Using AI to analyze data (Excel/CSV) correctly

The failure modes of naive spreadsheet prompting

When practitioners drop a CSV into a chat interface and ask, “What stands out?”, they are combining three hard problems: parsing, aggregation, and interpretation. Models may silently misread headers, infer data types incorrectly, or ignore edge cases such as missing values and outliers.

We model tabular analysis as a pipeline of narrow steps, each of which can be validated independently. Instead of a single sprawling prompt, we run:

1. Schema extraction.
2. Sanity checks and cleaning.
3. Question-specific transformations.
4. Interpretation with explicit uncertainty.

Step 1: Schema extraction

The model first summarizes the dataset:

• Column names and inferred types.
• Ranges, missingness patterns, and obvious anomalies.
• Row count and any partitioning (by user, by experiment, by time).

This summary is returned as a small JSON document that can be inspected, logged, and diffed. If the schema is wrong, later analysis is automatically suspect.

Step 2: Cleaning with reproducible operations

Cleaning instructions are expressed as operations against columns: filter predicates, imputations, deduplications. The goal is to produce a minimal script (SQL, pandas, or spreadsheet formulas) that you can run on the original data.

The model is not trusted to directly mutate the canonical dataset. Instead, it proposes a pipeline which is applied and version controlled like any other piece of code.

Step 3: Focused questions

Once the schema and cleaning steps have been validated, you pose narrow questions:

• “What is the distribution of revenue per active account by region?”
• “Which cohorts show statistically meaningful churn deltas after week four?”
• “Which metrics spike before an incident?”

The model reasons over pre-aggregated views or summary tables, with explicit units and definitions.

Step 4: Interpretation with uncertainty language

For each answer, we ask the model to:

• State the conclusion in plain language.
• Reference the exact columns and filters used.
• Describe at least one alternative explanation or confounder.

This pattern nudges the model away from overconfident narratives and towards a more analytical, scientist-like tone.