AI for data analysis at your company: how to get insights from your spreadsheets without being a data scientist

Asking questions in natural language lowers the barrier to analysis, but it doesn’t eliminate responsibility for the data. A convincing answer might use the wrong column, mix periods, ignore missing values, or attribute a cause where only a relationship exists.

Value comes from preserving a chain of evidence throughout the conversation with AI: source, definition, filter, calculation, result, and review.

# What can AI do with a spreadsheet?

Depending on the platform, plan, and workspace configuration, a tool may:

• describe columns and data types;
• calculate statistics and compare groups;
• identify trends, missing values, and outliers;
• explain or propose formulas;
• classify and summarize text comments;
• create tables, charts, and narratives;
• run code to transform data;
• answer follow-up questions;
• document assumptions and steps.

The documentation for ChatGPT for data analysis says it can work with spreadsheets and other files, perform calculations, and create tables or charts. It also recommends structured data, clear headers, and one record per row.

Copilot in Excel can produce summaries, trends, outliers, formulas, and other elements in the workbook. Microsoft warns that generated content should be reviewed and verified.

The AI function in Google Sheets can generate, summarize, categorize, and analyze text in ranges specified by the user. Its limitations and availability depend on the account and settings.

These capabilities are useful, but a list of features does not prove that a tool is suitable for sensitive data or critical decisions.

# What should AI not conclude on its own?

An assistant should not turn a correlation into a cause, invent business definitions, or decide which data is “correct” when sources conflict. It also should not:

• approve payments, credit, or hiring based solely on a generated output;
• combine periods or currencies without disclosing it;
• infer sensitive attributes about people;
• replace required financial, legal, clinical, or statistical review;
• publish a report without an accountable party and source version;
• present a projection as fact;
• hide excluded rows, missing values, or changes in methodology.

AI can suggest an interpretation. The organization must retain authority over the definition, calculation, and decision.

# How to prepare the spreadsheet before analysis

File quality determines how easy it is to verify the answer. It’s a good idea to prepare a working copy and keep the original intact.

Before uploading the file, remove any columns you don’t need. Replacing names with identifiers doesn’t always anonymize the data: dates, locations, or combinations of attributes may still identify someone.

# Three ways to use AI to analyze data

# Upload a file to an assistant

Useful for one-off exploration, new questions, and code-backed analysis. Before uploading information, the company should verify which product, account, retention, permissions, and terms apply.

# Work within Excel or Google Sheets

This keeps the analysis close to the file and can make it easier to work with formulas, tables, or columns. Even so, check workbook permissions, feature availability, and whether the output stays linked or is a static copy.

# Integrate a data source or platform

This is useful when the analysis needs to be repeated, respect permissions, use shared definitions, or feed a dashboard. It requires more design: identity, refreshes, logs, testing, and an owner for the semantic model.

The best approach may change. A file works for testing the question; a connection or API may be necessary once the result becomes part of operations.

# Seven-step analysis method

# 1. Formulate a decision question

“Find insights” is too broad. Use a question with a period, population, metric, and expected outcome: “Compare margin by category between the second and third quarters, using net sales and direct costs; identify the three largest changes.”

# 2. Define the metric

Explain what net sales, active customer, resolved case, or margin means. Specify inclusions, exclusions, currency, taxes, and cutoff date.

# 3. Describe the file

Name the sheet, relevant columns, the unit represented by each row, and the relationships between tables. If a data dictionary exists, include it without unnecessary sensitive information.

# 4. Request a data quality assessment

Ask it to identify data types, duplicates, missing values, impossible values, and format changes before calculating. Decide how to handle each finding.

# 5. Ask for the method, not just the answer

Request the filters, formula, code, or intermediate table. A verifiable result makes it possible to reproduce the calculation outside the conversation.

# 6. Compare with known cases

Test a sum, period, or segment whose result you already know. Also include an exception to check how the tool handles incomplete data.

# 7. Document the decision

Record the file and version, question, method, filters, person responsible, date, and limitations. If the analysis is repeated, turn this evidence into a regression test.

# Which questions produce better results

A useful instruction combines context, calculation, constraints, and format.

Avoid asking for strategic recommendations before checking whether the data can answer the question. A descriptive analysis can show what changed; explaining why may require additional sources or a causal design.

# How to verify an AI-generated analysis

The review should be proportional to the impact. At a minimum:

• confirm that the number of rows matches expectations;
• check filters and periods;
• reproduce key totals using an independent formula;
• inspect a sample of records;
• compare denominators, currency, and units;
• look for duplicates after joining tables;
• review missing values and exclusions;
• distinguish calculations, interpretations, and hypotheses;
• keep the code or steps when available;
• requests review by the person who owns the metric.

A chart doesn’t validate the analysis. It can accurately depict an incorrect calculation.

# Security, privacy, and permissions

Before using a tool, classify the data and confirm that the account is authorized. Ask:

• which product and plan is the person using?;
• who administers the workspace?;
• what data can be uploaded, and for what purpose?;
• how are access, retention, and deletion controlled?;
• does the data include personal, contractual, or confidential information?;
• does the file contain more data than necessary?;
• who can reuse or share the output?;

The NIST AI Risk Management Framework proposes governing, mapping, measuring, and managing risks throughout the life cycle. For business analysis, this means the file, method, tool, and decision form a single system of accountability.

Don’t upload real data to a personal account to demonstrate an idea. Use a synthetic sample, anonymized if that’s sufficient, or an authorized business environment.

# Hypothetical example: sales by region

A director wants to understand why the quarterly margin fell. The file contains orders, returns, costs, discounts, and region.

The responsible workflow would be:

1. define margin and the recognition date;
2. review duplicates, returns, and currencies;
3. reproduce the official report’s total;
4. compare equivalent periods;
5. break down the change by volume, price, discounts, and mix;
6. identify segments that explain the variation;
7. separate facts from possible causes;
8. review the result with finance and sales.

AI can speed up exploration and produce code or tables. It can’t know on its own whether a discount was an authorized exception, whether the accounting policy changed, or whether the cost data arrived late.

# When to move from a spreadsheet to an integration

Consider a connected data source, a semantic model, or an API-based workflow when:

• the same question comes up frequently;
• multiple people need a single definition;
• manual file handling leads to version errors;
• permissions should be inherited from corporate systems;
• the result feeds into another action;
• monitoring and traceability are required;
• the volume exceeds what can reasonably be reviewed in a spreadsheet.

Integration shouldn’t automate a poorly defined metric. First consolidate the question, source, and validation; then decide on the architecture.

# Frequently asked questions

# Can I analyze Excel files with ChatGPT?

Yes, ChatGPT supports spreadsheets in compatible experiences. Availability depends on the plan and configuration; review the code, assumptions, and results before using them.

# Do I need to know how to program?

Not for exploring basic questions. For recurring or critical analyses, it’s helpful to have someone who can review formulas, code, statistics, and business definitions.

# Can AI automatically find trends?

It can suggest trends and outliers. You should verify the period, method, seasonality, missing values, and relevance before interpreting them.

# Is it safe to upload business data?

It depends on the product, account, configuration, contract, and type of data. Classify the information and use only authorized environments.

# Can an AI response replace a dashboard?

Not necessarily. A conversation helps with exploration; a dashboard maintains definitions, updates, and monitoring. They can complement each other.

# When does an API make sense?

When the analysis needs to be repeated within a process, connected to systems, subject to controls, and produce its own records.

# From a quick question to a verifiable decision

AI makes analysis more accessible, but trust depends on evidence. Preparing the data, defining the metric, and verifying the method help prevent speed from creating false confidence.

ia¹ helps assess the use case, data, platform, and integration path. Request a feasibility assessment.

# Primary sources

• OpenAI — Data analysis with ChatGPT
• Microsoft — Get data insights with Copilot in Excel
• Microsoft — Get direct answers to data analysis questions
• Google — Use the AI function in Google Sheets
• NIST — AI Risk Management Framework Core