I think it is fair to say for the most part, a simple bar graph – whether horizontal, vertical, or stacked – meets most needs. The benefit with bar graphs is that you can easily compare categories either in a single or multiple points in time (granted, line charts are better at changes over time). There are several chart guides and catalogs out there that aid in the selection process and these are typically the places I go to directly when I need inspiration or validation for my own selections (I admit to my own biases). What I have been experimenting with lately is working with AI to explore how to visualize data differently. This post describes my process, including the steps I took, the prompts I shared and iterated, the things I kept and discarded (and why), as well as learnings.
The work I do using AI varies. Sometimes I use it to help with formula generation in Excel, other times with creating a more efficient CASE expression or simplifying an overly complicated query (figuring out whether a Union or a subquery is the better way), other times with converging documentation saved in multiple places on SharePoint and Outlook. Recently, I’ve experimented with creating visuals that support a story I am trying to tell with data. I intentionally worked with a simple scenario because a) starting small and specific allows me to have an existing solution in mind for which I can easily explain the process, b) concurrently evaluating AI on a pre-existing solution helps me document differences between my non-AI approach and the AI-assisted one, and c) I will always have the original solution upon which to fall back should the AI one fail.
To be clear, I do not use AI to create the final visual (graph or dashboard). I use it to gather ideas that I can later – should I choose – create myself in tools like Tableau, Power BI, Excel, or even Canva for infographics. I retain ownership of the actual development, hence, final product.
The Scenario
This is a scenario that I created using fictitious and structured class enrollment data (synthesized using AI) and unstructured survey data. I wanted to find a way to visualize any correlation between enrollment numbers and survey responses. After having aggregated the enrollment numbers (for HISTORY 218 - Disease & Society in World History) that I chose to group by major vs non-major, I created a basic line chart in Excel. I went with a descriptive and slightly provocative headline that I iterated a few times before having committed. Each iteration has its merits. I won’t dive into those here, but feel free to consider what those are and share your thoughts in the comments below.
First iteration: illustrating trends broken out by major
Second iteration: combining non-majors to emphasize the message
Third iteration: is a slope graph better to emphasize the drop?
Ultimately, I decided that slope is not a great chart because it hides patterns and masks the big picture. My argument is not how big the drop was, but that a major change of the way the class is scheduled and offered has impacted enrollment.
Fourth iteration: back to the reliable line chart!
I then focused on one survey question (“What was the single biggest reason you did not enroll?”)
First iteration: breakdown by reason
Second iteration: rolling up the top three reasons that relate to scheduling/ offering
Thinking about my final deliverable – a single slide – I came up with a headline: “HIST 218 isn't losing interest — it's losing access. Every bit of the enrollment decline is non-majors, blocked by a once-a-year, Fall-only morning schedule.”
To support this headline, I wanted to consider my visualization options, which I normally play around with in Excel or Tableau. This time, I ventured out to Copilot in Excel.
Practical Application
With the content and context already there, I created my prompt:
I have a fictitious course enrollment data set I used to create multiple iterations of visuals that will help to support my headline that I generated based on an analysis of my data. I want to show 2-3 visuals in a single PowerPoint slide considering my takeaway “HIST 218 isn't losing interest — it's losing access. Every bit of the enrollment decline is non-majors, blocked by a once-a-year, Fall-only morning schedule.” My audience is a group of higher ed staff who work with data as part of their jobs in different capacities. Review the Excel workbook that includes the data and visuals. I haven’t settled on the most effective ones yet, hence, the iterations. Generate 2-3 charts that might work better at delivering my takeaway. Render the data, including pivot tables and charts, and explain in 2-3 sentences why you selected them, why they might work better, why they might not.
One thing to note: generative AI models like Claude Opus 5 do an impressive job at creating usable Excel workbooks that include formulas and pivot charts and tables. So does Microsoft Copilot set to select the model automatically. I chose the latter. It took several minutes for Copilot to compute and finally render anything.
The Rendering (but not the "reckoning")
This is what Copilot finally rendered and at first glance I say, “not bad!”
Copilot leveraged the data set and visuals it found in my workbook, and responded to my initial prompt with the following:
- It logged 9 steps to build the single slide layout that I asked.
- It created multiple pivots tables that were used to create the charts.
- It built 3 visuals, including major vs non-major enrollment trends, interest score distribution (which was not part of the scope in my prompt).
- It verified the data in the original set and included additional context.
- It let me know why each chart works well to establish the story of the enrollment decline primarily due to scheduling reasons. It also warned me that the survey information does necessarily explain the cause for the drop in enrollment.
- It concluded with guiding me on how I might incorporate this in a new slide (something I did not explicitly ask for).
My Own Touches (because the first version is the AI’s, everything else is mine)
While Copilot generated some good recommendations, it was just the start. I continued to refine the prompt a few more times, remembering the takeaway I wanted to stick to. Another part of this might have been for me to ask AI to find the headline for me, but I prefer to keep that part of the process to myself. There is substance and importance in this part of data analysis: the discovery. That is not to say that in the future I might not leverage AI to assist with finding the story within the data (perhaps a future blog).
Prompt refinement #1:
I do not want to focus on the interest score part of the survey or determine cause. remove the visual labelled "72% still rate the topic 4-5"
Prompt refinement #2:
I want you to flip the order from largest to smallest count on the bar graph in the right. Remove the chart axes headings as well as the gridlines. Only keep total counts next to each bar. Change color palette to dark teal and gray, accounting for accessibility. Use the same color palette for the first visual, with teal for the non-majors.
With each refinement (and there were 3-4 more but I will spare you the details), I was provided with a reasoning step, but something else that I hadn’t expected…
Changes AI made without consulting me
What I noticed was that with each refinement, other parts of the worksheet changed, without me asking Copilot. For example, the heading at the top moved from “HIST 218 isnt’ losing interest – it’s losing access” to “HIST 218’s enrollment decline is concentrated among non-majors.”
Final AI Iteration and how I refined it for my final slide
This is the final iteration Copilot came up with following my refined prompting. What I appreciate is the speed with which it can make aesthetic adjustments on command rather than me needing to do them manually. What I don’t like is that it made changes to heading and subheadings without asking first. This matters because those changes affect the context of the data, the visuals, and ultimately the story being told. If users miss these subtleties, the final deliverable can shift in meaning, potentially undermining credibility if someone later catches the change and the user is caught off guard. Also, not a fan of how it tried to fit in both charts in a narrow dashboard looking frame. While the format is visually awkward for listing survey responses in a bar graph, it creates a bigger problem for the line chart by distorting the scale and making the differences in counts appear more dramatic than they are.
Listed below are the changes I made intentionally myself (manually) to the visuals, in no particular order:
Line chart adjustments
- Removed the redundant axis titles. The chart title’s job is to explain what it being counted and across which timeframe. Less is more.
- Removed the legend for less visual noise. Included measure descriptions at the end of each line, following the color association (dark teal for non-majors and gray for history majors).
- Changed the default shapes for the series values in order to make it more accessibly discernible. Kept the circle for non-majors but selected a square for majors.
- Kept the new chart title; it is descriptive enough. It would be unnecessary to include a reference to HIST 218 because the audience already sees that in the headline above.
- Added a vertical, dotted line in between 2022 and 2023 to highlight when the change in scheduling policy occurred and highlighted in yellow the moment following the change in order to emphasize subsequent enrollment changes (drops).
- Changed the color palette to something more modern yet minimal but with accessible colors.
- Kept the original slide title but altered the subheading.
Bar graph adjustments
- Rolled up the top three reasons into a single one for simplicity. All three relate to offering any way, but I made sure to include that decision in the citation below the chart.
- Changed the color palette to something more modern yet minimal but with accessible colors.
Added qualitative content
I went back to the survey data because I wanted to see whether I could incorporate an element of empathy in the single slide. I decided to pair this qualitative information with the quantitative charts, remembering Cole Nussbaumer Knaflic’s exclamation: "Numbers provide credibility but the quotes and images create connection."
The Final Slide
After the previous adjustments were made to the source workbook, I then finished by compiling everything into PowerPoint. Here is the finished product:
Considerations and Learnings
1. AI might suggest chart types that are not typically used or understood by common business users. They might be visually interesting (or perhaps just complicated) but more useful for statisticians than your average business analyst or academic advisor. Make sure to clearly delineate who the consumers of your visual are so as to guide the AI’s chart choice based on this necessary context.
2. Remember that the first AI responses are not the final ones. You will most likely need to refine your prompts multiple times before you can get back something truly useful and relevant to your needs.
3. Avoid sharing any sensitive information in any AI tool, including possibly even home-grown versions supported by your institution. Always make sure to check [frequently] with your institution’s AI and IT governance policies.
4. Provide the AI with as much institutional knowledge and business context as possible because it does not understand any of that without clear direction.
5. When asking the AI to create the visual, remember that it will focus on the data layer first. If you don’t provide it with the story you have chosen to focus on, it will compile something that may not jive with your message.
6. Not all AI tools are created equal. What you ask Claude, ChatGPT, or Copilot to do will almost always render something uniquely different. The version you use also plays a big role in the quality of the visual or image. For example, the free version of ChatGPT may struggle to render a crisp chart, whereas the basic version of Claude might produce a cleaner result with the same prompt. In general, paid tiers are more likely to produce the strongest results. My recommendation is to test 3 tools with the free versions. If you have access to a paid subscription (hopefully at zero cost to you), then I’d start with that.
7. Remember that generative AI works with pattern recognition and its output is basically a set of probabilities (Kiran, 2026; Neo VA Solutions, 2026). In other words, the text we see is the model calculating probabilities one token at a time, choosing the next most likely option based on the patterns it has learned. The result of these probabilities may not be the right output for you. Don’t forget about what Geoff Hinton, the godfather of AI, said about AI getting things wrong in a June 2024 interview:
“What the big chatbots do is they look at strings of words and they're trying to change the weights in the network so that they can predict the next word. And all of the knowledge they have of all the strings of words they've seen is in the weights on those connections. And when you get them to recall something, what they're really doing is regenerating it just as you are with people. And so they're always constructing these memories and there's actually no difference between a real memory and a fake memory except the one's right from the point of view of the person constructing it. You don't know which is real and which is fake. You just say what seems plausible to you. And the chatbots do the same. Now the chatbots are worse than people at confabulating, but they're getting better.”
Final Thoughts
Could I have created this slide without AI? Yes, confidently, but probably because I have several years of experience in data visualization. I am also aware that this was a very simple data set and ask. I think if I am provided to work with more complicated data sets, I might use AI to help in several different ways, including summarizing it, cleaning it, visualizing. I’d expect that I would have to refine even more. The more complex the data and task, the more important it is to watch for where AI may miss the mark against our defined scope and context. Recently, I have also started using AI to create dashboard mockups that I may later build in Power BI or Tableau, knowing that the AI output itself is not a sustainable final product.
At minimum, AI can help surface possibilities we may not have considered, although a thoughtful colleague can do the same. Reach out to someone in your office or to a fellow analyst through the HEUG Connected Campus community of practice. Analysts such as @Scott Nishizaki, @Gaurav Ohri, @Pravin Mate, @Sagaya Alphonse, @Jacqueline Ang, @Sathish Sundaramoorthy, and many others have experience AI and/or analytics tools and can offer valuable perspective.
One final suggestion (if you did make it through this lengthy post): join me in a two-part Data Story webinar and explore applications of design and narration that might enhance your own data visualizations. There is no AI in this one but you will find some of what was discussed here in this presentation.
Register now!
Data Stories by Design: Narrative Foundations for Clear Presentation (Part 1)
Data Stories by Design: Narrative Foundations for Clear Presentation (Part 2)
Did this post inspire you, challenge you, or leave you with questions? What stood out most? If you’ve experimented with AI-assisted visual selection or data storytelling yourself, I’d love to hear what you tried and what you learned.
Sources:
Kiran. (2026, January 8). How probability drives every AI answer you see. Medium. https://medium.com/@Kiran_crispy_/how-probability-drives-every-ai-answer-you-see-b9c881e06397
Neo VA Solutions. (2026, February 17). How AI generates responses: From prompt to prediction explained. https://www.neovasolutions.com/2026/02/17/how-ai-generates-responses-from-prompt-to-prediction-explained/
The Naked Scientists. (2024, June 25). Geoff Hinton: Why does AI get things wrong? https://www.thenakedscientists.com/articles/interviews/geoff-hinton-why-does-ai-get-things-wrong