AI can't do data analysis? Claude Cowork thinks different. [Un-edited]

0:02Okay, so we're gonna do something completely unscripted.

0:06Um I'm just using a webcam off my thing.

0:10I thought I'd like

0:11give a little window into not just how I use AI but also some neutral a clawed co-work I haven't I haven't tried it so

0:20So it might just be uh I don't know.

0:24Let's just let's just make the video and then figure things out.

0:26Um by the way, I am making videos to just I'm just not posting any.

0:31Go with a different tactic for 2026.

0:34stay tuned but on this particular one I am gonna post this very quickly to LinkedIn and we'll go from there now the number one thing that I think

0:44people potentially don't realize is some of the best ways to use AI do not involve these fingers.

0:51Does not involve you typing the prompt into

0:53the tool.

0:54A lot of the really pro users of AI have actually stopped typing prompts and have started talking prompts.

1:00So here's an example.

1:02I use a tool called Handy.

1:03This is a Mac tool.

1:04It's free.

1:05can go and download it.

1:06Other text uh speech to text tools are available.

1:11Um the way this works is if you want to talk it

1:14uses uh a model called parakeet v3 i've actually been using parakeet v3 to translate sorry to transcribe all my videos for the last

1:24four months.

1:25So if you've seen any sort of video on my channel, the the the transcript for that, the subtitles are generated using Parakeet V3.

1:33It's really powerful, it's really fast and it's really accurate.

1:37about to see how that works here.

1:38You can download other models that you've maybe heard.

1:40WISPO is a very common one, but Parakeet V3 is sensationally fast.

1:45So fast that it's actually good enough to transcribe what I'm saying in real time.

1:49and put it in.

1:50Let me show you.

1:51So I have handy set up to this shortcut option Z.

1:56So when I hit those two things, what you'll see is uh down

2:00here where my cursor is at the bottom in red you'll see a little thing pop up telling you that it's transcribing so let's go ahead and try and do that so here we are you can see that it's transcribing and when it's done it transcribes and then it

2:13pastes that somewhere now because I didn't have my cursor in this chat field it wasn't actually transcribing anything so now I'm in cloud core work what I've done is I've pointed it to the tableau C

2:25superstore folder specifically the GB version so for 2025.

2:303 there's a folder in there this is in my Tableau repository if you want to sort of try that

2:34this yourself.

2:35I've just gone in there and I've just said hey you know what I'm gonna point it to this folder doesn't have to know what's a TDS file but we can also see if it can understand the XLS files that are in there and if it

2:45Can't we ask it what it should do?

2:47So no pre scripting done.

2:50I've just installed cloud called co-work.

2:52That is all good.

2:53Okay.

2:53The final thing I need to explain is called co-work is only available on the Mac

2:59plan of chat of not chat gpt of Claude the max plan is not the $18 subscription this is the $90 a month subscription it's the it's the subscription that's tiered towards a

3:10developer who uses this and I use it because I get massive context windows just as an example the settings here allow you

3:20To see usage, so you can see I barely, barely run up against the limits.

3:27But when I was using the pro version, I would sometimes on like three out of five days of the week

3:34I'd hit the limit and I'd have to basically wait until the next day to finish something really complex.

3:39Maybe I'm making a video, maybe I'm working with a script, maybe I'm doing some uh formatting, maybe I'm doing

3:45code I would hit the limits so so fast that does not happen here with Claude Max and so um I bought Claude Max thinking maybe I'll regret it and I bought it and

3:55What I'll say is I have built tools that more than pay for the subscription of uh Claude Max very easily just to manage the channel.

4:04I've used it to move a whole bunch of content around my website on web fly.

4:08I've used it to build APIs to hook together my YouTube and my Webflow account and everything is good.

4:13So I feel like all this setup is important because it just explains the context behind it.

4:17what I'm about to do.

4:18Okay, so all of that is out of the way.

4:21Let's actually have a go at this.

4:22So let's go to new task here.

4:24You can see I'm connected to that folder I mentioned.

4:26It's the same one here.

4:27And what I'm going to do is I'm

4:29to ask it to tell me what's in this folder.

4:30So I'll go here and select option Z and I'll ask it, can you tell me what's in this folder?

4:36You see the um the transcript actually got

4:40me saying option B option Z and I'll just delete everything that came before that.

4:45I pressed that way too quick.

4:47Uh Grammarly and other AI tools also in there so let's just go ahead

4:51Hit um correct that hit blah blah blah blah blah.

4:53Let's go.

4:54Let's ask it what's in this folder.

4:55So it's loading skills, it's loading the agent, sending the request, looking at my folder.

5:00Uh I'll check what's in the folder.

5:01folder selected.

5:02I mean I could have done this in seconds but the point here is it's can it understand the context and absolutely yes it can okay so three TDS files two excel files world indicator

5:13indicators.

5:14hyper.

5:15Hyper file is a very old format, but it's there.

5:18These appear to be tableau example data set.

5:20The superstore file contains fictional retail sales data commonly used for learning Tableau and the world indicator file

5:26contains global statistics.

5:28Would you like me to explore any of these files in more detail?

5:32Let's try.

5:33Yes, which of these files can you analyze?

5:36Okay, you can see it's uh automatically transcribed that.

5:39And this is a much faster experience.

5:41I'd still be typing if I was doing things, and I'm bad at um typing.

5:45So I can analyze all of these files in different ways.

5:47XLS files.

5:49it can pass that uh the data source file tds files these are XML based metadata files that define how Tableau connects to data sources I can read the structure and see definitions hyper extract is

5:59This is Tableau's proprietary column in the database format.

6:03Direct reading is challenging without Tableau specific libraries.

6:06However, I also have access to the Tableau MCP tools.

6:10Very smart.

6:10They can query publish data sources directly

6:13if any of these are published files to your tablet server i could query them through that connection now that is chef's kiss right like it's like like a very nice bit of context threading and yes it's

6:24possible ah I love that it it found a tool I didn't ask it to do that and it's just gone and connected it up would you like me to analyze the excel files check your type

6:34server both right what interests you the most and I could sit here and say all of it and that's a potentially kind of silly thing to do but in this particular case we'll just focus on the Excel file so

6:47um let's just focus on one of the excel files you see here i typed it i don't know why i should have just said it uh more loud um but anyway let's just

6:59hit enter and my fingers my fingers sort of took over because I got excited that actually does happen when I start like going oh tell me more so let's look at the excel files so asking for clarification which of these excel

7:10Would you like to analyze?

7:11So the clarification is nice.

7:13Um, we can just go ahead and analyze the American one.

7:16Let's just look at this.

7:17So let's let's let her have a look.

7:19And data analysis is one of these things that people say that AI can't

7:22do but I also sometimes think it really depends on the tool and how you're doing it because I do generally think that in some cases the tool you're using really really matters and so here you can see it's running some stuff it's actually failing I don't know if you're seeing this

7:36Um you can see here um it's uh it's it's tried to do a couple of things and it's failed and it's tried to do something else.

7:43This one has worked and it's gone and got the sheet.

7:46So it's got a very um like um

7:49Let's try and run and let's see what happens mode.

7:53So don't like errors are normal, errors are part of the course, but when it figures itself out and does what

7:59what it needs to do.

8:00This is a really sort of great step.

8:01So here we go.

8:02Here's what's in the sample superstore cell file.

8:05Now the really scary thing is I've not opened any analytics store here.

8:08I'm just talking to this LLM and it's you know

8:10Essentially telling me hey, you've got three sheets, orders, people, and returns, main transaction data, regional managers, returned orders.

8:19Okay, so let's go back.

8:21Orders sheet conclude.

8:23columns include order info, blah, blah, blah.

8:26I get a bit of context.

8:27So what I can start to do is um say, okay, um tell me which are the um worst selling

8:35subcategories right so what is the worst selling subcategory in the data um break

8:46This down by yeah.

8:49I still I did it again.

8:50I did it again.

8:51I completely forgot to just talk it through.

8:53So let's try this.

8:54What is the worst selling subcategory in the data set and bread

8:58this down by year okay and now but yeah it's kind of funny so this is this is ai being ai it kind of hallucinated what I said

9:08said it meant year not yeah but I'll take it autocorrect stepped in which is fine let's hit that so let's see what it

9:16does.

9:17I know this data set fairly well, but we can also go validate this in Tableau, right?

9:22And we can go and check this out.

9:23So it's working, it's running the command.

9:26And I think the critique most people would say is hey I think I can do this faster.

9:29probably could but hey um I don't have Tableau open um I'm using a $90 tool $90 a month tool by the way Tableau subscription is actually a little bit more

9:40than that on a monthly basis um to to kind of get the full suite so this is actually you know pretty valuable can it do as much as tablet obviously not but the point here

9:50is I'm just getting on with this analysis.

9:52So we have year, sales, profit, quantity, and orders.

9:55Okay, so worst selling subcategories fasteners with a total sales of 8,132, the entire data set

10:02Fastness is by far the lowest performing subcategory.

10:05The next lowest is labels, has nearly 50% more at 12695.

10:11Is this correct?

10:13Big question.

10:13Let's go find out.

10:15So I'll open up Tabo 2025.

10:183.

10:19This is the 25.

10:203 version.

10:21That's the folder we're in.

10:22So I've just

10:23spun it up in the background of course what will happen is it will crash it actually has crashed behind the scenes it's like oh you've changed your license because you've not opened me in a couple of days so let's uh this this

10:35bug is just persistent on the Mac and then I'm gonna have to go in and license it.

10:40So let's go, let's go license Table A cloud.

10:44And then this time, okay, this this is already taking longer

10:48than um Claude but we're gonna get the answer we're gonna get the answer and I'm just gonna upload this in one take I'm not gonna edit this video because I do not have

10:58Tableau community leaders.

11:00I was not planning to edit a video today, and if I don't get this out today, this will not happen.

11:05Let's get my password manager out and go to

11:10Um community.

11:14Oh, table.

11:17Blah blah blah blah blah search.

11:19Okay.

11:20Where is this?

11:23Where is my login?

11:23Where's my login?

11:25No.

11:27It's not sorting it by most used anymore.

11:30That's so sad.

11:31Um

11:32Most relevant?

11:34Here we go.

11:34That must be it.

11:35Here we go.

11:36Found it.

11:38Copy.

11:40And good old Salesforce Two Factor.

11:43will will come in to challenge me like it should and I will have a number which I've got four seconds to enter go go go go go verify

11:54Yeah, we got the tick.

11:56My version of Tableau will go get licensed.

12:00Okay, we're licensed.

12:02We're legit.

12:03Yay.

12:04Okay, continue.

12:06There we go.

12:06Tableau.

12:07Um, superstore.

12:08I'll just open the the the sample workbook.

12:11Okay.

12:11This is one that's built for me.

12:13me okay and what we'll do is oh we'll just build it out won't we let's not let's not muck around so superstore what

12:24we want is subcategory where is subcategory da da da da da product subcategory we want it by year of order date interesting I don't know which year it used so that's actually quite

12:37Quite an important detail and then sales.

12:40Uh sales is there.

12:43Boom.

12:44Okay.

12:45And then let's do

12:48this we like this but go on let's you show me let's get these as lines and then we can see

12:59what it thinks this is maybe a quick way of doing this so it says subcategory fasteners so where is fasteners fasteners is the lowest performing one there it is right

13:10Right at the bottom.

13:11And then let's actually interrogate these numbers, okay?

13:15Let's interrogate these numbers.

13:17So it's saying in 2022 had 801.

13:24In 2023, 567.

13:27Correct.

13:28In 2024, 1001.

13:30Okay, it's rounding up 1193.

13:346161, 6162.

13:37Okay, all correct.

13:38Okay, all correct.

13:39It's even gone and got profit.

13:41We can test that as well.

13:42But this narrative that AI gets things wrong, like

13:46I'm sorry, but that's just not what's happened here.

13:49It's just not true.

13:50Um, but anyway, um, it's also doing the analysis, year-on-year growth, 29%.

13:56So okay.

13:56Let's see if that calculation is correct.

13:59Let's keep only fasteners, okay?

14:01And um what we can do is we can do uh table calculation, percentage.

14:08difference and we can see here that it's 29.

14:122% correct, 76.

14:156% correct, 550%

14:1815% correct, all correct.

14:21Okay.

14:22And it did it much faster.

14:24Didn't have to open Tableau for this.

14:25Okay, so that analysis is pretty good.

14:27And I asked a very specific question: tell me the lowest selling

14:30Subcategory.

14:31Very easy to check.

14:32You can validate this yourself.

14:34You can try it out.

14:35You'll have to set up everything that I've set up to validate this.

14:38So here you go.

14:40And then here's the thing: the key insight.

14:42Well, fasteners is the worst.

14:44seller overall it's showing strong growth momentum particularly in 2025 where sales jumped by over 500 percent interestingly the 2025 quantity 270 units is similar to previous years suggesting the sales increase

14:57came from higher price fastener products rather than volume growth.

15:01Like that is top-tier analysis.

15:05This is

15:07Like you might say, oh, that's pretty obvious.

15:09I could have figured that out.

15:11But the point is that not many people are good at asking questions in this critical way.

15:16And this makes that

15:17So it piques your interest.

15:19So you're like, huh, what is that?

15:21So would you like me to dig deeper into what's driving the 2025 spike?

15:26So yes.

15:27Um remember I should use my fancy tool.

15:29Yes.

15:30Uh dig in deeper and tell me more about what's driving this sales.

15:34Uh break it down so that I can understand which customers were ordering

15:38this product much longer prompt and uh I'll just hit enter I'll let it figure it out okay so I've taken that question I've asked an extra level beyond what it's at so I wouldn't just break it down actually I want

15:49to know which customers are ordering that product so it can understand a bit about that product but I haven't given it a customer type I haven't given it anything it's just analyzing and by the way when ai is doing this these like little steps these are actually just running python in the background

16:03To do the analysis.

16:05They're not trying to do sort of AI maths.

16:08It's actually literally creating the computation in the background.

16:11And the computation is

16:13accurate if that makes sense.

16:15The summary is what is being done by AI in this particular sense.

16:19So AI is sort of doing this thing of weaving these complex things together.

16:23to form this experience so the mystery is solved one anomalous order the 2025 fastness spike is driven by a single unusual order from Greg Gerthree

16:34Greg Gertrie, October the 5th, Staples 4, unit price 1299 total is that.

16:41This one order accounts for 80%, 84% of what

16:44all fasteners sales and it's being very factual giving me all the numbers normal stables pricing is this most of the orders are that grek is way over this is almost certainly a data entry error or possible

16:56possibly a different premium product miscategorized as basic staples so it's it's saying this is so out of the ordinary that we shouldn't be excited by this so the fact that first of all it started by doing this it dug into it and it was like actually no this

17:08doesn't look normal.

17:09The customer profile for Greg is the segment, the region.

17:12This is all the extra detail.

17:14I didn't even ask for the customer region, but it's gone ahead and got it anyway.

17:17Anyway, the real picture without the anomaly, if we exclude Greg's outlier, 2025 sales would be 962.

17:24Actually, below 2024 is 1002.

17:26The subcategory isn't growing.

17:28It just had one

17:29bad data point would you like me to find this as a potential data quality or explore aspects of the data set?

17:35I'm like I'm saving like oh my god

17:39Like this is good.

17:41Like this is very good.

17:42And if you don't agree with me, you know what?

17:45Let me know in the comments.

17:46Let me know in the comments and tell me this is not good.

17:49Okay.

17:50Let's dig into this.

17:51I'm not going to assume this is right.

17:53I know people will want to check.

17:55So let's go ahead and dig into this.

17:57So the way I do this in Tableau is I'd click on this data point and then we'd

18:00see um a bit of information now what i do need to do is go into the um individual line items and if we sort this from largest to smallest

18:10We should see here you go the 5199.

18:13So you can see that there this is actually this order, Greg Guthrie, 5199.

18:19So what I can do is I can show Phils.

18:22go and get the customer name so I'm doing this just inside of the data window so customer name and I can go ahead and get the date order date as well.

18:31As soon as I do that, I see Greg is there and there is his order there right there.

18:37Perfect.

18:37Okay.

18:38So this is like.

18:41Again, correct.

18:42So let's go ahead and remove Greg Guthrie from this uh data set.

18:48And we can do that by just filtering

18:50him out so let's go ahead I've I've validated that Greg is indeed there sorry about the size of Tableau I've got like a really large monitor in this case um if I go ahead and get the customer where's the customer customer customer customer customer

19:01Customer, customer, customer.

19:03Um where is it gone?

19:07Where is it gone?

19:08Why can't I see it?

19:09Customer name.

19:10Oh my god, I'm so blind.

19:11Greg Greg Guthrie

19:13Yeah, Greg GU3.

19:15There he is.

19:16We'll uh just uh exclude um hit apply boom there we go in fact

19:23if you take it away yeah let's see what it said let's see what it says if we exclude gravity's order the sales would be 962 actually below 2024 1000

19:36Subcategory isn't growing.

19:38So here, of course, we want to know if the number it got was correct.

19:43So it's saying it's down five, whatever.

19:45What we can do is we can just add sales onto

19:48uh detail here again and then put sales on the tooltip and then we can actually see that number here 946 oh so this is something interesting

20:00So we've taken uh Greg's outlier.

20:04We've taken all of Greg's order, uh Greg's orders out.

20:08What what I have to be sure of

20:11of though is does Greg have multiple I uh like order IDs have I taken that too much it might have actually done a much more

20:20accurate analysis and taking the order out but not greg's other orders which could have been correct so let's go ahead and take this out let's go ahead and do this so greg goo three

20:31Hit apply and then I can go ahead and put the order ID.

20:35Oh, Greg does actually have multiple orders.

20:38If we go and put the sales in there, we go and put the subcategory in there.

20:42This is almost definitely the issue.

20:44So let's go ahead and put the subcategory in there.

20:48We just want the fasteners.

20:49Okay, hit apply and hit again.

20:51Okay, and now we can see this is actually the order that I want to take.

20:57So I'm gonna take that, copy it, go back to

21:01my previous analysis, remove the customer name, and instead I'm just going to filter that one order that I just found that I know is erroneous.

21:09This one here we're going to exclude that, hit apply, hit OK.

21:12Okay, and what did I do there?

21:16Um what did I do there?

21:18What did I do wrong?

21:19What did I do wrong?

21:21Did I oh I unticked extra

21:23exclude so hit apply sorry hit okay fine that's uh user error now it says 961.

21:2975 961 962 which is again rounded up

21:34Bang on accurate correct.

21:37I don't think I need to say anything else.

21:40If you're using AI for data and analytics and you're not getting

21:44in the right results lean into it there is a way to get it to work and by the way i'm doing this on my laptop in a folder with excel okay i'm not gonna even save tableau i'm gonna close tableau

21:55this point.

21:56I've done a piece of analysis that would have taken me a lot longer.

22:01And by the way, I'm also doing a demo instructional everything.

22:05There is so much more that this tool can do.

22:07I have not used Cloed Core work before

22:09Before today, I literally just enabled it.

22:11And if you want to see what kind of impact that has had on my credits as an example, if you go to my usage, you'll see it has very marginal impact.

22:21And this is why I

22:22like it because actually if you'd run this um as part of um you know let me just yeah if you'd run this as part of like a normal tool you would definitely have maybe used

22:31five percent or something like that and so you wouldn't be really able to sit here and just keep going with this experience but um yeah I I I'm I'm really

22:41impressed with this and we could go on and on and on.

22:45So here's a new challenge.

22:48What if what if

22:50We gave it access to a whole folder full of tons of data.

22:56Okay?

22:57What if we gave it bookstore and the bookshop demo data set, which has to have a data model built?

23:03We have to get everything together and we're gonna ask the same questions we'd ask of that data set.

23:09For example, which authors haven't published a book?

23:11Yeah, the only way you could figure that out is to bring the data set together.

23:14Um who had the lowest sales across all the data set?

23:17It'd have to union the sales data, bring that in, merge it with the authors, and then merge it with the book and ask it to bring back the titles and the ISBN.

23:26so that we know it's got the other information from other tables.

23:29That is a much bigger challenge.

23:31We'll try that in tomorrow's video.

23:33Thanks for watching.

23:34See you soon.