Hello

This week I spoke with Klaus and Merlijn from Infotopics on The AI Analyst. You might know them as the team behind Apps For Tableau. Now, you need to think of them as the creators of a new platform, Trulyp. Here’s 3 things I learnt from the episode.

First, what’s coming up:

---

Thu 3 Sep · 4pm BST; The AI Analyst

📺Next time on on The AI Analyst? “Make with AI. Make mistakes with AI.” with .

We’re going to go philosophical: how is AI changing how we build, what we build and who we build for. Elijah has innovative, sometimes controversial opinions. I can’t wait to dig in. Register here.

---

Thu 10 Sep · 4pm BST: Register here.

📊Then it’s Chart Chat. Are you familiar with the latest academic research into our field? We’re going to share our favourite recent papers and you’ll learn how they can change the way you work. Register here.

Trulyp Studio changes who can build a custom Tableau extension

Trulyp includes an extension generator: you describe what you want in plain language, and it produces a working, deployable Tableau extension, in about two minutes. You want a multi-page dashboard, as an extension? Just describe it and it becomes an extension. Merlijn describes this in the video below:

If you also watch the demo section (YouTube), you’ll see that Infotopics have, as they always do, paid great attention to detail in the way this works: the prompt can get you most of the way there, but Trulyp also gives you full settings control through a traditional GUI.

If you’ve always shied away from the complexity of building a Tableau extension, now you don’t need to.

These extensions can also live outside the Tableau ecosystem, so you’re not even tied to forcing users to go to the Tableau Server. Trulyp can stand alone.

Subscribe now

The Analyst’s Role Is Undergoing A Revolution

Klaus’ thoughts on how our roles are changing is important. Like Klaus, I learnt my skills by spending hours (days?) chipping away in Tableau, searching for the best articulation of my data, the perfect “trick” to show things just so, the ideal way to focus on my story.

But now? If we can code anything, why spend time in such nitty-gritty detail?

I agree with Klaus with his diagnosis. In the episode, I then asked “But, Klaus, what does it mean if nobody learns the fundamentals anymore?”. And we ran out of time to answer it. 😜 What do you think? Do leave a comment with your thoughts!

Can you trace your AI’s decisions?

As someone who openly admits to falling asleep when talk turns to governance, I did get somewhat overexcited at Trulyp’s auditable AI tracing. I love this feature. You can go to any AI conversation within Trulyp and fully trace the steps it took to reach a decision.

Trust in AI analytics is one of the hardest things to build. Could this be the way forward?

Final Thoughts

Trulyp is attempting something the industry has been trying to do for decades: make powerful analytical tools accessible without dumbing them down. The GUI controls alongside the AI prompts is a great sign that they’re not pretending the prompt is enough. “AI on your terms,” is how Klaus describes it.

I’m still sitting with the question I asked Klaus that we didn’t have time to answer: if AI can do the craft work, what happens when people never learn the craft itself? I don’t know the answer to that yet. If you do, drop a comment, I’m curious.

And finally, do join me in the next couple of weeks:

• Thu 3 Sep; The AI Analyst“Make with AI. Make mistakes with AI.” with .

• Thu 10 Sep: Chart Chat: From Research To Practice

That’s all for now. Thank you to Infotopics for sponsoring this episode. Find out more at: https://trulyp.com

Take care,

Andy

A month ago I described four problems with chat box-based conversational data analytics. Just because OpenAI gave us a chat box interface to its LLM, it shouldn’t have meant that AI analytics followed the same lead.

First, a promo for The AI Analyst on Wed 26 Aug.

I’ll be with Klaus and Merlijn from Infotopics, as they explain why they’re building Trulyp, bringing all their Tableau extensions into a new, AI-driven platform. Register here.

The four AI analytics UX problems I highlighted were:

• Conversations with data aren’t linear

• Single threads give you one answer at a time

• There’s too much eye movement to different places

• Data gets squished in conversation panes.

Since that post went out, I’ve spoken to several vendors and dug into a few more on my own. People are thinking hard about these problems, and there are novel ideas to solve the problems. While none have solved everything, the direction of travel is encouraging. Here’s what I’m seeing, organized by the four problems.

Problem 1: Data conversations are non-linear

(Reminder: data exploration is “squiggly”; it is the ability to jump back and forth, up and down, in and out as you seek the best articulation of the data. Long, linear conversations are awful to navigate)

“Follow up”

Veezoo’s “Follow Up” feature lets you ask a specific question about any chart anywhere in a conversation. The follow up then appears at the end of the main conversation flow. Instead of having to use text to refer to a previous interaction (“Can we go back to that chart about sales by country”) you can force a new “thread” directly from the chart.

Touch-to-edit

Another solution is the ability to instantly switch from chat to direct interaction with a visual. At this point, if your AI platform does analytics and returns static images, you’re behind the curve. This is where the main LLM platforms (ChatGPT, Gemini, Mistral, etc) are: they mostly only return text summaries and static images. It’s like looking in a shop window but being unable to go in and browse.

If you give me a chart inline in a conversation let me interact: filter, edit, switch, rearrange. Veezoo lets you change filters and ask follow ups, Thoughtspot’s Spotter lets you change the tokens and filters, Hex lets you expand a chart into a detailed explorer view.

This matters for non-linearity because you can examine the output without relying on the text box. You can stay in the flow, adjusting and exploring, in whatever way feels most appropriate.

Going back in time

Golden Analytics has built a visual history timeline that lets you see your session history and jump back to any previous point with a click. It’s simple and well executed. This is a native part of the platform, i.e. not just when you use AI conversations, but it shows a way to skip forwards and backwards very quickly. It also aligns with Golden’s hybrid approach to AI: the “Slider of Autonomy”, more of which below.

Is it a chat? Is it an app? No, it’s Hex.

Hex has a different approach. Every conversation is also a Notebook. And an app. And it has a searchable table of contents. This is the most direct fix I’ve seen for the “scrolling back up to find that earlier thing” problem I described in Part 1.

For example, if you’ve done a long chat about, say, marketing campaigns, you can switch from chat to notebook and search all the cells for a particular moment in the conversation. That section can be repurposed and reused elsewhere.

Problem 2: Single-threads are a let down

(Reminder: an LLM could provide infinite answers. Most chat bot experiences provide one at a time)

I think this is one of the hardest to solve. Do we want threads? Do we want something akin to a chatroom with multiple responses to each question?

Can we build threading?

Researchers Matt Hong and Ana Crisan explored threaded exploratory data analytics as a design pattern. They designed a system that let users branch questions into individual threads. Their research found that users still ran into frustrations with this approach, which is worth knowing. They struggled with managing branches and global context. Threading feels like a promising approach; research shows it is not yet a silver bullet.

“Chat rooms”: give me ALL the threads

One of the most interesting innovations I’ve seen was the winning entry in Veezoo’s Analytics Cup this summer. Entrants were provided a comprehensive World Cup dataset.

The winning entry by Oliver Yang solved this problem. He made a system of 6 different analytics agents who would investigate a given question. His agents were tuned as cynics, romantics, historians, tacticians, etc. For any given match, each agent would examine the data and return a story based on their configuration. I love this. Just like real life, one question generates a whole different set of valid opinions.

Problem 3: Don’t make me move my eyes

(Reminder: The sidebar AI paradigm (data on the left, chat on the right) forces your eyes to travel constantly. Eye movement is expensive: it flushes your short-term visual memory, and when you go back to your chart, you’ve already forgotten what it looked like before the change.)

The solutions here cluster around one principle: bring the AI closer to the data.

Link the displays

Ridge does have a side pane to let you “chat” alongside your dashboard. But they’ve done something cool: the chart and the dashboard connect and instantly update as you interact.

Whenever you interact with the dashboard or a chart in the conversation pane, all displays filter, immediately. Ellie Fields at Ridge describes it as “the best of dashboards and AI chat, together.” It’s a really fluid and intuitive approach to the challenge.

Put the AI where my eyes are

Do you use Microsoft Office? You know how the contextual tooltip bar appears whenever you select some text? It’s such a great little UX trick: you’ve just selected the text, so the app knows your eyes are literally looking at the selection. Therefore, why not show the most commonly used formatting options right there, rather than send the user’s eyes and mouse up to a toolbar? It’s one of my favourite UX features of all time.

And it’s one I called for when I spoke to the Tableau dev teams about the UX problems of AI Agents.

Golden has implemented the AI chat box directly adjacent to the relevant chart — not in a global sidebar, but right next to the data you’re asking about. You simply click on what you want to change and describe your intent. It’s exactly what I requested in my talks with the Tableau team back in 2024. I’m so happy to see this; I think it’s a great solution.

Subscribe now

Problem 4: Data gets squished in conversation panes

(Reminder: Narrow side panes for AI-driven data chats create unreadable charts)

This is my shortest section. Why? Because the solution is simple: give data more space.

You can make the AI agent the full focus. Thoughtspot and Veezoo have this approach. Just like ChatGPT, Claude or Gemini, the chat is the window.

Hex has a tabbed layout that lets you toggle between full-screen chat, or full-screen notebook with the chat in a collapsible panel. You’re never trying to do both things at once in half the space. You can choose and change as you wish.

Beyond the chat box

Everything I’ve discussed so far assumes the chat box is here to stay and the challenge is to make it better. What if the chat box is just a bridge technology?

Golden’s Francois Ajenstat proposes a “Slider of Autonomy,” inspired by Andrej Karpathy. Instead of optimizing the chat UI, why not reduce our reliance on it? Francois argues that AI should be embedded behind buttons and interactions, allowing users to choose how and when to invoke it. If you want to click, drag, and drop, you should; if you need AI, it should be there, immediately, contextually ready.

After publishing my first post on the UX problems of AI, Nikita from Supersimple got in touch. He showed me what they’re building. They’re moving past the chat concept entirely, opting for an infinite canvas of blocks (charts, tables, and variables) that function as a workspace. Here, the “conversation” isn’t a separate pane; it’s the record of changes you’ve made to the canvas.

No single vendor has cracked all the problems; we’re too early in this new world to have worked out all the solutions.

The innovation I see is mostly from startups, not incumbents. They’re the ones with the least amount of legacy code, giving them flexibility to adapt quickly as we all learn how to best bring AI and LLM technology to data analytics.

I’m excited to keep watching the industry. If you’re building something in this space and have an idea worth sharing, I’d love to hear from you.

Thanks to the teams at Veezoo, Hex, Ridge, Golden, Thoughtspot and Supersimple for sharing what they’re working on. And to scholars exploring this new frontier.

Finally, do join me on 26th Aug for my next episode of The AI Analyst as I talk with Klaus and Merlijn from Infotopics about their new Trulyp platform.

I hope you’re well.

I’m getting ready for a busy late summer and autumn. Here’s where you can find me. At every event, my aim is to synthesise and share the latest trends in this wild world of data.

The AI Analyst: Why We’re Building Trulyp

On Tuesday 26 August, at 4 pm BST (11 am ET), it’s The AI Analyst? Episode 8. I’ll be joined by Klaus and Merlijn from Infotopics as they explain why they’re building Trulyp.

Trulyp “brings all trusted Infotopics extensions together with AI, generative apps and data insights.” Sounds good, but I will be digging much deeper than that! How are they building to aid governance, to embrace AI but not overload us with it, and to help all their customers?

Register here.

The Duel of the Data Experts, Big Data London

I’ll be at Big Data London (23-24 September) and hope to see you there.

I’m very excited to be hosting a new event: The Duel of the Data Experts. We’ve got 6 leading names in the industry, and divided them into Team Optimist and Team Pragmatist. They will debate 3 hot topics. YOU decide the winners!

The topics are the big dividing lines in our industry right now: governance, interfaces and humans. Is AI the panacea we’re promised? Or is it just another tech that will overpromise and underdeliver?

I’m really excited about this. It’s been an idea I’ve had for years and am grateful to the Big Data London team to bring it to life.

See you there.

Communicating Data for Impact

If you’re in or near the north of England, do come along to Data Decoded. I’ll be keynoting again and delivering my “Communicating Data” workshop on day zero, Monday 12 October. The workshop is a steal at just £295 per person. Data Decoded itself is… free. Yay!

I am building out a great season of shows throughout Autumn, too. Do subscribe to stay informed on each of these

Subscribe now

• Chart Chat (this is our provisional lineup, you can always find the latest at

  https://www.chartchat.live/)

  • 9 September: 3 Academic Papers You Need To Read

  • 15 October: Topic TBD

  • 5 November: Election Viz: a review of the charts from the US mid-terms

• The AI Analyst

  • 26 August: Why We’re Building Trulyp

  • 3 September: “Make with AI. Make mistakes with AI” featuring Elijah Meeks

  • 1 October: What’s happening with the Tableau ecosystem? featuring Kirk Munroe

• Keynotes

  • I can’t announce anything yet, but there’s more in the pipeline for speaking opportunities

Take care, and thanks for reading.

Andy

My main flows are for the US Wildlife Strikes and Road Fatalities datasets. They’re somewhat sober datasets, but each has an excellent mix of dimensions and measures, geo- and time-based trends, and lots of valuable stories to reveal. Each one is updated regularly. [Note: both these also use CSV inputs and CSV outputs; I’ll discuss enterprise data sources later]

My data flow needs to continue, but the technology needs to change. Much as I have loved Tableau Prep over the years, it’s time to move on.

Prep flows are machine readable

Subscribe now

I turned to Claude Code.

Me: “Could you convert a Tableau Prep flow to a Python script? I want the Python file and a whole suite of tests to validate the script.”

Claude [paraphrase]: “Why, of course. A Prep .TFL file is just a zip file. Inside there’s a JSON file describing the entire flow. Why don’t I make this a skill for you?”

We began with the latest Road Fatalities data. Claude got to work building and testing the skill. On its first pass, 155k rows were wrong (out of 750k) when compared to the output from the Prep file. Second pass: 8 rows. Third pass: 0 rows. It was a byte-perfect match.

What this is actually about

A data pipeline is tied to one platform’s logical steps. Every remap to a new technology is hard and validating success is vital. The process of making this skill identified issues that were introduced as the data evolved year by year. Those errors weren’t with the flow, they were with the underlying data. They impacted my insights.

I’m a lazy analyst: I don’t go back and review pipelines; I review dashboards, and those always concealed the errors that had crept in.

So the skill also checks the data and the flow together to capture possible issues.

The skill is useful but the bug report it wrote is worth more.

Try it on your own flows

I’ve packaged the whole method — parse the flow, translate it, verify it against real output until the diff is zero, then audit the result for exactly this kind of semantic drift — as a Claude Code plugin. You can install it using these lines in Claude: (it’s also on github)

claude plugin marketplace add acotgreave/tableau-prep-to-python
claude plugin install tableau-prep-to-python@howtospeakdata

Point it at a .tfl or .tflx and it’ll give you a Python script, a plain-English account of what your flow actually does (useful in itself if you inherited it), and a list of the quirks it faithfully reproduced that you may well want to remove.

There are caveats:

• I’ve tested this on three different Prep flows. Each succeeded. But that doesn’t mean it will work with yours.

• Real verification needs a reference output, so dig out the last output file your flow produced before you start.

• I’ve only tested on text file inputs and outputs. The skill will try to deal with “proper” databases, but I haven’t tried it on those.

There’s one final deliberate design choice: where it finds data issues, it does not silently correct them. It reproduces Prep exactly by default, tells you what’s wrong and how many rows are affected, and puts the fix behind a flag. You may have published figures based on the old numbers. That’s your call to make, not the tool’s.

Go and open one of your flows. I’d love to know if this works for you.

Also, are you migrating away from Tableau? Have you tried this same process for TWB files? I might try that project next: Desktop Dashboard to HTML app, anyone?

Final reminder: join me on “The AI Analyst?”, Thursday, August 6, 4pm BST.

Enrico Bertini and I will be discussing how we’ve used LLMs and GenAI to create data viz teaching assets, and lessons we’ve learnt.

I was joined by on The AI Analyst (LinkedIn, YouTube) to hear her thoughts on the latest in embedded analytics and why she’s building Ridge AI.

Before we get to Ellie and Ridge AI, here are my up and coming livestreams. Do join me!

• Chart Chat, Monday Aug 3, 11am EDT: we’ll look at how data teams have mapped air quality and wildfires around the world in this period of exceptional heat. Sign up here.

• “The AI Analyst?”, Thursday, August 6, 4pm BST: and I will be discussing how we’ve used LLMs and GenAI to create data viz teaching assets, and lessons we’ve learnt.

Here are three big takeaways from our chat. I’ll share my own thoughts at the end.

Embedding is still a holy grail

“If you give people data where they’re working, that’s helpful,” said Ellie. Embedded has been a goal as long as IFRAMEs have existed but it’s still not reached its potential.

Here’s how Ellie explains the benefit of embedded analytics:

Embedding technology has matured. It’s ready!

One of the problems of embedding has been the technology. Users of BI tools are simply not patient enough to wait 10, 15, 20 seconds for a display to load. Below, Ellie explains how the technology is finally ready to provide millisecond responses.

Subscribe now

Pricing AI analytics is evolving

Pricing a BI platform is harder than ever. Do you price by user? Compute? Something else? It’s a big topic in the industry.

Here, Ellie explains why they’ve chosen to price per “Ridge” (embedded asset) you create:

My reflections

At Tableau, it seemed that every year, in our company kick offs, we’d have leaders from the product and the sales teams say, “This is the year of embedded analytics.” And it never seemed to take off.

The key problem, as I saw it, was the dreaded spinning wheel you’d see while waiting for a view to render. If a user actively logged on to Tableau server, then it was just about accepted that you might wait for a view to load. Not ideal, but a price of using a BI platform.

But if you’re coming to my platform, and an embedded element takes an age to load, you’d judge me, not the underlying embedded tech.

For that reason, it’s the matured technology aspect that wins the day for me. Reducing friction has to be a prime objective for any vendor.

What do you think? Are you ready for embedded analytics yet? What are your experiences with embedded?

Take care and stay safe everyone.

Andy

Happy Friday everybody,

A quick note before the weekend to let you know what’s coming up.

tl/dr:

• Chart Chat 72 is on July 23 (register here)

• I am doing a currently Top Secret keynote at Big Data London in September

• A workshop and keynote at Data Decoded in Manchester in October

Chart Chat #72, Show and Tell: July 23

Next week on Chart Chat 72 (July 23rd) , , and I will share some data projects we’ve worked on recently.

I’ll be talking about the ups and downs of vibe-coding my Add To Playlist explorer (a work in progress you can see here).

Click here to register.

Subscribe now

Big Data London: 23 Sept, London

Tickets are available for the biggest data show in London: Big Data London.

I’m going to be involved in one of their main keynotes. Right now, I can’t share what we’re doing, but we’ve got a new concept to inform, educate and entertain you all, with a cast of hundreds [er, I think you mean 7, Andy - Ed.].

Data Decoded: 12-14 October, Manchester

I’m going to be in Manchester for the second edition of Data Decoded, 12-14 October. I’ll be doing a Data Communication workshop on the Monday, and a “Staying Human In The Age of AI” keynote during the main show.

Register for the workshop here.

That’s all for now. Stay safe and well and I’ll see you next time.

Andy

Subscribe now

In November 2022, OpenAI decided, as a research experiment, to stick a text box on top of its LLM, call it ChatGPT, and make it publicly available. It became the fastest-growing app ever. The rest of the tech world reacted and we all decided that, in order to interact with an LLM, you just needed a single-threaded text box chat interface.

In the data analytics world, everyone copied it. At Tableau, we raced to build a “Copilot”. PowerBI did the same. Everybody did the same.

Right from the start, this felt like the wrong way to do AI analytics. In January 2024, I gave a presentation to Tableau’s engineering team about my concerns.

Now, in July 2026, the same UX problems persist: data gets simultanously oversimplified and the interface gets harder to use.

Let’s take a deeper look. In my next newsletter, I’ll share solutions.

The problems

1. Conversations with data are not linear

Verbal conversations are linear. Conversations with data are not.

It’s 11 years since I wrote about the Squiggle of Visual Analysis. tl/dr: data exploration is the ability to jump back and forth, up and down, in and out as you seek the best articulation of the data.

Have you ever explored data in ChatGPT, Claude, or Gemini, and got frustrated because you’ve had to scroll back up the page to find an earlier interaction? That’s the manifestation of this problem.

Single-threaded data analytics in a text box is not the squiggle. Scrolling up and down in a chat interface and trying to refer back to an earlier moment is annoying and inefficient.

2. Single-threads give you one answer at a time

A single thread in conversations with LLMs are fine if you’re just shooting the breeze. Surely this is limiting? An LLM could provide infinite answers to any questions you ask. How many useful insights do you miss because the LLM only gives you one?

When I’m in a room with lots of humans, each of them bring their own characteristics and biases to a data related question; many might be surprising and useful. I want that in my analytics too: it’s the squiggle in action, in AI analytics.

3. Don’t make me move my eyes!

One of the best things about working about Tableau was that it was built to exploit the power of our cognitive system. The goal was to keep users’ eyes where the action was. This avoids “change blindness.” What is that? Watch the video below: can you spot the difference in the two images?

It’s hard to see because of the “flash” of grey between the two images. That’s all it takes to flush your brain’s memory cache of what it just saw. Amazing, right?

The same happens if you move you eye from one part of a screen to another. This is why we always tried to keep your eye on the action in Tableau.

The problem with side-bar AI conversations is that your eye is always moving and missing the action. For example, imagine you want to change something using the Copilot in PowerBI. The red line in the image below traces what your eyes might do:

You’re looking at data... then you type a question (no longer looking at data).... then you wait a while (do you go look at your phone for a moment?)… and finally you go back to the data display. What changed? Who knows, because you’re weren’t looking there.

4. Data gets squished in conversation panes

While the first three problems are about how the analysis gets done, this last one is about how it gets presented.

A lot of apps use conversational panes at the side of a main data canvas. This seems logical: “Let’s build the app so the datas in the main view and the conversation to the side,” seems like a reasonable product direction.

The glitzy demos say “As you converse with the data, you get charts inline showing you the details.” Two things tend to happen in these demos. Either

• the resulting chart is unrealistically simple

• the resulting chart is unreadable and the demo moves on very quickly

Real world charts get complex quickly and there simply isn’t enough screen real estate in these panes to show the detail. If we actually expect users to consume the information in those panes, we need to fix the display.

What do we do about this?

What do you think? Are there other UX problems with the current approach to AI data analytics?

I proposed ideas to the development team about ways to minimize the problems. Two years later, I’m seeing a lot of vendors implement various novel ideas, too.

I’ll explore all of those in part 2.

Chart Chat: show and tell my latest project

It’s Chart Chat 72 on July 23rd. We’re all going to share some recent projects, along with the highs and lows of each. I’ll be sharing my Add To Playlist explorer (a work in progress you can see here).

Subscribe now

Hello all,

There’s a token paradox that baffles me.

Organisations are seeing an explosion in costs.

• Uber burnt its entire annual AI budget in just four months.

• At Nvidia, the cost of compute is more than the cost of employees

• Software startups are spending upwards of $10million on inference and LLM products

At the same time token costs are plummeting.

As someone who doesn’t have a background in economics, this seems like a paradox to me. If the underlying cost is going down, how come it’s blowing budgets?

I like to see this kind of thing visually, so I turned to Claude Code to visualise the paradox. It created this lovely app to let me tweak the results.

First, there is “Today’s reality” (according to Claude: 50% decline in token cost a year, 200% growth in usage):

That results in 7.6x higher costs in 5 years.

What if there is an “Agent Explosion”? (Usage up 600% a year, costs decline 30% a year). Then it gets really scary:

Subscribe now

Notes

• This is a vibe-coded app built to help me see the paradox in my mind. There are a thousand caveats about this kind of model, but it helped me. I’m writing this post not to reveal some new insight in the world of economics but to show how genAI tools can assist with thinking processes by the rapid creation of learning tools. You can go play with the app yourself to see difference scenarios.

• If you’re reading widely on AI or are already an Economics expert, the paradox is not news to you.

• In researching this article, I also discovered Jarvon’s Paradox, a well known paradox similar to what’s happening with token cost and usage.

Have you used vibe-coding to help understand any concepts you’ve struggled with? Do let me know in the comments.

Before you leave:

• Do sign up for “Why we’re building Ridge AI” on my next episode of “The AI Analyst?” with Ellie Fields.

• Further ahead, it’s Chart Chat 72 on July 23rd. We’re all going to share some recent projects, along with the highs and lows of each. I’ll be sharing my Add To Playlist explorer (a work in progress you can see here)

I’ve written many posts over the years recommending data-adjacent books. I enjoy hearing from people who go on to read them and the recommendations you send me; I’ve discovered many a great book thanks to you.

My summer 26 list (below) might seem to push the boundaries of being “data adjacent.” They all belong to a theme that everyone connected to the technology industry must consider: power and inequality.

Magnifica Humanitas, by Pope Leo XIV

You’ve probably read many other people’s takes on the Pope’s encyclical on AI. You might have even popped the URL into a chatbot and asked for a summary.

I decided to sit on my back deck and read it. I recommend you do too. An AI summary will not get close to capturing the clarity and depth of this important AI assessment. [FYI, at 47,000 words, it’s half the length of To Kill A Mockingbird].

“When such power is concentrated in the hands of a few, it tends to become opaque and evade public oversight, increasing the risk of distorted forms of development that give rise to new dependencies, exclusions, manipulations and inequalities.” (paragraph 95, Magnifica Humanitas)

The encyclical is, among many things, a statement against the centralisation of power.

• It’s in favour of a more communal sharing world.

• It’s against a restriction of people’s rights, opportunities, status, and freedom.

• It’s in favour of regulation that works for all, not the few.

I am all in favour of the Pope’s viewpoints. I continue to use AI and work with it, but as I have written before, I despise a great deal of Big Tech’s theft from us: of our knowledge, of our natural resources, and of our attention. The encyclical is a worthwhile counter to the flood of pro-AI slop.

[Note: I see irony in this recommendation. The Pope is right to call out the problems of centralised power, but the Catholic Church is hardly innocent of these problems. Take your pick from the exclusion of women, cover-ups of child abuse, banking scandals, etc.]

Nobody’s Girl, Virginia Roberts Giuffre

Why is this book in a list of books for data people? Everyone should read it. All should bear witness to the abuse perpetuated by Epstein, his associates, and his enablers. For that reason alone, I include it.

Words I write here can’t convey the horror of reading the abuse Giuffre suffered. It’s a tough book to read. But reading it is nothing compared to what Giuffre went through during the periods of abuse, and after, as she fought for justice.

Sticking to the theme of this newsletter, her book also overlaps with themes that the Pope writes about: power corrupts.

If we’re to work in the data industry and make use of AI tools, we are part of a machine that supports the collection of great wealth and power.

Reading this book forces me to confront the real-world consequences of unchecked power. I highly recommend this.

Subscribe now

All That We See Or Seem, Ken Liu

Want some fiction to lighten the load? Well, this is a near-future dystopian sci-fi novel, so you’re still going to be challenged. In this book, Julia, a young hacker, follows the trail of Elli. Elli is an “oenirofex”, a dream artist, who weaves an audience’s dreams together through a shared landscape.

Do I recommend this for the plotting and the language? No. The twists and turns are clunkier than a Dan Brown novel. The dialogue is bad. And there are caldera-sized plot holes.

But that’s no problem. Partly because Dan Brown novels are fun, so why can’t this one be a romp, too?

The sci-fi in the book is excellent. We’re taken to some time in the future. It feels like it’s set only 4 or 5 years ahead of today. AI has taken over, people have withdrawn into talking to their own agents, and surveillance is inescapable.

Before becoming a full-time writer, Liu was a software engineer and a corporate lawyer. As my personal interests increasingly focus on the impact of tech on society, his meta-commentary on the topic resonates with me. He also explains in this interview that he tries to remain neutral as an author and let the reader draw their own conclusions.

What are you reading?

That’s my list - if you’ve read any of those, let me know your thoughts. If not, what have you been reading this year that’s been useful for all of us in the data world?

Before you leave:

• Do sign up for “Why we’re building Ridge AI” on my next episode of “The AI Analyst?” with Ellie Fields.

• Further ahead, it’s Chart Chat 72 on July 23rd. We’re all going to share some recent projects, along with the highs and lows of each. I’ll be sharing my Add To Playlist explorer (a work in progress you can see here)

I’m excited about 2 livestreams this week coming up:

Golden Analytics: Ask Me Anything with Francois Ajenstat, Mon 1 June

Join me live with Francois and find out everything you need to know about the new kid on the analytics-platform-block. Got questions? Submit them here.

Supercharge v Sabotage: The Two Faces of AI.

Chart Chat on Thursday June 18 will see us discuss our latest feelings about AI in analytics. Should be spicy! Register here.

I hope to see you online at one or both of these.

Best wishes
Andy

Subscribe now

It’s been a busy couple of months and will continue to be so. I’ve got some exciting livestreams coming up (and ones you can catch up on).

Also: if you’re looking for a keynote speaker, a trainer, or strategic advice, book some time on my calendar.

What to watch

Golden Analytics: Ask Me Anything (LinkedIn)

On Monday 1st June, I’ll be joined live by Francois Ajenstat, founder and CEO of Golden Analytics.

The format is Ask Francois Anything, so use this slido to submit questions. Francois has also told me there will be some exclusive announcements, so join us for the big reveals he has in store.

• Here’s the event page on LinkedIn.

• If you have a question, or want to vote up existing questions, use this slido.

Chart Chat: Supercharge v Sabotage

Our next episode of Chart Chat is on Thursday 18 June: “Supercharge v Sabotage: the two faces of AI data analysis”. Following last month’s hallucination errors, we’re going to share our feelings about this AI stuff. Is it good for our field? Register here.

Further ahead…

• In late June, I’ll be joining Veezoo to announce the winners of the Veezoo Analytics Cup. There is still time to enter their contest for a chance to win a MacBook or iPad. Entries close on May 25th.

• Early July: Ellie Fields, co-founder of Ridge, will be on The AI Analyst to discuss how Ridge is tackling the AI-driven analytics landscape. We haven’t fixed the date yet, so make sure you’re subscribed to be notified.

  Subscribe now

Catch up

• “The Future of Data is Fake” - I chatted with Mark Brocato, Head of Engineering at Tonic, about all things synthetic data. I’ve had great fun using Fabricate for making fake data. Until using it, I hadn’t realised the size of the synthetic data market. Watch the show here.
  For a free trial of Tonic Fabricate, click here.

• “Staying human in the age of AI” - it was an honour to open the Data Innovation Summit in Stockholm, where I shared my thoughts on the impacts of AI. Spoiler: There are serious risks we need to consider.

• “What will define the next decade of business-first innovation?” I sat on a panel discussion about AI’s impact on analytics. The full watch is here, but if you want to see me get spicy over how CDOs overmeasure success, watch this section. 🙂

What to read

As a subscriber, you’ll have seen my previous newsletters on AI/Humanity and AI Hallucinations. There was even more to write about, which I covered in the last Chart Chat newsletter, too:

Chart Chat

Chart makeovers... with an AI sting

Hi, Andy here with you this time. I hope you’re all well…

Read more

4 months ago · 12 likes · Andy Cotgreave

Thank you for being a reader of the newsletter. I appreciate all the comments and ideas.

If you’d like to work with me, let me know. I’m available for keynotes, strategic advice, and data communications training. Do book some time on my calendar here.

Until next time, take care,

Andy

This post is a summary of my keynote talk at the Data Innovation Summit in Stockholm on 6 May. Welcome aboard if this is your first time here. You can find me on LinkedIn, and do please subscribe to this newsletter, too:

Subscribe now

This period of AI disruption feels like the Rock’Em Sock’Em robots bashing around in my head. On one hand, the blue robot is fighting for me to realise that AI is an amazing tech that is going to disrupt in a good way. On the other hand, I have the red robot fighting to convince me that the existential risks are too huge to ignore.

This internal tug-of-war shows up in my daily work as a data professional: should I lean into experimenting with the latest AI tools and features, or slow down to scrutinize their impact and safeguard against mistakes?

These constant decisions keep the debate between excitement and caution alive in nearly every project I take on.

Excitement abounds. We’re all vibe-coding and doing things we never really dreamed possible. The example I shared is my homage to my favourite radio show, Add To Playlist (http://addtoplaylist.howtospeakdata.com/).

With Claude Code, I wrote an app that scrapes all tracks ever featured and summarises them with charts and a searchable catalogue.

And yet there is a never-ending story of things going wrong. The one I featured is about the Clawbot agent whose job was to find and fix errors on GitHub. What did it do when a human being rejected one of its changes?

It researched, wrote and published a hit piece about Scott Shambaugh, the human moderator of matplotlib. That post is readable by all, and will now itself be subsumed into the knowledge of LLMs.

More reading about the Clawbot agent:

• This was featured in the tech press (e.g., The Register) and on podcasts (e.g., Hard Fork)

• Here’s where Scott denied the push.

• Here is the Clawbot agent’s response.

If that’s the quandary, the good and the bad, how can we stay human?

I shared ideas in three critical areas:

Data Storytelling and Analysis

I talked about the risk of letting AI Assistants frame your story on your behalf. You must stay in control of the framing.

Here are links to the things I talked about:

• Claude has great add-ins for PowerPoint and Excel.

• Thinking Fast and Slow by Daniel Kahneman describes System 1 and System 2 thinking. I recommend prompting in system 2 mode: it’s a way to provide as much context as possible to your AI Assistants right at the start.

Semantics and Context

Semantics is more important than ever. I referred to my episode of The AI Analysts with Charles Schaefer from Hex when we talked about Hex’s Context Studio. It’s a part of the platform that monitors users’ questions and flags ones that Hex cannot answer based on semantic definitions. Those flagged conversations are sent to the data team so they can extend the semantic layer.

“Like it or not, data people are being turned into ‘context curators.’ That’s your role now,” was Charles’ excellent quote from our episode of The AI Analyst (see all episodes here).

Governance: local and global

At an organisational level, I shared two great resources:

• “That meeting you hate may keep AI from stealing your job” is an excellent article by Noam Scheiber at the New York Times.

• Leader, Lab, Crowd is an organisational framework described by Ethan Mollick.

And at the global level, I shared these resources:

• Do read The Empire of AI by Karen Hao. It changed my worldview.

• The Economist did a great piece asking if we should trust just 5 men with AI.

My final call is for us all to consider how we regulate and educate ourselves to mitigate the risks of the dark side of AI. It’s not easy.

Summary

I’d love to know your thoughts.

Do you have the two voices on your shoulders, telling you that AI is simultaneously exciting and terrifying? How do you manage it? How do you see us staying human in this age of AI?

That’s all for this one. Do follow me on LinkedIn and subscribe to this newsletter if you haven’t already done so. I’ll be continuing my adventure trying to make sense of AI in our data industry.

Subscribe now

It’s too easy to be wowed by AI Analytics and miss mistakes. Mistakes that then propagate through to incorrect decision-making.

Our latest Chart Chat episode highlighted how even seasoned professionals get carried away. My trust in AI Analytics had been rebuilding after a previous heinous error (see below), but I admit, this has really set me back. I’m back to Very Wary mode.

Can you spot the error?

It was a “One Chart, Four Different Ways” episode. Steve, Jeff, Amanda and I were all remaking a chart about active workers and retirees from Voronoi. Three of us used AI Assistants to remake our chart.

Steve used Claude Code to build an attractive dashboard with BANs. Through several versions, ending up with the image below as his final makeover.

It’s lovely. In our episode, we discussed how impressed we were with Claude.

After every episode of (do go subscribe to our newsletter!) we have an “After Party” - an open invite Zoom where attendees can share their conversation. After this episode, BJ Price explained how she is still concerned about hallucinations. Someone then asked, “Steve, are you sure your version didn’t contain errors?”

A quick check, and Steve realised that the Peak Worker Surplus calculation was wrong:

Claude had indeed made an error. The surplus in 2008 was 2.4 million.

In each iteration Steve made, the error was passed on. Wowed by the attractive styling, neither he nor us spotted it.

But there’s an even bigger error we didn’t spot

In writing this post, I now see an even bigger error. It’s even more obvious than the subtraction error and none of us spotted it in the building of the viz, during Chart Chat, or during the After Party.

Look again at the chart. When is the “peak worker surplus”? (i.e., when is the gap between the blue and red lines the highest?)

Claude’s BAN claimed the peak was 2008. It’s clearly not: it’s 2001. (5.1 million)

I’m really shocked. Shocked at our inability to spot this.

Imagine doing this for real data. You do the work and share your insights. Decisions get made based on 2008 not 2001. Errors like that could easily be very expensive.

What happened under the hood?

The BANs were hard-coded in the HTML. It’s not clear where Claude made the mistake, but once that hallucination was on the dashboard, the mistake was permanent.

Subscribe now

I thought we’d got past this

I’m building an “AI for Data Analysts” course (details soon) and had been using an old example of hallucinations that happened in ChatGPT 4.

In this case, when I’d asked ChatGPT for a chart based on a really simple dataset, it created a line chart with the months sorted alphabetically. To make it worse, ChatGPT labelled the x-axis with the numbers 1-12, further hiding the error. I only caught it because it was a fake dataset and I knew it should show sales going up.

My course contact said we can drop the section on hallucination because the ChatGPT 4 option is old, and the models have gotten so much better since then.

I agreed with him. We dropped it.

Now I have a new example proving we cannot drop the hallucination risk section.

What do we do?

I’m somewhat shell-shocked this morning. I have been increasingly gung-ho about AI-driven analytics. I am shaken by seeing a glaring error that none of us spotted. And in seeing Steve talk through the iterations he took, he’s representing the work any one of us could do: iterate through a solution, without deeply checking every result.

What’s the solution? Manually test every number? Ditch the AI Analytics?

I’m honestly not sure. We’re all trying to work harder and faster. How do we make time to check everything?

What do you think? Are you checking all your numbers? Are you seeing hallucinations?

Before you go, here’s a quick plug for Dashboards That Deliver. Right now, it’s at a steeply discounted price (60% off) on Amazon UK and the US. If you’ve been on the fence about buying it, now’s your chance!

I hope you’re well.

This is a short newsletter with links to things I’m up to in the next month or so.

On Thursday, at 4 pm BST (11 am ET), it’s Chart Chat #70. We’ll each be sharing our own makeovers of a chart from Voronoi. These episodes are always great ways to explore the subtleties of data storytelling and how new tech is making things easier.

You can register here or join us live on LinkedIn and YouTube.

I’m excited to be delivering the opening keynote at the Data Innovation Summit in Stockholm on 6 May. The topic is about staying human in the age of AI. Why Rock ‘Em Sock ‘Em robots? You’ll have to wait and see.

You can attend live or virtually. Details here.

I will also be continuing with The AI Analyst series. I have 3 episodes in the pipeline. One’s with : we’ll be talking about using AI to help teach data visualisation principles. I can’t yet tell you about the others, but I’m excited to bring the guests to the show. I’ll be announcing those on Substack and on LinkedIn.

Subscribe now

If you attend one of these events, do come say hi.

Take care,

Andy

“Millions of dollars are spent yearly collecting data with the expectation that the data will automatically cause the correction of the conditions studied.”

That’s true, right?

Pop quiz: what year is this quote from? Clue: it’s not from the 21st century. Read on…

• Chart Chat is live on 30 April: we’ll be doing 4 makeovers of a chart about retirement numbers. 👉Register here and join us live on LinkedIn and YouTube.

Would you fall out of your seat if you knew it was 1914?

Yes, 1914. The quote is from the first page of Graphic Methods for Presenting Facts by Willard Cope Brinton. Over 100 years ago, there was deep frustration that money invested in data was not reaping benefits. Brinton was a frustrated engineer who wanted the world to store, analyse and present data effectively.

His book could still be a manual for any contemporary leader.

Here are four ways Brinton’s book could inform a contemporary data strategy:

The Data Layer

1914 was the start of the age of big data, not the early 2010s as the big vendors would have had us believe.

Back then, it was punch cards and tabulating machines.

The image above is a census card from the 1910 US election. One card per citizen. Can you imagine being the person signing off on that data model? Can you imagine getting the cards back and realising you’d missed out a category? These days, we can build the data model as we go, fortunately.

It makes me smile to think that a modern analyst could type ‘SELECT * FROM Population where [State]=“Florida”’ and instantly retrieve records. In 1914, someone would have to physically go to the physical data warehouse and physically select all the records by physically checking their state value.

Once you have your punch cards, a modern tabulating machine could process three thousand records per hour!

The Semantic Layer

In 2026, semantics is more important than ever. If we’re going to get AI to work, it needs context. I’ve talked about this with Charles from Hex and Christine Haskell on my show The AI Analyst.

Defining data was just as important in 1914, and index cards were Brinton’s method.

A single index card could contain one year of data for one measure. The data could be drawn as a line chart and shown as a table. In the image above, you can see the data has been signed and approved (governance was important too!).

But there’s a large blank space. Brinton said this could be used to write in extra context and, if necessary, written on a separate index card, linked by card id. That’s essentially what we do today: define data and extra contextual information in a semantic layer.

Subscribe now

The Analytics Layer

Remember when Tableau disrupted the market? Its underlying VizQL technology allowed us to drag and drop fields around so we could see data any which way. Brinton foresaw that with his index cards.

If someone wanted to see, for example, 3 years of sales data, the analyst could retrieve the index cards, lay them on a table and move them around. Horizontal or vertical arrangements would let you see the data in different ways. Brilliant. Brinton knew that data needed to be flexible to be analysed.

Oh, and what kind of person could do this work? “Any man of ordinary mental calibre could do this work.” Good to know!

The Action Layer

We invest in data to drive action. Brinton knew that visualising and presenting data is a vital part of the data strategy. “There is danger in giving too much information to executives of small brain capacity,” he said. Ouch. I like to think he meant “very busy leaders” 😊

Brinton had opinions! He advocated for line and bar charts and hated pie charts. “Banish them to the scrapheap,” he said.

He shared a photo of a New York municipal parade where one of the trucks had city performance charts on it. Today, it is a mere button press to share data with the entire planet. In 1914, you’d need a cart, a horse and a driver!

So nothing’s new?

Obviously, AI feels like a massive disruptor, but if you’re in charge of a data strategy, the four layers remain the same. You don’t need to read Brinton’s book (but I recommend you do), but as you develop your data strategy in conjunction with vendors and stakeholders, these are the four considerations that must be kept in mind, whether it’s 1914 or 2026:

• Where do you store your data?

• How do you define it?

• How do you analyse it?

• And how do you drive change with it?

I would love to know your thoughts.

In the meantime, do subscribe if you like this content. Also, I’m really active on LinkedIn. And do join us on Chart Chat on 30 April for some data storytelling makeovers.

.

joined me on The AI Analyst to talk about GenAI’s impact on data storytelling. Her core philosophy is that data professionals must stop positioning themselves as “chart makers” and evolve into “decision shapers.” She spoke on that topic at Outlier last year and has written on Substack about it.

AI can give us instant output, but it cannot think and reason.

I want to highlight three key insights from our conversation:

The Dangers of Frictionless AI

Creating things (charts, art, newsletters, etc) involves friction. That friction has value.

AI outputs, fast and polished, look authoritative. This creates an “illusion of precision” and an automation bias. It can cause users to blindly trust the machine’s verdict without double-checking the facts.

Gabby advocates for “mindful friction” (a principle she attributes to Projects By If). These are deliberate stops in the workflow that force users to question the context and ensure accuracy before decisions are made. Traditionally, they happen naturally: if we’re sketching ideas, the very act of drawing them, on paper or on screen, is slow. It allows for pauses to deliberate and inspire.

AI removes those pauses. I’m finding myself falling foul of it already.

I’m working on some vibe-coded data-driven apps right now, and it’s hard to stop and reflect while I’m in creation mode. The dopamine-release I feel when typing a prompt and coming back to see the result is addictive. I fear, though, that I’m not stopping to think enough and instead am chasing the next iteration.

Gabby advocates for us to build the friction back into the process.

Subscribe now

Data Professionals Must Become Guardians

If a data storyteller sees themselves as mere chart creators, their role is at risk: AI does that task faster. Gabby calls on us all to be strategic educators and guardians of the process. We must be gatekeepers of the knowledge.

I see this in conversations I’m having with new AI-first companies I’m talking to. They recognise that AIs are no better than well-intentioned interns: enthusiastic, with basic knowledge, likely to make big mistakes. You’d never put them in charge of your main client’s portfolio. Our roles are needed to help this new generation of software and the people who will use it.

Use AI As a “Sceptic”

I admit to liking this one and have used it myself. At Figures and Figures, Gabby’s design studio, they set their AI Assistants to critique their work. It’s another perspective for generating creative ideas.

Don’t cut humans out of getting this feedback, but the AI can give you another useful perspective.

I really enjoyed this conversation with Gabby. What do you think? Is genAI changing how you do data storytelling? Let me know.

Thanks for reading! And now, some updates on what I’m up to in the next couple of months.

• I’m delivering my “How To Speak Data” course at Data Decoded in London on Tuesday, 21 April. Come along to learn all about the LEVERS of chart making. Register here.

• I’m delivering the opening keynote at Data Innovation Summit in Stockholm on 6 May: Why Human Insight Is More Important Than Ever in the Age of AI. I can’t wait. It’s going to be spicy. 🌶️

• The next episode of The AI Analyst? will be on May 20 with . We’ll be talking about using GenAI to make dataviz teaching tools. Details here on LinkedIn.

That’s all for now. Take care,

Andy

Claude’s new add-in for PowerPoint and Excel is powerful. It makes Microsoft’s own Copilot look embarrassingly weak. For data storytellers, it’s good enough to build data presentations. As always, the upside is matched by a downside: if we’re asking LLMs to create editorialised data stories, who is deciding which way to spin the stories?

Let me share two ways I used it this week.

First, do join me for episode 3 of “The AI Analyst?”. I’ll be discussing the impact of AI on data storytelling with . We are live on Wednesday March 25th at 2pm GMT.

Building data stories with the add-in

Currently, I’m writing a course on analytics for Datacamp. Part of the course covers how formatting can frame data stories. I wanted to create a set of slides based on ’s excellent Tableau Public dashboard about How Americans Spend Their Time (itself based on a chart from Our World In Data).

Even though it’s a relatively straightforward data story, it’s a skilled job to execute in Tableau or PowerPoint.

I had the data open in Excel. I pasted a screenshot of the dashboard into the Claude’s PowerPoint chat box, along with a detailed prompt explaining that I wanted one slide for each of the charts:

And… it worked. First time. I had 3 slides, each framing the story in a different way, as directed:

Now, you might think, “Well, that’s taking an idea that’s already been created and recreating it.” Fair enough.

I tried another experiment. I gave the PowerPoint add-in this fictional data about burglaries:

• Solved: 11,977

• Unsolved: 5,988

And asked Claude to create two slides based on the data, one telling a positive story and the other a negative one. Here’s what it built (first time, no editing)

The risk introduced by this power

What are the implications of all this? If you don’t ask for editorial guidelines, is there a risk Claude will make editorial decisions for you?

I did the experiment again, providing Claude with the data and giving it a terrible prompt: “Give me a data story based on this data.” I was gratified to see this response before it completed the task:

I am glad this safety is built into Claude. We must remain in control of editorial decisions. Otherwise implicit bias inside the LLM’s models will emerge.

I redid the “bad prompt” exercise with Gemini (see it here), and it happily took a positive spin on the data, without any questioning whether 33% failure rate might actually be a bad thing: “Based on the data provided, the clear ‘story’ here is a remarkably high success rate for law enforcement,” it said. No room for nuance or other spins, then?

Gemini also has a long way to go to catch up with Claude’s design chops:

Use the power responsibly

I’m very impressed with Claude’s add-in. As always, with LLM advancements, there are risks as well as rewards. It’s a tool to help you be productive, but it is not a substitute for critical thinking and care in your data stories.

P.S Poor Copilot…

I tried the exercise using Microsoft’s Copilot.

It’s awful that you’re in PowerPoint, and a request to build a slide won’t build a slide in the deck you have open. You have to download a deck.

The deck it did build was appalling. It’s not even 16:9. It looks like it’s from 2003:

What’s your experience? Have you used the add-in yet? Let me know what you’ve done.

That’s all for now.

• Do join me and on on Wednesday March 25th at 2pm GMT for “The AI Analyst?”

• And, do subscribe to this newsletter if you haven’t already.

Subscribe now

Claude continues to turn my assumptions about AI’s role in data analysis upside down. Mockups and prototypes? Claude’s changed the game with a feature that, in my experiments, is fast, reliable and productive.

Prototypes and mockups with Claude

Claude now creates custom visuals and diagrams inline in its chats. It’s not image generation, and that’s really significant. Instead of static images, we now have SVGs: editable, copyable, scalable. There are so many ways this can help data analysts. The first thing I tried was creating mockups.

In Dashboards That Deliver, a core part of our development framework is the Discovery and Prototype phase (you can find resources on that on our site). During this phase, you should be working with users and presenting them with many ideas. Low- to medium-fidelity assets are really important at this point.

I gave Claude my spreadsheet and this prompt:

And Claude returned these three ideas:

I also requested mobile layout ideas. Claude was suggesting a hierarchy of views, too, with navigation paths:

Depending on where you are in the development phase, you can make the diagrams higher- or lower-fidelity.

How would you do mockups in the past? Figma? Pen and Paper? PowerPoint? Tableau?

All of those are still valid approaches, but this is a fast, new way to test out ideas. Previously, I’d have to place each section manually. Now I can describe a vision and iterate.

I’ve begun reworking all my personal Tableau dashboards as web apps, using Claude Code, and these diagrams are already helping me move faster.

Don’t use this as a shortcut to not speak to users, of course: they need to see, understand, and provide feedback on what works and what doesn’t for them.

What do you think? Have you used these features? Are they game-changing for you?

Do join me and on on Wednesday March 25th at 2 pm GMT for “The AI Analyst?” We’re going to discuss how these new features impact our work as data storytellers.

Finally, do subscribe to this newsletter if you haven’t already:

Subscribe now

That’s all for now. Take care, Andy

Dear ,

[CEO of Superhuman, the makers of Grammarly, currently presiding over a name-stealing shitstorm of your own making]

Your non-apology for using people’s names in your Expert Review feature doesn’t even pass the AI test in your own product. 34% of your text has patterns that resemble AI text.

When you are under fire for using AI to mimic human voices, releasing an apology that sounds like AI wrote it reinforces the exact criticism you are trying to deflect. Now I believe even more that you’re prioritising synthetic efficiency over human authenticity.

Yours,

The Humans.

What’s going on?

We all want experts reviewing our work, right? I do. When I write about data, or AI, or presenting, I love getting feedback from pros. Guess what, Grammarly made it available to me for free!

Above, you can see the feature in action. All I needed to do was click on “Expert Review”, and named experts in the topic your text is about would provide ideas.

I received expert advice from Cole Knaflic on my draft data storytelling course. Thanks, Cole (and , Stephen Few, ); I guess I don’t need to buy your books anymore.

Grammarly’s Expert Review feature says AI’s quiet bit out loud: LLMs take content from people without permission or payment. They use that knowledge to line their own pockets. By putting the names to the content, Grammarly made it more explicit.

The fallout has been swift. Unsurprisingly, journalists were angry (Casey Newton, Stevie Bonifield, and Julia Angwin, who has initiated a class action lawsuit against Superhuman).

Grammarly took the feature down this week. Fine, but who in that organisation thought it would be a good idea in the first place? Superhuman’s CEO, Shishir Mehrota, wrote an appalling AI-slop apology that acknowledges nothing. It reads like something from Lumon Industries.

The apology is badly written. I put it through Grammarly’s own app, and it identified it as AI-written. It’s a terrible advert for a writing company. In the face of human backlash, the CEO can’t even be bothered to create something that has a sense of humanity.

This debacle really got under my skin. Any LLM can do an expert review in the style of Cole and Alberto if I ask it to. Their knowledge has already been sucked into their knowledge bases. So why is this one different?

I think it’s because what they did was just so blatant. “We will use your names, without consent and without payment, to drive profits to our own organisations. Simultaneously, we will throttle your income streams. Thank you. Have fun. Have a free trial.”

Perhaps I got cross because this is also the week I chose to delete ChatGPT (yeah, I know the QuitGPT movement says to cancel subscription, but I went a step further).

I’m sick of unregulated vulture capitalism (do read ’s excellent book). I’m sick of tech bros and billionaires working with corrupt regimes at the expense of others (also read Empire of AI by ).

I’m not against AI. The technology is astonishing, exciting, and disruptive. It’s also not going away. I will continue writing about the discoveries I make with it.

However, it’s also dangerous and unregulated. Twenty years ago, we didn’t do anything to regulate social media, and that has got us into a mess now. I don’t want us to do the same with AI.

What can one Substack account do against the massed private interests that control the economy? Probably not a lot, but I’m bored of being quiet. One step at a time.

What’s my action? Well, I’m going to write more about this aspect of AI. I’ve quit ChatGPT, X and Amazon. I’m not 100% sure what my next steps should be. Any ideas?

Subscribe now

“Like it or not, data people are being turned into ‘context curators’. That’s your job now,” said Charles Schaefer from Hex on my latest episode of The AI Analyst on LinkedIn

(sidenote: should I switch to doing videos here on Substack?)

If we’re all becoming “context curators”, as Charles claimed in my latest episode of The AI Analyst, what’s the impact? Data analysts got used to being the visible “heroes” of those great dashboards or data stories. Becoming a context curator pushes the analyst behind the scenes, enabling conversational AI to take the stage. How does that make us feel?

Subscribe now

My conversation with Charles focused on semantic layers.

“If you believe semantic layers are the secret to unlocking AI on data, I think that’s just a myopic view,” he said.

He acknowledges that semantic layers allow naive LLMs to understand your business. However, businesses move faster than semantic layers get defined. Rigid semantic layers prevent true self-service in organisations by blocking ad-hoc, novel questions from being answered correctly.

The Context Puzzle

This “context puzzle” is the bigger challenge data teams must consider: how can users get answers to questions the semantic layers didn’t anticipate?

Hex solves this with Context Studio: when users ask novel questions, the platform searches for relevant information and if a certified definition is missing, it defines an answer based on what it can find. These “missing context” queries are flagged so that data teams can spot them. They can work with the people who asked these questions and update the system’s knowledge to create new, trusted metrics.

This is why data analysts are becoming “context curators". They work less on specific visual assets (dashboards, reports, presentations) and more on enabling successful conversations.

My reflections

I think Charles is right: the role of the analyst is changing. There’s enriching and satisfying work to be done for the analyst of the future: enhancing data models to enable self-service is a worthy role.

Is that a change data analysts will embrace? The data analyst role, for the last 10 years, has been about being the hero who builds visible, useful assets (dashboards, presentations, stories, etc). “Context curation” puts the analyst behind the scenes. That’s going to change who comes into this field, for good or bad.

What do you think? Is the future no longer better visualisations, but better definitions? I’d love to know your thoughts.

You can get an extended free trial of Hex here.

Hex supported the episode of The AI Analyst as a brand partnership. Rest assured, I take my responsibility to bring you useful, authentic content seriously. I pledge to work only with organisations and people I trust and who I think bring value to our field. If you ever think I am breaking that promise, please let me know.

In other news:

• Chart Chat #69 this week is all about our favourite Olympic Charts. Do join us for a data graphic geek out.

• I am teaching my How To Speak Data course at Data Decoded in London in April. I’m also delivering a keynote at the main event. Come say hi.

• I’m also keynoting in Stockholm in May at the Data Innovation Summit on the topic of Humanity and AI.

• Do you wish you had a one-page design cheat sheet for making great dashboards? Of course you do! Guess what? We have one in our members’ area on DashboardsThatDeliver.com.