Can AI tell a better data story?

Hi, Cole here. Recently, I put the same data and series of prompts into five different AIs. Today, I want to share with you what I found, highlighting where things went well, some issues that came up along the way, and sharing my thoughts on how you can use AI for data storytelling and communication. Before we turn to AI, let me tell you a

little bit about what I shared with it. It was a scenario from my book Storytelling with Data, Let's Practice. Specifically, we're looking at data that shows up a couple of times in exercises 2. 7 and 4. 4. The example is a real one. It's from the banking sector, and the original human designer made a graph that looked like this. Notice here we're plotting bank index. It's unclear from this exactly what that is. It turns out it's actually a satisfaction score, and that's plotted against the vertical Y axis. We have time on the X axis with data points called out in a legend at the top for financial savings and the peer group average. In the graph, we see those plus many more dots, which are the other banks flowing into that peer group average. I'm explaining this just to give you a little context, which will be helpful as we explore the iterations from AI. I did not provide this to AI as a starting point. Rather, I gave it this. This is the underlying data from the graph we just saw. I figured this way, it can start with a blank slate. I'm not biasing it with any original view of the data. Rather, I want to see how it interprets and visualizes things. At the onset, I didn't provide any context, simply this data table to see what it would come up with.

My first prompt was simply to use this data to create a data story. Interestingly, not one of them asked me for any information. There were no questions about what the bank index was or what it meant, about who my audience would be or what they cared about or about my goals. I should mention I have used chat GPT for the better part of a year, maybe even more at this point. So, it knows who I am, uh the books I've written, the content of those. I asked it to suspend that knowledge for this conversation. I don't know if it did or not. It's possible that history influenced things. Though, maybe not because if we look first to how chat GPT responded to my prompt, there was not a single graph in its output. After listing a few observations about the data, this was the summary it provided. Uh the banking sector has experienced steady growth, commentary on improving index values over time. That's interesting because as mentioned, I didn't tell it what the index is. It assumed it's a positive thing, which may or may not be the case. It turns out it is here. And we see that reinforced through the rest of the summary. Gemini made the same assumption, calling the story one of resilient recovery. Gemini also included a lot of descriptive text, which I did not show here. This was common across all of the AI I tested except Claude. Claude was the only one that included a graph in the initial data story. This was accompanied by some observations outlined in text. Now, these elements were separate in the conversation flow. I've pulled them together here. One thing that caught my attention in the text is that second bolded statement says, "Financial savings lags behind peers. " When I have used this example in workshops. It takes participants a significant amount of scrutiny before they get there. Something that took my tool mere seconds. So, that is impressive. However, if we continue reading, we see for the 8-year period, financial savings is tracked below the industry average. No, that is not the case. If we look at the graph at the left where financial savings is clearly above the industry average from 2022 through 2024. Copilot integrated some external context in its key takeaways with reference to the pandemic. And finally, Grok, after detailing a lot of observations, pulled the story together in this executive summary, one of remarkable resilience. Interestingly, that wording's quite similar to Gemini's resilient recovery. Who did it best? They all used a lot of words, mainly descriptive observations. And I found myself wishing that they had asked me for more context or confirmation of some of the assumptions they were making. Claude, as we saw, was the only one that included a graph in its initial output. And at first look, it was pretty decently designed. So, I think given that, I'm going to give this first round to Claude. But I do also want to compare how they each graph data. So, let's look at that next. My second prompt was to

graph the data. ChatGPT focused on financial savings and the industry average. It also zoomed in quite a bit. We could pick apart both of these decisions, but if I simply needed a quick visual of these two trends, this is a reasonable starting point. Gemini created three graphs. The first shows the top five performers together with the industry average. The middle one is industry average versus financial savings. That view looks quite a lot like the one we saw from ChatGPT. Even the color choices here are similar. And then that final one maintains the focus on our two lines of interest, industry average and financial savings, but also layers in the rest. I do really like this final view. We saw Claude's graph briefly already. One thing it did we haven't seen yet are some summary stats at the top. Also, this wasn't the only graph. This is the trend over time. You could also click on 2026 rankings and the biggest movers tabs. Though those weren't super interesting. Back on the main graph, I like some of the design decisions here. It's relatively clutter-free. I like how all of the lines are grayed out except the industry average and financial savings. I do wish the formatting of those two lines were reversed, so the industry average would be dotted and financial savings solid. It actually took me revisiting this a couple of times before I saw an issue. The lines in the graph have been smoothed. I actually hate this. If the lines weren't smoothed, each line segment connecting the points would visually show the relative change via the relative slope of the line. The smoothing distorts this. Bad call, Claude. Copilot did not have that issue. It chose to focus on the industry average and financial savings only, eliminating all of the other lines. You'll recall we saw that in Chat GPT's graph as well. Also, we've zoomed in here quite a bit. This isn't the end of the world, but does mean small changes get amplified, which may not be what we want here. Finally, Grok graphed all the data with sparing emphasis. This looks pretty good, though the text box explanation at the upper left feels a bit redundant with the legend at the bottom left. When it comes to graphing the data, who did it best? I'm going to disqualify Claude this round due to the smoothing. Chat GPT and Copilot's zooming isn't wrong in a strict sense, but it does make small changes seem like a big deal, which I don't think is right for this data. That means it's close between Gemini and Grok, who created very similar views. Given the redundancy of annotation and legend in Grok's version, Gemini created my favorite output from this round. With my next prompt, I shared the context that I wished these AI assistants had asked me for. I defined the ambiguous bank index, shared why I was looking at this data in the first place, and asked for specific output, an exec-ready slide. Here is the

specific prompt I gave. Let's see how they do. ChatGPT's looks pretty slick at first glance. I love the takeaway title displayed prominently at the top. Financial savings gained ground, but is now falling behind competitors. I also appreciate the sections at the right. Key takeaways, why this matters, recommended actions. When you read through that detail, it has some pretty good interpretations and suggestions. That said, I don't love the icons in the annotations. There's also a good amount of redundancy. Multiple levels of titles that say basically the same thing. Every data point is labeled plus the Y-axis labels. I'm nitpicking. It is definitely impressive that I could give some data and a couple of prompts and get something like this in return. I'll look next at Gemini. You'll recall it created my favored graph of the bunch. So, I was looking forward to seeing what it would do given this additional context. It landed here. It's denser than I expected. Again, I don't love the icons. The text when you read through it is also super jargony. On the positive side, I like the color choices and the general organization of the graph on the left and the words on the right. We saw that in ChatGPT's creation as well. Note also that there is a strategic recommendation included at the bottom right. Analyze competitor breakout strategies to regain greater average performance in 2027. Not sure exactly what that means, but I like the fact that there is a recommendation. Looking back at the graph, I also like how the competitor bank spread is shown as shaded gray. However, when I look closely and compare to some of the other views of the data, it turns out that gray area isn't actually the spread across banks. If it were, it would be a wider band than this and include a low outlier in 2022. I initially thought maybe it was showing the 10th and 90th percentile or something like that, but it literally says all 22 banks in the legend. So, that's outright wrong. Let's see what Claude comes up with. Claude turned things into a scorecard. The graph on the right's decently designed. It even got rid of that smoothing that had aggravated me in the previous view. I also like the takeaway title. Financial savings is improving, but the gap persists. I appreciate the headlines and supporting at the bottom. Next up, Copilot. This was a big departure from the graph it created previously. I like the takeaway title at the top, but after that, there's a lot of space devoted to relatively little text. At first, I thought it skipped graphing the data altogether, but then I noticed there is a graph at the bottom. Strangely though, the shape is totally different from the actual data. I thought at first this might be plotting only the beginning and end points, but looking at a little more closely, I think it's actually a data hallucination issue. Also, I assume when looking at something like this that the same color used in multiple places means the same thing. But if we look at the gold color used throughout the right-hand side, it's used for totally different things. Let's see what Grok has for us. This visual was accompanied by some supporting text that I've not shown here, including the headline, "Financial savings has delivered strong satisfaction gains, outperforming the industry average since 2022 despite a minor 2025 dip. " And we can see the gains and dip that it refers to annotated on the graph. Grok goes on to list bullets and a key story and a recommendation. ChatGPT and Gemini were the others who did this. Here, the recommendation was to continue investing in customer experience initiatives to solidify our competitive advantage and drive further gains in satisfaction scores. Who made the best exec-ready slide? This time, I'm disqualifying Copilot due to its graph issues. Gemini would have been a contender if it weren't for the issues with the shaded region. So, I'm going to say, given the takeaway title, relatively clean design, and recommended actions, I'm going to go with ChatGPT. Though, I really do want to get rid of those icons. One observation I'll share at this point. Going into this, I expected that with each prompt, things would get monotonically better. Meaning, the more I provided, the better each iteration would get. That doesn't actually seem to be the case, though. Sometimes we take steps forward, but embedded in that, sometimes are some steps backwards, as well. This underscores a general point of caution, which is just that even if everything is good at one step in the process when you're working with AI, don't assume that that's going to continue to be the case. We need to have a discerning look and a critical eye at every step during the process. We're not done yet. I gave my AI friends one final prompt. Do it like I do it. Specifically, revise

the slide to follow best practices from the book Storytelling with Data. I called out a few specifics. Eliminate clutter, focus attention, use a takeaway title, support a single key message, and recommend next steps. Let's see what they do here. As a reminder, here's what ChatGPT had given me for the exec ready slide. Things definitely feel a little lighter when we move from that to its storytelling with data inspired version. Partially due to the change in font and also some additional simplifying. Many of the icons were removed. Hooray! I like the annotation it added to the graph, highlighting in green the period of financial savings outperforming the industry, then in red where momentum was lost. It also reduced the text at the right, got rid of the heavy backgrounds and shading, which makes it easier to focus on what remains. Including at the bottom right, next steps. For those next steps, the suggestions focus on looking at drivers of our own past increases and decreases, which is probably a reasonable thing to do. Here's what Gemini's exec ready slide looked like. Note the descriptive title, that gray shaded competitor spread on the graph that we've discussed, the icons at the right that I complained about before. Gemini's storytelling with data version has a takeaway title, the shaded competitor spread, even though it's still mentioned in the legend, seems to have gone away. Uh but on the bright side, so have the unnecessary icons. The color scheme shifted from blue and green to blue and orange, which feels on brand for storytelling with data. I also like the bold next steps at the bottom right. Though I'm not totally sure looking externally as a next step makes as much sense as the internal focus that ChatGPT suggested. Gemini also included a script to go with the slide and tell the story verbally, from underperforming to turnaround to falling below average. Check out that final paragraph here. Our singular focus must be to reverse this trend. We recommend a strategic audit of the breakout leader so we can implement learnings and regain our above average status. Claude had originally put together a scorecard for the exec-ready slide. It did a total revamp when I asked it to apply storytelling with data principles. Note the takeaway title. Looks like there's some misalignment with the annotations and lines on the graph. That's a limitation of all of the AI that I used, which is that it provides an image back. So, you can't adjust individual elements. Even Co-pilot, who said it could provide a PowerPoint slide, I was never able to actually get it to do so. Back on Claude's slide, note also the recommended next step at the bottom to conduct a root cause analysis of the 2025 decline and use that to identify issues and how to address. You'll recall the dark slide with the tiny and incorrect graph at the bottom that Co-pilot provided. Well, in its remake, it doubled down on this, making that wrong data big, which was an unfortunate decision. It did suggest some interesting recommendations. Those are enumerated at the bottom. Diagnose top drivers of satisfaction decline through targeted customer research, benchmark specific experience touchpoints against the top five performing banks, and develop a 12-month action plan to close the gap. Here's a reminder of Grok's exec-ready slide. In its storytelling with data remake, it removed the background shading, added a takeaway title, and some bullets at the bottom left, plus a recommendation at the bottom. Continue customer satisfaction investments to maintain leadership and drive further gains. I found the maintain leadership part strange since financial savings is below the average in 2026. The final bullet at the left says, "Consistent leader since 2022," which isn't actually the case. This is another instance of some backwards steps relative to the last iteration. Who did the best job applying

storytelling with data principles? Well, to answer that, I feel like there's one more option we have to look at. My option. The version I designed before any of my back and forth with AI looked like this. Has a takeaway title followed by a graph that is designed exactly the way I want it, sparing emphasis to focus attention on what matters most, annotation to describe what we're seeing, and a next step, a discussion prompt that is targeted for my specific audience. None of the options I explored with AI did all of this. What have we learned? It's not a great idea to simply give data to AI and ask it to create a final product. That probably doesn't surprise anyone. And while I wouldn't be comfortable presenting anything but the version I created, AI did make some interesting observations and choices as I worked with it. It raised different approaches and ideas. And that's one thing we can get with AI, another perspective, brainstorming help, feedback, potential next steps to consider, ideas for how we might tell a

story, and how we can connect it with our audience. There are definitely ways to use AI in the data storytelling process to make things more efficient and robust. Actually, if that's something you're interested in learning more about, be sure to check out our July 2026 live event where we're going to show how you can partner with AI throughout the data storytelling process. Learn more about that at storytellingwithdata. com/ai. I'm curious to know how you're using AI as you work with and communicate data. Drop me a comment below and be sure to check out additional resources and that live event I mentioned at storytellingwithdata. com/ai. Thanks very much for watching. I hope to see you soon.