Can Gemini Perform Data Analysis? (How Data Analysts Should Approach It)

Since Gemini can now perform data analysis directly in Google Sheets, I decided to try it out right away.

At first, I thought about writing a light, tutorial-style article, but as I started running it using people analytics as a subject, it became increasingly fun, so I challenged myself to create a demo video.

When I uploaded the finished demo video to X, I received positive feedback. Thank you to everyone who watched it!

Now, can Gemini perform data analysis? It worked much more smoothly than I expected. First, please take a look at this demo video.

In the approximately 4-minute video, Gemini is active for about 3 minutes. Since it is edited for the video, the actual operating time was about 10 minutes.It is surprising that it can do this much in 10 minutes.

About a year and a half ago, I uploaded an analysis example using ChatGPT's Advanced Data Analysis to note, and it is clear that technology has advanced significantly since then. I tried Gemini this time, but other LLMs have surely advanced as well.

Compared to when I tried it before, I was happy to see that analysis could proceed with rough requests. For example, in the demo video, with just one sentence, 'Analyze the relationship between overtime hours and engagement,' it...

• Considers and decides on an analysis strategy.

• Exploratorily finds data columns from ambiguous names.

• Analyzes using correlation coefficients and scatter plots, taking into account that they are quantitative variables.

It breaks it down into these actions.

At this point, you might wonder what will happen to the job of a data analyst in the future. However,looking at it in detail, I still have the impression that it is difficult to leave analysis to be done automatically without supervision, and I felt that accompaniment by an experienced person is necessary.

I will consider why I felt that way, and how it would be best to use Gemini as a data analyst.

Note: I am running this on an enterprise Google Workspace, but as of April 4, 2025, it is treated as an Alpha version. Please be aware that there may be various changes in the official version.

It performs simple analysis quickly

I started using Gemini in the Google Sheets side panel on the morning of April 1st, and it was so much fun because it worked more smoothly than I thought, so I spent the whole day playing with it.

As shown in the demo video, it takes about 10 minutes to run this analysis scenario. However, I did not get the desired results on the first try, and I ended up doing about 20 retakes to make the video.

Also, there are points to be careful about regarding the output of the demo video, and those issues are explained in detail on the Kuni-Lab technical blog. It is over 10,000 characters long, so please take a look when you have time.

First, let's consider the advantages of data analysts using Gemini.

The first is that it understands ambiguous requests, so you can quickly test rough ideas.

Especially in people analytics, various ideas emerge during discussions, so the ability to capture those ideas and input them into a prompt for quick verification is very attractive.

It is amazing that it can analyze more flexibly than using fixed BI tools, and even organize insights and actions.

An analyst with moderate adaptability

On the other hand, the challenges I felt when performing data analysis with Gemini can be summarized as follows.

• The analytical approach can change every time you run it. It often chooses suboptimal analytical methods.

• When visualizing data, there are times when it interprets the graph in detail to provide insights, and times when it does not.

• Data processing methods become inconsistent when given vague instructions.

In short, sometimes it works as intended and sometimes it doesn't, and even with the same prompt, the results can change. While refining the prompt can resolve some of these issues, it is not always effective.

Compared to a real person, my impression was that it is an "analyst who works quickly but has moderate adaptability".

Gemini can handle vague instructions and use a wide range of analytical approaches, so it is not a complete beginner.

However, even though it possesses the knowledge of a mid-level data analyst capable of handling complex and advanced requests, it sometimes produces output that makes you wonder, so it is still unsettling to leave it entirely to its own devices.

In other words, a human needs to verify the analytical approach, process, and results, but evaluating the analytical output appropriately requires a certain level of skill.

It still seems difficult to expect it to function like outsourcing data analysis tasks to other teams or internal vendors. In this regard, it has not yet reached the level of the Deep Research services that have been appearing one after another recently.

Therefore, it is necessary for someone with a certain amount of data analysis experience to guide Gemini while working alongside it.

Challenges for the person working alongside it

After playing around with it, I noticed that there is aslightly tricky problemfor the "partner" who makes requests to Gemini.

That is, the task of utilizing prompts is far removed from the act of understanding data while working with it manually. Even if Gemini provides a good response, I felt that it is difficult to deepen one's intuitive understanding.

At this stage, rather than leaving it to run on its own like Deep Research, it is necessary to deepen the prompts step by step. Therefore, the partner themselves must also understand the data and the background of the phenomena through the data.

However, the behavior of generative AI is basically passive; it does not perform tasks beyond the user's questions, and its inquiries to the user are minimal. Therefore, I felt that ingenuity is required to expand the partner's ideas.

As an image, it requires the ability to effectively manage an associate who has advanced knowledge and implementation skills but lacks practical experience. This means it is difficult for general managers or beginners, and that the user acting as a partner needs a certain level of practical experience.

Having literacy to use generative AI effectively

To summarize what has been discussed so far, while I believe that using Gemini can indeed help you handle data analysis efficiently, it is premised on the fact that the user (the generative AI's partner) themselves possesses data analysis literacy.

Some might think, "That sounds exhausting," but in a sense, it might be a healthy thing.

When you think about it, the basic subjects we learn in compulsory education and high school may include things that we rarely use as-is in the real world, but they can also be said to be a condensed collection of knowledge that should be known in modern society. It is what we call "reading, writing, and arithmetic."

In 2022, "Information I" joined those basic subjects. Computer science has become one of the pieces of knowledge that should be known in modern society.

Looking at the Information I textbook, you can see that it includes the basics of data science. For example, I have the high school textbook my daughter used at hand, and it contained the following content.

Seeing this, I was surprised and also thought, "That's great, I wish I had learned it this way."

What a fun era it has become... that might be an exaggeration, but at least for those interested in this field, it is a good world line.

In other words, the younger generation is going out into the world with basic data analysis literacy, and they will likely be able to use generative AI well. If that happens, it would not be surprising if people emerge who can "run analysis well" using generative AI.

Of course, since generative AI is still in development, there is a possibility that analysis tools like Deep Research will emerge before we know it. In fact, generative AI is already beginning to enter the field of programming.

There are diverse approaches to data analysis

However, unlike programming, a certain ambiguity lurks in data analysis.

If requirements are solidified, one can define a program that becomes the correct answer, and even if there are various ways of coding or designing, the output must follow the specifications. Such tasks are where generative AI excels.

On the other hand, in the world of data analysis, especially business data science, it is necessary to consider ambiguous hypotheses and complex mechanisms from incomplete data. In other words, the approaches are diverse.

Inferring mechanisms or predictive models from data as a result in a state where the causal structure is not clear is, mathematically speaking, like an "ill-posed problem." Therefore, data analysts and data scientists are constantly innovating every day.

Even if generative AI becomes able to handle data analysis at the level of a mid-level data analyst, the phase of breaking down business problems into analysis tasks and solving difficult questions through ingenuity will remain. This is because these phases require "intent".

Therefore, while looking forward to the coming development of generative AI, I believe that raising your data analysis literacy and honing your applied skills will not be in vain.