Why Can't AI Analyze Excel Data? The Surprising Weakness of Smart Models

Imagine this: you’re at the office holding an Excel file containing tens of thousands of rows of sales data. You ask the smartest AI in the world today to analyze this ‘table,’ which summarizes who sold what, when, and how much, broken down by product every month. But the AI replies with something completely off-base, saying, “Hmm, this data just reads like an ordinary story.” Why does an AI that should be accurate at numerical calculations make such mistakes?

Recently, Large Language Models (LLMs)—AI trained on vast amounts of text to converse like humans—have shown incredible abilities in summarizing sentences we write, analyzing difficult research papers, and writing complex programming code. However, they actually lag behind traditional statistical methods used for over a decade when it comes to analyzing ‘tabular data’ like Excel sheets or databases Source 10, Source 11.

Why is this important?

In modern business and research settings, the vast majority of core data exists in tabular form. Important decisions regarding financial reports, customer purchase history, and clinical trial results are all made using these tables of numbers. If the most advanced AI cannot properly understand this core data, companies must continue to rely on outdated analysis tools and fail to fully enjoy the benefits of cutting-edge AI. To become the ‘smart assistant’ we expect AI to be, it must overcome this hurdle of numerical data analysis.

In simple terms: AI reads tables as ‘sentences’

Let me use an analogy to explain why AI struggles with tabular data.

The technology called the ‘Transformer’—the core structure of AI that extracts meaning by identifying relationships between words in a sentence—was originally born for ‘language.’ Simply put, AI is trained to find the ‘flow of a story’ that moves from left to right when reading text.

However, when it encounters tabular data, the AI begins to read by forcibly converting (serializing) the table into text, as if reading a novel written in a foreign language Source 9. It goes, “Row 1, Column 1 is sales, Row 1, Column 2 is product…”

This is where the problem occurs. A table is not a ‘story.’ A table is a two-dimensional space where rows and columns are connected independently or in very complex ways. AI instinctively tries to read sentences that have a sequence, but a table is a multi-dimensional chunk of information that has nothing to do with sequence. It is similar to trying to find your way by looking at a map while only reading a list of names written in order, rather than looking at the map itself. Source 9

Furthermore, while AI is accustomed to basic arithmetic like addition when analyzing data, it is clumsy at creating complex logic on its own, such as ‘grouping and aggregations,’ which are essential in actual data analysis Source 3, Source 8. AI has yet to ‘learn’ the logical analysis level of a human creating a pivot table in Excel.

Where we stand: AI analyzes by ‘guessing’

Currently, many AI models rely heavily on the names of columns (identifiers) written in the table rather than deeply understanding what the data itself is Source 12. For example, if it sees a column named ‘Sales_Amount,’ the AI guesses, “Ah, this is the sales amount.” However, if you change this name to meaningless characters like ‘col_01,’ the AI’s performance drops sharply Source 12. In other words, it is not deeply interpreting the actual data values but rather guessing based on the name tags (metadata) left by humans Source 6.

Because of these limitations, traditional machine learning methods based on Decision Trees are still much faster and more accurate at analyzing tabular data in the field Source 11.

The way forward: A true data analyst

Moving forward, research into ‘Data Language Models’ that allow language models not only to excel at text but also to understand the structure of the table itself will become active Source 6. The day will come when, if we ask while looking at tabular data, “What is the product that sold the most here?”, the AI will not just guess by looking at name tags but will accurately perceive the structure of the table and mathematically aggregate the results to answer.

However, right now, rather than entrusting 100% of important management numerical analysis to AI, it is wise to use it as a supplementary tool in fields like text summarization or code generation.

MindTickleBytes’ AI Reporter View

Because language models have learned the world’s knowledge through text, they treat tables full of numbers as a ‘strange language.’ But the moment AI learns how to combine mathematical logic with linguistic insight, our work efficiency will accelerate at a speed incomparable to now. Until then, please use AI only as your ‘genius assistant.’

References

1. [Source 3] Large Language Models Engineer Too Many Simple Features for Tabular Data (https://arxiv.org/html/2410.17787v1)
2. [Source 6] Large Language Models(LLMs) on Tabular Data: Prediction, Generation, and Understanding – A Survey (https://arxiv.org/abs/2402.17944)
3. [Source 8] Large Language Models Engineer Too Many Simple Features for Tabular Data (https://arxiv.org/html/2410.17787v2)
4. [Source 9] The Illusion of Generalization in Tabular Language Models (https://arxiv.org/html/2602.04031v2)
5. [Source 10] Small Models are LLM Knowledge Triggers for Medical Tabular Prediction (https://arxiv.org/html/2403.01570v3)
6. [Source 11] Language Models Are Good Tabular Learners (https://openreview.net/forum?id=r8tMECbxOl)
7. [Source 12] Data Language Models: A New Foundation Model Class for Tabular Data (https://arxiv.org/html/2605.06290v1)