10Okay?

31:10So bigger prompt, very abstract.

31:12Let's see if you can do it.

31:15Okay, so this is more of like, hey, show me your homework.

31:19Okay, it just came straight away.

31:21Let me map out the data model.

31:23Oh, nice, nice.

31:24Here we go.

31:25So author and book.

31:28Okay, a book ID goes to edition ratings and award.

31:32Addition goes to ISBN I oh my god

31:38That is wild.

31:39That is wild.

31:40Let me just take a screenshot of this.

31:42I'll show you why in a second.

31:44Let me just do this.

31:46Okay, then we can go back to here.

31:48Maybe go and close this down.

31:51And just let's just compare these two and see how off it was.

31:56So book info

31:58Interestingly, it didn't need the book info.

32:01So it didn't even bring it into its analysis.

32:03Which if you were doing this as a query, SQL query, that is exactly what Tableau does behind the scenes actually.

32:08When it fires off the query to go and get this data, it sees that it doesn't need anything from

32:12info so it doesn't include it in the info.

32:15Now the book links to the author via a publishing author ID.

32:19That is correct.

32:20Primary key, foreign key.

32:21Very good.

32:22It links to series via series ID, primer key, primer key, very good.

32:28And it's actually showing that here.

32:30Then it goes to book ID to edition, correct.

32:33Oh, but this is interesting.

32:35You see it has book ID and book ID.

32:37Oh, very interesting.

32:38And what it doesn't tell me is how it's brought the book ID information together.

32:45So

32:46Does it spell this out anywhere else?

32:49Let's see.

32:50I mean it's done a lot of like wow, this is a as a way of just discovering data, this is actually not bad at all.

32:58So

32:59Imagine your day trial, you just want like a simple place to get started.

33:03So here we go.

33:05It didn't have to bring the info and the book information together, so it never needed to do the join calculator of book ID and book ID one

33:12So it got away with it.

33:13Everything was just using a single field.

33:16So this is actually all correct.

33:18Addition and everything is fine.

33:20Now info, additional metadata.

33:23And let me ask it the question.

33:25So

33:26How would you bring in the info table if I wanted to know the genres?

33:35And

33:38of the books we've looked at.

33:41Let's ask it that one specific question to see if it trips up on this join calculation.

33:48At least a drawing calculation of what it's called.

33:50So it's updated.

33:52Let me first check what's in the info table to understand.

33:54A few steps.

33:55It read the file.

33:57Well great step.

33:58Now I can see the structure

33:59Here's how the info table fits in.

34:03Okay, so you see here it does slip up.

34:07Book ID one, it says the foreign key for book ID.

34:11That is not the case.

34:12So first part of the composite book ID, oh no, it does get it right.

34:19It does put composite key here.

34:20Can you see it just didn't place it correctly?

34:23Wow.

34:24Yeah.

34:24Composite key.

34:25Book ID.

34:26Book ID one.

34:27Very smart.

34:28So TM uh the genre okay.

34:30So key in the book ID in the other tables like TMI is split into two columns here.

34:35To join you need to concatenate book ID, book ID one.

34:39How did I assume it I was wrong?

34:41Okay

34:41How to join our existing fields.

34:43It's actually got the Python fine.

34:46And it didn't need to do the Python here.

34:48It just wrote it out, which is a little bit shit.

34:51Don't know why it just didn't show me the Python.

34:54But that's fine.

34:55Now I can show you the complete picture.

34:58Okay.

34:58So there you go.

34:59It takes all that.

35:00Genres for the books we discussed

35:02Oh wow.

35:03So let's actually go and check what book have we got in here.

35:06So The Spark and the Ashes.

35:08Did it do that correctly?

35:10Yes, because this is technically part of it, so we can put it in here.

35:16Genre.

35:17Young adult for the spark and the ashes.

35:19So let's see if it got that right.

35:21The spark and the ashes.

35:24Where did that go?

35:26Young adult is in there, but it did not.

35:30It missed one of the books to be fair.

35:32It missed one of the books, so that's fine.

35:36Let's check the one that we did start with.

35:39What is that called?

35:40The startling Actually is that even one of the first ones?

35:44Or is it hallucinated new books?

35:46This potentially could be hallucinations, let's just see.

35:50Okay, so there here we go.

35:52What is one of the first books?

35:54A startling end of this.

35:57Okay.

35:58We did do the Spark and the Ashes by Ursula Karine Neen, but it didn't

36:08It didn't put it in there.

36:10So that's a good that's a good detail there.

36:12So the hallucination here has missed a book that we did do analysis of.

36:17Okay.

36:18The startling end of Mr.

36:20Heedhu.

36:22is okay.

36:24But some of these were not part of the original analysis for sure.

36:28One, two, three, four, five, six, seven, eight books now.

36:31There were only five books and I don't believe any of them have suddenly turned up in our analysis.

36:38So let me just go and do like Ultara check.

36:42The actual conversion issue is never okay.

36:46It did bring some other books into this check, but they weren't the same three.

36:52Okay.

36:54It's hard to know where it thinks those books were coming from.

36:59So okay, the revenue power houses, uh, here they are.

37:04So let me just

37:07This is the the thing I I don't even know if there's this sort of I will say that

37:17I like I don't have perfect memory.

37:19So actually another issue here is that it's spitting so much data out of out at me that I'm not even able to remember the full context of everything it's said to me.

37:28So I like

37:30In a weird way, gosh, am I am I subpar compared to the computer?

37:36Like should I be worried?

37:38I'm trying to take a screenshot for every book.

37:40I'm just gonna leave it on screen, just so we can see this and check.

37:45Your best sellers, blah blah blah, land of skin's best.

37:48Okay, so these are all the books, right?

37:51So I have on screen every book

37:54So let's go check and critique its last list.

37:57Okay, so the last list it came up with.

38:02Okay, the startling end of Mr.

38:05Dingy.

38:05That is in there

38:06Hashtag Quaka Selfie, that's in there.

38:10Kala Kala Avenue.

38:11These three are all up here.

38:13Banana slug and the glass half full.

38:17Not in there.

38:18Fine

38:19The Malamara King is there, Rhisworth is there, Port Miron is there, 9083 is there.

38:27So indeed, none of these are hallucinations.

38:30They were all in the analysis.

38:32The issue was my prompt, right?

38:34So what did I say in my prompt?

38:36How would you

38:38Okay, so oh no, don't drop it in.

38:41No no no no no I don't want that, I want that.

38:43So the issue is this prompt, so

38:46I can't edit this prompt.

38:47You can't sort of go back in this sort of reasoning approach.

38:50So let me say this.

38:51How would you bring in the info table if I wanted to know the genres of all the books we've looked at

39:01in this conversation if I can put that in in in quoted speech mark.

39:07Does it come back with all of them?

39:08One, two, yeah, should be.

39:10Should be

39:11Okay, so it does that, it does that now.

39:13Okay.

39:14Okay.

39:15It does them now properly.

39:17So banana slug is in there.

39:19It's actually done it in order of the analysis that we did.

39:21So these were the first five.

39:23These were the next four.

39:26Spot on.

39:28And then Lace and Brandy and Lady Water Sky Space.

39:33Now this one.

39:34Ah, these were here.

39:36I just closed it.

39:37That's my bad.

39:39There's a third one in there that it didn't grab, I think.

39:44Don't check your ego and it's never a glass.

39:47So let's just take a screenshot.

39:49This

39:52And I scroll all the way down.

39:54This is kind of like very tedious.

39:58Lady Water Sky Spice, it's never just the glass is not in there.

40:02Don't check your e-bow is not in there either.

40:05Lacey and Brandy.

40:06Lacey and Brandy is not in any of those.

40:09Oh am I no, there, that was the end.

40:13Okay.

40:14It did a pretty good job, but it it did miss a few.

40:17And that's me being a pedant.

40:18It technically wasn't correct before.

40:20When I say all it's missing a few, it just it just doesn't have that thing.

40:24So again

40:27Should you trust a process that can't remember all the work that it's done?

40:34That's sort of a question, right?

40:35Because it it has missed some.

40:38That's partly hallucination.

40:39That's also partly context.

40:41There's also there's there's a term for it which is called known as like I think it's like context bleeding or something like that.

40:47in the AI industry where it just slowly loses bits of information the longer the conversation carries on.

40:52And the longer, the bigger the context, the more of these you start to get.

40:57So maybe I'm also starting to hit the limit of like, oh

41:01This is the analysis I'm doing, blah blah blah.

41:03If I was to go and start a new analysis, what I should do is actually go to a new chat and start again, give it the same folder, because I don't have to create a copy of these files and then start that analysis again.

41:16But it's almost very inefficient because what you'd love to do is keep this context of the way it's brought the data together and just be able to constantly have a conversation with the data source.

41:24So this is something that Tableau and all those tools do really well.

41:27You set up your data model, you can just keep going back to it and building on it.

41:30And you can't necessarily guarantee that it will do these upper parts reliably every single time.

41:35It's not deterministic.

41:38to an extent probabilistic based on what it's seeing in the far, it might come to a different reasoning in a different line of questioning.

41:44It's just just the way the technology works.

41:46Interesting.

41:47Okay, I've been going for 44 minutes.

41:49I said I wouldn't edit this.

41:50This is probably a good time to stop this.

41:51Um hopefully the timestamps help, but I just wanted to do this experiment just to see where this goes.

41:58I I do think

41:59I do think this is something really interesting, like in a huge way.

42:03All the other ones, you know, when you go into AI and you ask them to do maths and it gets it wrong.

42:07Like that was never the that was never the goal.

42:11At least from a data perspective, it's this stuff, right?

42:14And the interface is not great.

42:16I've not seen a single chart.

42:18It's all text.

42:20But then again,

42:22Like a lot of people just like to read stuff, and if this is going to be this accurate, um great.

42:27If you can ask it to do charts, great.

42:29In fact, let's ask you that.

42:31Hey, can you do a deck on

42:36All of this analysis, right?

42:39Including charts and captions to explain the trends.

42:45Okay

42:46Let's just see what it does in one shot.

42:48Let's sort of close this off.

42:50Can you create something that goes off and does it?

42:55Now, interestingly

43:00It's gone and written a prompt based on my prompt.

43:03It's like, yo, that wasn't enough, Tim.

43:07It's used the skill to beef up my prompt.

43:10It's taken that prompt.

43:13And literally read written a much better prompt.

43:16Wow, okay.

43:19That's the prompt I should have written to get a really reliable response.

43:23It's visualized all the slides and it's gone through the steps.

43:31Wow, it's really gone ham on this.

43:34Like

43:35It did that very fast.

43:36There's gonna be mistakes in that.

43:39View steps, okay.

43:40It's read HTML, something marked down, which is this thing.

43:44It's read the CSS markdown.

43:45Now I have a full understanding of the workflow

43:47It kind of wrote itself marked down in HTML.

43:51Let me create a comprehensive presentation with charts analyzing the bookshop data.

43:55Update to do list, create presentation workspace.

43:58Updates to do this, generate comprehensive chart, data for presentation.

44:02Now we'll create the HTML size with a bold and editorial design.

44:05I'm choosing an editorial magazine aesthetic with a deep teal navy color scheme

44:10and Coralescence to make the data pop right in the file, creating a cs file so okay.

44:16Boom.

44:17What I expected to do

44:19is to put it in here.

44:20Oh, there it is.

44:21There's the presentation folder.

44:22It's busy writing the HTML.

44:25It's writing a sl uh a HTML page for every slide.

44:29This is like ridiculous, isn't it?

44:31I'll just leave this on screen so you can

44:34See what it's doing.

44:35It's a bit broken because it's doing this one slide at a time.

44:39And this is no doubt going to use a ton of AI credits, like each and every one of these HTML pages

44:45The prompt to write it is a query off the called code and it's just going to kind of chug on this.

44:51So let's see.

44:54Slide five.

45:01Slide six

45:12This is pretty impressive.

45:14Even just I know this is gonna be garbage when I open it, I'm sure it will be, but like

45:21I think I think I keep saying to people, this this stuff doesn't get worse.

45:24It only gets better

45:26And it's easy to look at it now and say, oh, it's not quite there yet.

45:31Until it's there, right?

45:33And then suddenly you've got a different conversation.

45:36So to my to my point of the post I made on LinkedIn Alien, like last week, lean into this stuff.

45:41You need to understand

45:44What all of these steps are, like where are the gaps?

45:46A bit like I just did in this video.

45:48I kind of went through and I said, hey, maybe it might be good at actually helping you drill into a data set, but you're not trusting it to choose the direction.

45:55or to choose a nuance.

45:56You might want to check the mats a few times just to make sure it works out, but it might actually help you drill into something where you already have a perspective, but don't have the time to learn Tableau and learn all these other tools.

46:08So

46:08So it's creating what is this?

46:11Slide eight?

46:12What are we?

46:12Slide nine?

46:13Yes, doing slide nine

46:16I don't know if you see it, it's uh happens very, very quickly.

46:19Oh, it's done the nine slides.

46:21Great.

46:22Updating to Lilius, creating

46:24presentation.

46:25js is like the JavaScript file for that.

46:29That's right, really like ridiculous.

46:31Yeah, wow, okay.

46:33HTML2ppTX is like an HTML.

46:35Must be a HTML library for PowerPoint slides.

46:38And presentation JS might must be the JavaScript that runs it all.

46:44And the to-do list, uh we can't see it, but um

46:49Oh actually you could see over here on the top right hand side and if we just limit these I can't really expand this more so building the PowerPoint presentation visualize validate and fix any visual issues.

47:01Oh this okay

47:03I've seen this happen before and it's absolutely dumpsifying.

47:06So it does actually write the HTML, then it uses that HTML to build the presentation.

47:12And the way that all works is it goes from HTML to PowerPoint.

47:15That's exactly how it works.

47:17And the validate and fix any visual issues is like a weird one because it's based on what it sees.

47:30So it's got node.

47:31Node.

47:32It's just installed a bunch of dependencies that I don't want to know about.

47:35Oh, the playwright browser download is bot.

47:38Let me use an alternative.

47:39I'll create the presentate directly with.

47:40PPTX Gen JS without the HTMLT PPTX inversion.

47:46Wow, that's a very

47:49And yeah, if you're sat there paying for AI cribs, this is gonna be costly.

47:54Um if I just go to my settings and go to usage.

48:00Yeah, we've used nearly 10% and I haven't used Claude before I started this video.

48:06So just this interaction, I'm on Claude Max.

48:09I get sort of five times more like you know capacity than the Pro Plan.

48:14And I've used ten percent nearly of my allocated usage

48:22within the session, our session is classed as a certain period of time.

48:26The weekly limits, two percent.

48:30Yeah, I probably wouldn't actually get through that.

48:34Because I've got until Thursday.

48:36But the current session is actually quite quite allowing.

48:40So let's go back to there we go

49:01Oh, here we go.

49:09Oh, it is there, it is there, it is there's a file.

49:12There's a PowerPoint

49:28This movie's a phone, okay.

49:31Okay.

49:34Drum roll please.

49:39Here we go.

49:41Bookshop sales trends.

49:43Insights, opportunities and recommendations

49:51I have to say, back in the day, people used to burst reports for this very reason, okay?

50:01And not gonna lie, like numbers accuracy aside.

50:07This is pretty scary.

50:08This is pretty scary.

50:10It's not a great looking deck.

50:11It's not branded.

50:12I didn't tell it a brand, but

50:15This looks like if I I I'd struggle to understand how by the end of this year

50:23This was not already performing at a higher standard than it already is now.

50:28Just even this, like getting the book reviews out, this is already sort of m sort of more

50:32sort of meeting ready than something like in Table A, right?

50:35Like Tableau itself is a product you have to browse, you have to go look at it, how the data connects.

50:39It's nice that it's done this.

50:40It's actually quite useful.

50:43Promote hidden gems, leverage Bianca Thompson, investigate growth drivers.

50:48Oh man, that's not bad.

50:51That is not bad at all.

50:55Yeah.

50:56And yeah, this is the second time I'm using this tool.

50:58I'm sure if I use this tool every day

51:02For a couple of hours to find out all the places, all the nooks and crannies where it's good and where it's not

51:10Yeah, this we're having a different conversation in a couple of months' time.

51:13Easily, easily.

51:15So yeah, I'll keep pushing on that.

51:17I'll keep sort of playing around with it, see what I can do.

51:19The thing that just is wild to me is that I didn't need a single business intelligence tool to go and do this.

51:26And you might say, well, yeah, you've got Claude Max, but then, you know, Claude Max is not just our data analysis tool.

51:32It can also write scripts, it can also write Python, it can also code, it can also help me with my writing.

51:38It's a multi-purpose, on-demand software idea generation, like creative tool.

51:44It's pretty broad what it does.

51:47So you can easily see why companies will be worried about pretty much all the industries collapsing into AI.

51:54A bit like how all the cameras

51:56collapsed into the smartphone.

51:57Camera industry is just now.

51:59$2,000 cameras and nothing else below that because your smartphone already does that basically.

52:04Okay, so I closed that on the PowerPoint deck, but then I messaged Kirk.

52:09And he asked me a question.

52:11He said, Hey, can it do the checkout to sales or issue for the top three books in the dataset?

52:18And it looks about right.

52:20I'm not gonna check this now because the video's too long, but I'll leave this on screen just so you can see, like a question that at least Turk and I think is difficult.

52:29for it to do, right?

52:31The checkout to sales ratio for the top three books in the data set.

52:34And it comes up with this, right?

52:36Okay.

52:37And then I ask you, hey

52:38How did you do this with the data?

52:40Can you show you're working?

52:41And so basically this is the steps it's taken.

52:43So the challenge is that checkouts and sales are stored differently.

52:46spot on.

52:47Checkout, Book ID, monthly checkout contains directly by book ID sales individual sales transactions by ISBM.

52:55So one is it's like order ID level

52:59Another one is monthly aggregated data.

53:02Sales are recorded by ISBN, but checkouts are recorded by book ID.

53:06I need to convert ISBNs to book IDs to compare them.

53:10Really smart, astute bit of analysis.

53:12Alright, so

53:13It does cells, converts them to ISBN, counts the cell per ISBN.

53:18Great.

53:18Next step.

53:19Takes additions.

53:21Gets the ISBN to book ID.

53:23Converts all those ISBNs to book ID.

53:25Looks up the book ID for each ISBN.

53:27compares the total.

53:28Okay?

53:29Boom.

53:29Then goes to checkout.

53:31Then counts the checkouts per book ID because the checkout is per book ID.

53:35Then it compares and it brings everything across.

53:38It's got like a little bit of code saying, hey, step one, you do this, step two, you do that, step three, you do that.

53:45It's doing it in Python.

53:46Get top three book totals, calculate the ratio.

53:49Yada yada yada yada yada.

53:50That's the that's the sort of how it does it.

53:52Then I'm saying, what can you do this in a time dimension to the top three?

53:57break it down by month for each book for the top book so the mala marking which is the top book apparently and it goes and it does it boom

54:08And we can go check this in Tableau.

54:10It would actually be a little bit tricky.

54:12I'd need to sort of make sure I spend a bit of time actually figuring it out.

54:17But

54:17I'm gonna leave this in the recording and in the next video we'll go through this.

54:22And the reason I came back is I'm actually now really interested.

54:26Throw me your worst.

54:29People always say, ah yeah, but you know, every business data is like a mess, right?

54:33Like it's never this clean, it's never that good.

54:35And true.

54:36Bookshop and superstore are super clean.

54:38So what I'll say is listen

54:39If you've got mock data that is you think is challenging, I already think this tool has proven that it it it it interrogates the data

54:49better than a a a junior to mid level C uh analyst.

54:55100%.

54:56Like no one can tell me that a junior

54:59or mid-level analyst can do the same level analysis as fast as it's done them as accurately and as eloquently and then display the information in a PowerPoint very clearly.

55:08uh with actions and next step as if it understands the business.

55:11Like no one person could do that out of the gate as quickly as this tool has done it.

55:16That said, it was an easy data set.

55:19So I would love if anyone has a publicly available data set with some challenging ways of bringing the data together that require really nuanced ways of working.

55:30I would love you to send the data in, put it in the chat, maybe a link or whatever, or reach out to me on LinkedIn, whatever.

55:38Don't

55:39Don't tell me what questions to ask.

55:41I'd love you to suggest a prompt that you think would throw it awry.

55:45And then we'll put that in and we'll work with that basically.

55:48Okay?

55:48So that's gonna be

55:50my suggestions but invites everyone else just to make this video go all the way up over an hour.

55:56Thanks for watching.

55:57See you in the next one.