Video |

Tableau | AI & ML | Industry trends | Data prep

Tableau GPT - Ai in Tableau Prep, Tableau Pulse & more | Announced at Tableau Conference 2023

Tableau GPT is exciting, but in an analytical context it's all good and fine until it gets something wrong, really badly.

Tim Ngwena •14 May 2023 •2 min read •7 min watch •17K

Part ofTableau Conference 2023 Tableau Prep Builder tutorial Jan-2020 Tableau Pulse

Watch on YouTube

Tableau GPT is built on a large language model trained to understand both Tableau's terminology and analytical context within organisations, sitting as an underlying layer across the platform.
AI capabilities were demonstrated in three places: the search bar for metric questions, Tableau Pulse for generated prompts and open-ended queries, and Tableau Prep for tasks like extracting an email from a JSON field via regex.
The model is likely trained on Tableau's knowledge base, documentation, forums and telemetry from Tableau Public, making its responses richer and more context-aware.
AI is strongest for prompting direction (what to ask, what to look at) but riskier for mission-critical work, where wrong answers can have serious business consequences.
Generated calculations should come with documentation and performance implications, ideally nudging you to push heavy logic back into the data stack rather than into Tableau itself.

0:00Hey, it's Tim here. Tableau just announced

0:02Tableau GPT. And in this keynote roundup I

0:06recently just

0:07launched, I explained that Tableau had to

0:09talk about GPT and AI technology, because

0:11every single

0:12company is talking about it. Let's find out

0:14more and get into it. Let's get started. So

0:17Tableau had

0:18to talk about AI, there's just no getting

0:19around it. Every single company that's

0:21invested in the

0:22technology space has to be talking about AI

0:25capabilities. AI has just mothballed

0:28through

0:28every single technology company. And

0:30although they're all still trying to figure

0:31out how it

0:32works or the different ways, the general

0:34consensus, and I'm not sure I'm totally in

0:37line with this,

0:38but the general consensus is every company

0:39thinks there's a way they can use this

0:41technology

0:42to enhance their products. And so what

0:44Tableau had to do is actually highlight the

0:47fact that

0:47they're not just jumping on this bandwagon.

0:49In fact, Tableau has been, and I've done a

0:51few

0:51videos already showing you how Tableau has

0:54already been using AI technology in all of

0:56their tools.

0:57And so this was a really sort of good way

0:59to set this up because they then announced

1:01something

1:01called Tableau GPT. Now GPT stands for

1:05generative, I can never remember this,

1:08generative something

1:09transformer. We're gonna have to literally

1:12come out of this recording, generative pre-

1:14trained

1:14transformer. Okay, cool. Now GPT stands for

1:18generative, generative pre-trained

1:21transformer.

1:22I had to literally go and Google that

1:24because I forget it every single time.

1:26But that technology was actually invented

1:28by Google some time ago, and we've actually

1:30had

1:30multiple versions of GPT. You might have

1:33heard of chat GPT version 4, which is based

1:36on GPT 4

1:37technology. Well, GPT 1, 2, and 3 have all

1:39come before it. You might have been

1:42familiar with GPT

1:433.5, which is what OpenAI introduced

1:46alongside chat GPT. So what it essentially

1:49means is that

1:50Tableau themselves have been building a

1:52large language model that not only

1:53understands the kind

1:55of terminology that we might use with its

1:57product, but also will better understand

1:59analytical context

2:01within organizations. And so this

2:03technology is going to be an underlying

2:05layer throughout their

2:06platform. If I sort of just jump ahead a

2:08little bit, you can see there's this sort

2:10of graphic,

2:10and I think this graphic explains it really

2:13well. Large language models and AI

2:16technology really do

2:17have to work across the platform because

2:19they need to sort of have a hook into lots

2:22of different

2:22metadata sources in order to really

2:24understand what people might ask. And then

2:27throughout the

2:27keynote, we saw different examples of how

2:30this works. The first example was obviously

2:32just going

2:33into the search bar and asking a simple

2:35question about a metric, and we saw a

2:37response. The second

2:38one was in Tableau Pulse, a Tableau Pulse

2:40being this new place for metrics to live, a

2:43new way of

2:43building metrics as well, not necessarily

2:45attached to data visualizations. And what

2:48we had there were

2:49prompts generated by Tableau GPT,

2:51essentially questions you could ask of the

2:54data source,

2:54but also you had a search box which could

2:57understand your queries, which could be

2:59relating

3:00directly to the data source, or you could

3:02even ask open-ended questions. For example,

3:04what else

3:04should I know about this particular product

3:07? They actually did a demo and showed us an

3:09example of

3:09that as well. The other place we saw it

3:11inside of the product was inside of Tableau

3:14Prep, and I think

3:15this is the use case that really we are all

3:17familiar with because this is exactly what

3:19we've

3:19been using ChatGPT for. They showed an

3:21example where they went into Tableau Prep

3:24and they had

3:25a field of data, and the field had emails

3:27contained within it. And I think it was

3:30JSON object,

3:30it was a complicated sort of data format,

3:33but it was all stored in one column. And so

3:36what they

3:36were able to do is ask Tableau GPT, "Hey,

3:39what regex can I use to extract the email?"

3:42or "How can

3:43I extract the email from this field?" And

3:45it came back with a regex response. I don't

3:48think the user

3:48actually asked for regex, it just came back

3:50with it. And so that's another use case. So

3:52if you can

3:53imagine that inside of the calculation

3:54window inside of Tableau Desktop, you're

3:56trying to solve

3:57a problem, just being able to simply go to

3:59Tableau and say, "Hey, I'd like to write a

4:00calculation

4:01that looks back four months." That's a

4:03classic example, right? Even though metrics

4:05now looks

4:06like it has the interface to build that for

4:08us, having this technology inside of Table

4:10au GPT is

4:11going to be great. And you could argue it's

4:13probably been trained on data sets, for

4:15example,

4:15the knowledge base articles, the

4:17documentation that goes back many, many

4:19versions. You've got

4:20Tableau forums, you've got knowledge base

4:23responses to specific known issues. All of

4:26these

4:26are really super powerful sources for this

4:28kind of technology. And if they've been

4:29training their

4:30technology correctly, they can start to

4:32pair that with telemetry because of course

4:34telemetry tells

4:35them how people are using this alongside

4:37what's going on inside of Tableau Public,

4:39which is another

4:40big source of telemetry for Tableau. Pair

4:42ing all of that together gives you something

4:44really rich.

4:45The difficult thing though is making sure

4:47that when it gives you an answer, it's not

4:49only useful,

4:50but it has to be right. And this is sort of

4:52the tricky point here. It's all good and

4:55fine to get

4:55excited about this technology. But, and you

4:58all know this using Chat GPT, when it gets

5:01it wrong,

5:01it has the potential to get it wrong really

5:03, really badly. And in an analytical context

5:06,

5:06in the business context, that can sort of

5:08have issues. And so where this technology

5:10is great is,

5:11you know, what questions should I ask? You

5:13know, what should I look at? That kind of

5:15stuff,

5:15the kind of stuff to prompt you in a

5:17direction, move you in a direction, give

5:19you some starting

5:19points, give you some visual aids, that

5:21kind of stuff is totally fine. But when you

5:24start to go

5:24down the road where you're writing

5:26calculations with it, that's great. But you

5:28also want a little

5:29bit of documentation on that calculation so

5:31that you understand exactly what's going on

5:32in that

5:33regex. And as you get closer and closer to

5:35sort of mission critical parts of your

5:38analytical stack,

5:39you want to know that it's going to

5:41actually help you get the right queries,

5:43the right responses,

5:44and even things like giving you the

5:45performance calculations. We've now got

5:47capabilities in Tableau

5:48that can tell you what's performing and

5:50what's not. We want this technology to give

5:52you the

5:53sort of criteria to say, hey, write a

5:55performance calculation. Or in that example

5:57with regex,

5:58you get a response saying regex, but it

6:00then tells you these are the performance

6:02implications. Have

6:04you thought about pushing this capability

6:06further back into your data stack, so you

6:08can do some a

6:08little bit of data modeling engineering,

6:10whatever you want to call it, and get this

6:12as a field

6:13ready prepared inside of Tableau rather

6:15than doing this here. That's the kind of

6:17sort of intelligence

6:18you need, especially when you're running

6:19big analytical workflows. So I'm super

6:21excited about

6:22this technology. I can't wait to see more

6:24of what it can do. But it's still early

6:25days, we only got

6:27a real small glimpse of how it's working. I

6:29can't wait for the first release of this

6:32inside of what

6:32will probably be a Tableau Cloud, but

6:34nonetheless, I can't wait. Let's find out

6:36more. And until then,

6:37I'll catch you in the next video.

6:38Okay.

6:39Bye.

6:48[ Silence ]

Rollup

Tim breaks down what Tableau GPT actually is and where it shows up across the platform, and flags why analytical AI needs a higher bar for accuracy than a general chatbot.

Tableau announced Tableau GPT at Tableau Conference 2023 as an AI layer running across search, Tableau Pulse and Tableau Prep. Tim reacts to the keynote demos and what they imply for day-to-day analytics work.

The Breakdown

Why Tableau had to talk about AI 0:18

Every technology vendor is under pressure to show an AI story, and Tableau used the keynote to argue it isn't just jumping on the bandwagon but has been building AI capability into its tools for a while.
What Tableau GPT actually is 1:02

GPT stands for generative pre-trained transformer, the same family of technology behind tools like ChatGPT. Tableau GPT is Tableau's own large language model, trained to understand both Tableau-specific terminology and the analytical context of an organisation, and it's designed to sit as an underlying layer across the whole platform rather than as one bolt-on feature.
Search bar metric questions 2:27

One surface is the search bar, where you can type a plain-language question about a metric and get a direct response back.
Tableau Pulse prompts and open queries 2:41

In Tableau Pulse, the new home for metrics not tied to a specific visualisation, Tableau GPT generates suggested questions to ask of a data source and also lets you type your own queries, including open-ended ones like asking what else you should know about a product.
Tableau Prep regex generation 3:00

In Tableau Prep, you can describe a data problem in plain language, such as extracting an email address from a complex field, and Tableau GPT returns a working regex solution, even suggesting techniques you didn't explicitly ask for. The same pattern applies to writing calculations in Tableau Desktop, such as asking for a formula that looks back four months.
What it's probably trained on 4:13

Tim reasons that Tableau's knowledge base, documentation, community forums and known-issue responses are all strong training sources, and pairing that with telemetry from how people actually use Tableau and Tableau Public would make responses richer and more context-aware.
The accuracy caveat 4:46

An answer has to be useful and right, and in an analytical or business context a wrong answer can go badly wrong. AI is safest for prompting direction, what to ask or where to look, and riskier the closer you get to mission-critical calculations.
What good AI-generated calculations should include 5:22

When AI writes a calculation, such as a regex, you want documentation explaining what it's doing and ideally a note on performance implications, including whether the logic should be pushed back further into the data stack rather than handled inside Tableau itself.

Worth Knowing

The keynote only showed a small glimpse of the technology in action, so much of how it performs in practice is still unknown.
AI-generated regex or calculations came back without documentation, which is a gap you'd want addressed before relying on them for critical work.

Use It When

Reach for this kind of AI assistance when you need direction, starting points or a first-pass calculation, but treat it cautiously and verify anything that feeds into mission-critical analytical work.

How this Rollup was made provenance & method

A Rollup is drafted by AI from the video's transcript, then reviewed and edited by Tim. Everything used to produce this one is listed below — the model, the exact prompt, and the source video — so the process is transparent and reproducible.

Transcription: On-device — NVIDIA Parakeet v3 for recent videos, OpenAI Whisper large-v3 for earlier ones. The transcript never leaves the machine or gets published.
Drafting: Claude Sonnet 5 in the cloud, from that transcript.
Prompt: The exact Rollup prompt (v2) — the full system prompt, unedited.
Source video: Watch on YouTube
Drafted: 5 July 2026 at 09:38
Reviewed & edited: 5 July 2026 at 09:41 · by Tim Ngwena

Model + prompt + video is everything you'd need to recreate a Rollup like this yourself. The one thing we don't share is the transcript.

Rights. The video and its transcript are the property of TN Media Ltd. Unauthorised use or download is prohibited. © TN Media Ltd.