Tableau Agents ( Einstein CoPilot ) for Web Authoring in Tableau Cloud | New in Tableau 2024.2

0:00Einstein Copilots, sorry, I mean Tableau agents are a new feature in Tableau 24.

0:052.

0:06Now, there's a lot to cover in this.

0:07First I'll cover the features, then I'll cover the naming.

0:10Let's get stuck in.

0:12So we're here in Tableau Cloud.

0:14This feature is only available in web altering.

0:16That's why we're starting here.

0:17If we go to the new option, go to Workbook.

0:20Um it loads up this view.

0:21We connect to our data for this one.

0:24I'm going to connect to my Tableau Pulse demo dataset.

0:28Let's go ahead and select that, hit connect, and boom, we're inside of the authoring experience inside of Tableau Cloud.

0:36Also, this feature is only available in Tableau Cloud.

0:39It's not coming to Tableau Server before someone asks

0:41Now you can open up Einstein copilot or tableau agent in two places.

0:46The first place is over here on the right hand side.

0:48You can see by clicking that icon it opens up the pane.

0:51And you're pretty much good to go.

0:52As soon as you select got it, it opens up that bottom pane here on the on the bottom.

0:57I kind of wish it was at the top, and it's sort of weird.

1:00You click up here for the icon and then you get sent down all the way here

1:04for this thing.

1:05Um I I think you can break UX a little bit.

1:08I know Chat GPT and everything sort of has this bottom up sort of window.

1:12I find that frustrating.

1:13I think you can be a bit innovative.

1:15If the icons there, put the text box here and just have have the feed updating below.

1:21That's fine.

1:21I can always read down, but moving my cursor down anyway.

1:25Bit of a nightmare.

1:25Anyway.

1:26The other way you can enable this is if you go into this calculated field option and you click on the icon, it also opens up this pane.

1:34But it does something slightly different.

1:35So let's actually get into it.

1:37Maybe there's a little clue there.

1:39You can see it's generated something.

1:40Let's get into what you can do with this capability.

1:43So the first thing I'll do is I'll highlight that this this Chat GPT window, this

1:48Einstein copilot window is contextual to each sheet.

1:53So the only way I found of clearing this view is to actually start a new sheet.

1:58When you start a new sheet, it starts a new conversation with Einstein Copilot, and there you have the context.

2:03So it's view specific or sheet specific in this exact sense.

2:07A view and tableau is something slightly different.

2:09The sheet is exactly what we're talking about here, and that's what we're looking at.

2:12So it's sheet specific.

2:14What can you do with it?

2:15Well

2:15The first thing you can do is you can click on suggestions and it analyzes your data set and comes back at you with some questions you can ask.

2:21If you're stuck, you don't know where to start, you can ask it some questions.

2:25When you click on those questions, it goes ahead and generates the chart.

2:28Let's see what it builds.

2:30And let's see how complex it is.

2:32Great.

2:33So we've got month of order dates, sales forecast, and sales.

2:36So it's built a line shop very easy.

2:38It's user dual access as well.

2:39Very complicated trick.

2:41Now I will just take a slight tangent here.

2:43and highlight that this this is already possible in Tableau of many many years ago and it's called something called Show Me.

2:50Show me can do this exact same chart.

2:52Let's just go ahead

2:53and show you how the only difference we show me is you need to pick the field so let's say you wanted to select order date our hold sales and our hold sales forecast and then you go to the top right

3:03Select show me and you need to be understanding of how show me works here.

3:07For me to see a dual access, I actually have to click this exact icon here.

3:11So let's go ahead and click that.

3:12And you'll see we get this now.

3:14The difference here it's used year of order day, but with one small change, I can actually go ahead and change this to month of order day.

3:21And that creates pretty much exactly the same thing as we've just seen.

3:24So if we go over here, that's exactly the same chart.

3:27So the only real difference here is that the year of order date is a default field that I actually selected.

3:32So when you select order date

3:33By default, it will pick year, but you can go ahead and change that manually so show me could already do some of the things we're asking, and I think it gives you a much clearer visual of what's going on, which made me wonder like

3:45Why isn't this just part of Show Me?

3:46Why aren't all the Einstein copilot tableau agent chart building things just in ShowMe?

3:54with a much much more interesting experience.

3:57It would be fascinating if as you hovered over these different charts, Tableau was sort of changing the charts in front of you, showing you how they work.

4:05And you could get a sense of how they're built.

4:07You know, really bring this canvas to life, really show people how to use the product.

4:11But nonetheless, um, I just want to sort of touch on that just to just to pull to the surface that

4:17What's really happening here is that you're typing text into window and it's using the rest of Tableau, some of the things we've already talked about to do what ShowMe was already able to do.

4:26That capability has existed.

4:28The difference is um the instruction and how it's being delivered to the product.

4:33Okay.

4:34Now that's not to simplify this because

4:36Yeah, Einstein Copilot can go a lot lot further.

4:39If I go back to my sheet two where we had a context already running, you can ask follow-on questions.

4:43For example, you can say, Hey, can you filter this to let's say 2022?

4:50and 2023.

4:52Now I've thrown a bit of a span.

4:54I've asked for two separate things using an AND clause in there.

4:57What I should see is two years worth of data.

4:59Let's see what it does.

5:00Remember it's still a dual axis, still keeping the context of what's going on.

5:04And it says, I'm having trouble with that.

5:06Could you try phrasing your request another way?

5:09I don't think I could.

5:10I think that's the simplest, simplest way to do it.

5:13And that shows an example of what's going on here with this.

5:17Tool, it's AI, it's gonna really struggle.

5:19That to me was a simple question, and let's say filter this to 2022 instead.

5:25This it should get now.

5:27I think the AND clause is what confused it there.

5:30And I think it was struggling to understand that I just want to see both of those two things.

5:34It's still having trouble.

5:35Okay.

5:36Now this has worked in the past.

5:38Let's say filter this to um

5:41uh south region.

5:45There you go.

5:46I've asked it for a uh value inside of dimension.

5:50So

5:50There's a region filled and okay.

5:53It's just not having any luck with me at the moment.

5:57Let me um

5:59Let me let me try a new session in a new window.

6:01So let's go ahead and let's go do this.

6:03Now because we built this chart ourselves, we could actually carry on.

6:07So let's let's let's start a new window here and say filter this to 2022

6:12Hit enter, let's see if it has any luck.

6:15Maybe Tableau Cloud has decided to go to sleep right now as I'm recording this very late in the evening

6:20I'm talking to fill the air.

6:21Come on, you're gonna work.

6:22I believe in you.

6:23I have faith.

6:24Yes, it worked.

6:25Okay, month of order date 2022.

6:28You can see that is there.

6:29Let's try something else.

6:31Let's say add

6:322023 to this two.

6:35Let's be very specific about that.

6:37So does it extend the filter to include 2023?

6:41Can it do that?

6:42Let's see.

6:44And that and for the record, this is not some sort of uh AI test, whatever.

6:49And what I'm trying to understand is look, how much natural language could this understand that analysts would use?

6:55And I think, you know

6:56Filtering is pretty standard.

6:57I think Tableau needs to iron out all of these sort of issues because I know it can do this.

7:01It's just it's just it's just struggling at this exact moment in time.

7:04That's not to take away from the feature.

7:06I think it's important to understand that sometimes these things work and sometimes they don't.

7:11Let's dig into that a little bit more so you can understand what I mean by that.

7:16Now the next thing you can do is you can ask it to help you with calculations.

7:20So

7:20What I'm going to do is I'm going to ask it to help me with this calculation.

7:23I've specifically written this in a way that I think an analyst would write it, but I know it's not going to get it right because of the way it's written to highlight an example.

7:30So

7:31Um can you write a calculation for me to find the

7:40average sales per customer by region.

7:42Now the reason I've written it like that is because an analyst would get that request and paste it in and write it exactly like that.

7:49But

7:49Here's the catch with this question.

7:51Average sales per customer by region is not a metric.

7:55It's a chart.

7:56I've actually described a chart and here's how to pass this out.

7:59Average sales per customer, that's the metric, by region is the dimensions I want to see this by.

8:06So what it will do here is I've thrown a dimension in that it can see, and what it's probably going to do is write a calculation to include the region.

8:13When the region is just a context for the chart.

8:15So let's see if this happens.

8:18And you can see that in writing calculations, what it should do is open up the calculation window, name the calculation, which is has correctly done

8:25It's actually given it nice camel case naming there.

8:28And then you can see it has actually fallen for that pitfall there.

8:32The bi-region bit shouldn't be part of this because the bi-region is basically what we're going to put as a dimension in the rows

8:39So if we were to fix this, we can go ahead and do this, and this will lead me to the next thing it can do.

8:44So let's go ahead and fix this.

8:45Let's remove this average from there.

8:47let's remove that from there let's remove region because we don't need it as part of our LOD if you're not familiar with LODs go check out my videos go check out SQL Bell's videos they're great uh if we hit apply

8:58We've got our calculation there.

8:59We've used it to help us do this.

9:01Now, if I highlight this and I say, look, okay, uh, can you help me explain this calculation?

9:10Calculation, you can ask it to understand the calculation.

9:15Now I type that in there in the chat window just to show you that it can look at the chat window, the sort of calculation window whilst you're working.

9:22This calculation finds the average sales for each customer.

9:25It does this by fixing the calculation of the customer name level, meaning it calculates the average sales for each individual customer, which is perfect.

9:31Now it's not actually the average sales.

9:33It's talking in metric terminology, but what it actually does is it calculates the average value of items in a customer's whole a record base.

9:43So if a customer has had multiple orders,

9:45it will go in and find the average value of all the items that are in that record because this data source is at a granularity it is not just orders is actually products in an order and you can buy multiple products

9:57of the same type in one order, hence that detail.

10:00So we've got that metric uh there now created.

10:03What we'll do is we'll go ahead.

10:05and put it in and we'll get our number and Tableau's gone and summed it up, but what we can do just to make this chart be more correct is

10:13put a customer name and I think what we'll do is we'll I'll put that in a really bad place.

10:18I should have put it there.

10:20Apologies.

10:21And then what we can do is rotate this 90 degrees.

10:23and then go look and see what the highest value is.

10:26Let's let's keep Mitch let's keep Mitch in this data set.

10:31There's Mitch.

10:32We'll go ahead and put the value on label.

10:34So you know, I'm just playing around with this.

10:36I'm gonna put region and this is what I was saying earlier.

10:40Look, by region.

10:41So customer by region is basically what the question was.

10:45And um see it's actually the same value, and that's because well my LOD is written at the customer level, so if I wanted to make this work,

10:54Then what I should have asked is look, can you calculate the average sales per customer region combination basically?

11:03So you're you're you're pairing customer and region to create this the sort of context.

11:08We can get this correct in our LOD going in here, select edit and actually adding this back.

11:13So it fix that the customer

11:15region level and let's go ahead and add that calculate the average sales if you hit apply you'll see that number now changes to reflect that so

11:22Now this is working.

11:23That's how you debug this calculation to make sure it's doing the right thing.

11:27If you want to, you can even go as far as going into the record.

11:31to find out what's going on at a summary level, but also looking at the full data.

11:35Now the full data will only show you what's actually available in the dataset in the current moment, but you can go ahead and show additional fields if you wanted to

11:44If you can bring in uh let's say the product name, you can see that these are two different products and not the same product.

11:49Um you can go look at the quantity.

11:51There's seven of that, there's three of this one.

11:53So there's a little bit of um, you know, I think investigation you have to do with all of these calculations, regardless of whether it's written them or not.

12:00The charts are sort of easy to see if it's done them correctly, but the calculations I think you do have to uh start writing them.

12:06Anyway, that is

12:08How it works.

12:08There's a whole bunch of other things to cover, but look, if you want to understand the full context of this, um, there is good documentation of this, and I'll come back to why it's still called in I'm still in co-pilot here in a second, but uh nonetheless

12:23You've got all the things it can build, all the charts it can build, the way the data works with Einstein Trustlayer, the different types of file formats that it can read, how to launch it, basically a lot of the things I've covered in this video, along with

12:36I think it has a tips section here at the bottom that gives you really good ideas for the prompt you're going to type so you can make sure it's really, really tight as well as some unsupported features, things it can't do just yet.

12:48Um and then there's also an FAQ, which is kind of nice.

12:51So it kind of goes into some of the some of the things your colleagues or you know people in your business might might ask.

12:59Um, its compliance, its sort of capability, how it works, who the LLM providers are, all of that is nicely detailed.

13:07Which leads me to, I guess, the final bit, which is let's talk a bit about the branding of this.

13:12Now, this is called Tableau Agents as a result of Dreamforce.

13:15And Dreamforce in that announcement just now two weeks ago

13:20They came out with this idea of something called agent force.

13:24And so this idea that AI is, you know, fine, copilots are fine, but actually we need to go to the next level where we have

13:30autonomous AI-powered agents doing things, exercising demands and requests based on user workflows on people's behalves, essentially amplifying the effort that people put into their daily jobs

13:42And this is great, but the the the fundamental problem with this is that uh the the brand change happened faster than pretty much the release of Einstein Copilot.

13:52So

13:52uh Einstein Copal I came out in 242.

13:55It been around in beta for a little bit longer.

13:59And then literally within I I'd say a month and a half, it changed branding to Tableau Agents.

14:05And you can see throughout this whole video

14:06There's bits where it's still called copilot, there's bits where it's called Tableau Agent.

14:10And that's that's something that just needs to get ironed out because I think people will search for this, they'll find copilot a lot.

14:16And Tableau talked about Einstein.

14:18in so many different contexts, so you will no doubt for the rest of the end of time find, you know, conversations about Einstein that could be mistaken for this.

14:27Anyway.

14:28The other thing is, I think Agent's a really bad name for this.

14:33I'm sorry to say this, but

14:36You know, I get the concept of agents within Salesforce.

14:39I totally get that.

14:40Because in Salesforce, that kind of ecosystem, a lot of the workflows in Salesforce itself.

14:46do speak to this sort of idea of agents, you know.

14:49Um the a lot of the demos they gave suggest you might call in look for a service desk and have an agent

14:55Check inventory and do a bunch of things.

14:57There's a level of autonomy that is expected to complete that workflow.

15:01But when you're in Tableau in the web authoring experience,

15:05You're supposed to be a wingman or wingwoman of me, the author.

15:10You're not supposed to be some autonomous thing going off and doing stuff in the background.

15:14There is no

15:15agent-like capability in what I've just shown you.

15:17It's more of an actual copilot.

15:20And to me, that's sort of the irony.

15:22I don't know why Tableau couldn't have gone down the route to say, look, we get agent force, we get agents

15:28But there is a world, uh hear me out, there is a world where agents and copilots can coexist.

15:34An agent in Tableau to me would be something that goes out and picks up a bunch of metrics.

15:40that it thinks you need to look at based on activity, it's learnt from you, behaviors it's learned from you, and surfaces them to you with some sort of explanation and narrative.

15:49It's a bit like um Google's notebook LLM where you can give it a script, you can give it a bunch of audio, a bunch of text, and it will create a podcast based on that.

15:57That to me is more of an agent, you know, surfacing things that it thinks are important for you based on your interest.

16:02Helping you.

16:03A copilot helps you build things faster, helps direct you, helps you um empower you.

16:09And that's that to me is such a much better branding.

16:11I know Salesforce wanted to move away from this in Microsoft.

16:15co-pilot, whatever thing.

16:17But I still think like well if you're gonna do that, you really do need to sort of understand what that word means.

16:23And I think if you go and Google and search, you know, agent versus

16:27um, you know, co-pilot, you'll find lots of good discussions about these.

16:31And and everyone's going towards this world of agents in the AI space.

16:34But I do think, you know, if you look at developers, if you look at a lot of people who are writing code with uh ChatGPT

16:40It's much, much more of a copilot.

16:42You're still in charge, you're still driving.

16:44You saw that I had to at times interrogate what the AI had done.

16:50uh correct it and that is very much what a pilot does versus a copilot, right?

16:55It's kind of our irony in that the experience I showed you was more of a co-pilot pilot experience than it was an agent

17:02Because if if it's an agent and all of that stuff happened autonomously, you'd end up with a whole, you know, bunch of junk on Tableau Server, Tableau Cloud.

17:10Anyway, it won't come on Tableau Server.

17:12So I don't know why I mentioned that, but anyway.

17:14I think that's my two piece on that.

17:16I just think this wasn't that's just something not right.

17:19And you know, it goes even further because I don't know, this is sort of not that I'm a journalist or anything, but

17:24It feels like there's a bit of a control find replace on Tableau Agent.

17:29Like everything that was previously called Einzang Copilot has been rebranded to just be called Tableau Agent Everywhere.

17:35Like agents weren't didn't exist in April 2024.

17:39Uh and here they are talking about it, right?

17:42And even if you go into the um article from back then,

17:46Um you'll see they've gone and done Tableau Agent Former Einstein Copilot basically on a bunch of articles.

17:51They've gone and done a whole sweep and audit and sort of changed things.

17:55And but a lot of the you know marketing uh unfortunately for this is still talking about Einstein Copilot.

18:01Like if you open this up, you can see it's still got Einstein in there.

18:05Now Einstein will remain as the branding because Einstein is sort of the AI machine learning brand within Salesforce.

18:15Einstein is in Salesforce.

18:17Einstein copilot was like copilot.

18:20Salesforce branding, man, it's just so bad.

18:23It changes faster than they can build a product, which is sort of ironic.

18:27But anyway, anyway, I think I've said enough about this.

18:29We've been going for nearly 20 minutes.

18:31Um I'd love to know your thoughts.

18:32I'd love to know what you think you're gonna be doing with Einstein Copilot.

18:36I'm really intrigued to see where this is going.

18:39And I think I have to call out this is just version one.

18:42There are other copilots in Tableau Prep.

18:44There's other copilots in data lineage.

18:46So I'll make separate videos on that to show you sort of how they work and what you can do with them.

18:51But hey, it's the first video in a while.

18:53Um I just wanted to cover this because it's something that I know

18:56is fairly topical.

18:58Let me know what you think in the comments.

18:59I'd love to know what you think of this.

19:02Are you going to use it?

19:02Don't forget you need Tableau Plus in order to use this as well.

19:06So I do think the user base for this will be limited at first pass.

19:10Um until everyone catches up with the availability and they start to see some actual value coming from this.

19:15Thanks for watching.

19:16I'll see you in the next one.