
What Can AI Actually Do for Data Analysis?
Carlos GarciaPublished Every major software vendor now sells you AI for data analysis, and every one of them describes it in roughly the same language: upload your file, ask a question, get an answer.
The marketing converges. The products do not. One of them runs Python in a sandbox. One of them can write a working Excel model. One requires organizational cloud capacity that your Pro license does not include. And one feature that spent a year in every "best AI for Excel" roundup was switched off in September 2026.
This guide works from vendor documentation only — what OpenAI, Anthropic, Microsoft and Google actually publish about their own tools. That turns out to be a different picture from the one secondary coverage paints, and the gaps are the most useful part.
What Can AI Actually Do for Data Analysis? The Short Answer
AI is genuinely good at three things with your data: writing and running analysis code so you do not have to, describing what is in a dataset you have not read yet, and producing the artifact at the end — a chart, a pivot, a summary, a model.
It is unreliable at exactly one thing, and it is the thing that matters: being right about numbers without being checked.
That is not a critic's framing. It is the vendors' own. Microsoft's documentation for Copilot in Excel says it "uses AI to generate suggestions and can sometimes make mistakes, misinterpret information, or produce inaccurate results," and instructs you to "Review, edit, and verify anything Copilot creates before you rely on it."
Using AI to build reports nobody acts on is just faster waste. Get a free audit and find out which of your analytics and search metrics actually move the business.
So the honest shape of the answer is this: AI collapses the time between having a question and seeing a candidate answer, and it adds a verification step you did not previously need. Teams that treat that trade consciously get real value. Teams that skip the verification step ship wrong numbers faster than they used to.
Free audit
Where do you stand in AI search?
A manual SEO + AI search audit of your site: how ChatGPT, Gemini, Perplexity and Claude read and cite it today, and what to fix first. Report within 48 hours, no credit card.
Get my free auditOr have it all done: AI SEO services, $999 once
What AI Can Actually Do With a Data File
Underneath the branding, there are only a few mechanisms at work.
Write and Run Code
The strongest capability is the least glamorous. ChatGPT's data analysis runs real Python in a sandbox that OpenAI says "is pre-loaded with hundreds of Python libraries." You are not getting a language model's impression of a correlation; you are getting a computed one, from code you can inspect.
That distinction is the single best predictor of whether an AI answer about your data is trustworthy. Code that ran can be audited. A number that was generated as text cannot.
Describe a Dataset You Have Not Read
OpenAI documents ChatGPT's data analysis handling t-tests, ANOVA, correlation and regression, ARIMA and seasonal decomposition, alongside data-quality checks for missing values, outliers, duplicate rows and incorrect data types.
For a file you have just been handed, that profiling pass is often worth more than the analysis. It tells you what you are working with before you commit to a question.
Produce the Finished Artifact
Charts, pivots and documents are where AI saves the most visible time. OpenAI says "you can create interactive tables and charts from your uploaded data," though interactivity is narrower than it sounds: "Currently, only bar, pie, scatter, and line charts are currently interactive in most cases."
Anthropic goes further on output formats, documenting that Claude "can create Excel spreadsheets (.xlsx), PowerPoint presentations (.pptx), Word documents (.docx), and PDF files," including "financial models in Excel with working formulas."
The Five Tools People Mean by "AI for Data Analysis"
These are the five that searches for this topic are almost always about. They differ more in access than in capability.
ChatGPT
OpenAI's help center calls it "Data analysis with ChatGPT," subtitled "A personal AI analyst for business users and experienced data professionals." The underlying tool is still named Code Interpreter in OpenAI's documentation.
Limits are documented and specific: "512 MB per file," and for tabular data, "For CSV files or spreadsheets, the file size cannot exceed approximately 50MB, depending on the size of each row." Note that OpenAI itself hedges that figure. Up to ten files per conversation, and the sandbox has no outbound network access.
Availability is the surprise. OpenAI's free-tier FAQ lists "Analyze and extract insights from your data" as a free-tier capability, with the caveat that "File uploads, image generation, voice, data analysis, and other tools have separate usage limits."
Claude
Anthropic documents uploads of CSV, XLSX, PDF, DOCX, JSON and more, with a gate worth knowing: "You must enable code execution and file creation in your account to upload XLSX files." The code-execution feature carries a 30 MB per-file limit in both directions and, Anthropic warns, "will use more of your limit compared to normal chats with Claude."
There is also a dedicated Excel add-in. Anthropic states that "Claude for Excel is generally available to Pro, Max, Team, and Enterprise plans" — so unlike ChatGPT's data analysis, there is no free route in. It does cell-level cited answers, assumption changes that preserve formula relationships, and native operations like pivot tables and conditional formatting. It explicitly does not support data tables, or macros and VBA.
Microsoft Copilot in Excel
Microsoft describes the core behavior simply: "Ask Copilot about your data, and it returns insights as charts, PivotTables, summaries, trends, or outliers," and it "can create new columns or rows that calculate values from your existing data."
Licensing is where it gets complicated. Microsoft lists four separate routes to eligibility, spanning consumer subscriptions with an AI credits plan, Microsoft 365 Premium, a commercial Copilot subscription, and Copilot Chat-eligible business plans. Microsoft adds that availability "may vary depending on your license and (for enterprise users) your organization's tenant configuration."
One hard constraint catches people: "Copilot editing is only supported when Calculation Options are set to Automatic."
Power BI Copilot
Microsoft's description is broad — Copilot "provides chat-based experiences" covering "tasks from on-the-fly analysis for business users to data analysis expression (DAX) generation for advanced creators."
The licensing requirement is the fact most published guidance gets wrong. Microsoft's requirements state "Paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher). Trial capacities and free SKUs aren't supported," and then says it outright: "A Power BI Pro or Premium Per User (PPU) license alone isn't sufficient—Copilot requires organizational capacity."
Two behaviors deserve flagging. Microsoft documents a general-knowledge fallback: "When a question is related to data in the semantic model, Copilot uses the semantic model to answer the question. Otherwise, it answers from the large language model's (LLM) general knowledge." In other words it may answer from the model rather than from your data. And responses are cached: an identical prompt on an unchanged semantic model returns the cached answer within a rolling 24-hour window, and "The clear chat button doesn't affect this."
Gemini in Google Sheets
Google documents creating tables and formulas, generating "data analysis and insights," building charts, and actions like pivot tables and conditional formatting.
Google is the least specific of the four on access, requiring only "an eligible Google Workspace or Google AI plan" without enumerating which plans qualify. The caveats are candid: "This feature works best with native Google Sheets files," and "The generated chart doesn't link or respond to changes in the original data set that was used to generate it."
Picking the right AI tool matters less than knowing what you are trying to measure. Get a free audit and start from the numbers that count.
The Features That Quietly Got Retired
This is why vendor documentation beats roundups. Two headline capabilities in this category no longer exist in the form they were written about.
Excel's COPILOT() worksheet function — a formula you could type into a cell to classify sentiment or summarize text — is gone. Microsoft's page states: "Starting September 14, 2026, the COPILOT function is no longer available in Microsoft Excel." It never left preview; Microsoft notes "The function was offered as a preview feature," available only through the Frontier and Microsoft 365 Insider programs.
If you have old workbooks using it, Microsoft says "any results that were already calculated remain in the workbook as cached values," but recalculation returns #NAME? because the function "is no longer recognized."
Claude's named "Analysis tool" has also been superseded. Anthropic's own article now carries the note: "The analysis tool is being replaced by more powerful code execution capabilities." Worth knowing for accuracy's sake, since most coverage described it wrongly: the original ran JavaScript, not Python.
The lesson generalizes. In this category, a feature's existence in a well-written guide is not evidence that it still ships. Check the vendor's page.
How to Run an AI Analysis You Can Actually Trust
The workflow that works is not "upload and ask." It is five steps, and the last one is non-negotiable.
- Profile before you question. Ask what is in the file — row count, columns, types, missing values, duplicates, obvious outliers — before you ask anything analytical. You will catch structural problems that would otherwise silently corrupt the answer.
- Name your columns in the prompt. Microsoft's own guidance is that "Naming the columns you want Copilot to analyze produces more accurate results." It is the cheapest accuracy gain available.
- Ask for the code, not just the number. Where the tool runs code, request it. A computed answer with visible code can be checked; a stated answer cannot.
- Give it less, not more. OpenAI is explicit: "Generally, loading in fewer, focused documents will lead to higher accuracy," and warns the approach "can struggle with complex tasks like summarizing very large documents or comparing multiple large files."
- Verify every number that leaves the room. Recompute the headline figures yourself. Anthropic's own best-practice guidance says to "Always review changes before finalizing your work" and "Verify that outputs match your organization's standards and your own judgment."
There is a sixth rule that is really a security rule. Anthropic warns: "Only use Claude for Excel with trusted spreadsheets. Files from external sources can contain hidden instructions that manipulate the add-in into extracting data, modifying records, or performing destructive actions." A spreadsheet from outside your organization is untrusted input, and this risk is barely discussed anywhere.
A verification step you skip is a wrong number you publish. Get a free audit and let us check what your analytics are really telling you.
When AI Is the Right Tool, and When It Is Not
Reach for it when you are exploring. First pass over an unfamiliar dataset, generating candidate hypotheses, writing the analysis code you would otherwise look up, drafting a chart to see whether the shape is interesting, cleaning a messy export. The cost of a wrong answer at this stage is low and the time saved is real.
Be much more careful in four cases, and here the vendors agree with the caution.
Anything audited or client-facing. Anthropic lists Claude for Excel as not recommended for "Final client deliverables without human review" or "Audit-critical calculations without verification."
Regulated or sensitive data. Microsoft advises against using Copilot for "decisions in sensitive areas such as finance, legal, or medical topics," and Google's instruction for its Sheets experiments is blunter: "Don't include personal, confidential, or sensitive information in your prompts."
Anything where the number is the deliverable. If a figure goes straight into a board pack or a filing, the verification cost cancels most of the time saved.
Recurring production reporting. A scheduled pipeline you can test beats a prompt you re-run and re-check every month.
Limitations the Vendors Publish Themselves
Collected in one place, the vendors' own caveats make a stronger case than any critique.
- Microsoft, on Power BI: "Without preparation, Copilot can misinterpret the data and return generic or inaccurate results."
- Google, on Sheets: "Gemini features may suggest inaccurate or inappropriate information."
- OpenAI, on uploads: "Remember that responses may vary depending on the type, number, and size of documents you upload."
- Microsoft, on Copilot in Excel: "As with any AI-generated content, review, edit, and verify anything Copilot creates for you."
Context ceilings are real and mostly undocumented in user-facing terms. OpenAI states that ChatGPT Enterprise "can process up to 110k tokens from uploaded documents in the context window," and that multiple oversized documents "Extract up to 55k tokens, divided evenly among the uploaded documents." Anthropic notes that "Additional token limits may apply based on the length of extracted content." Power BI caps prompts at 10,000 characters.
Governance is where enterprise teams get caught. Anthropic documents that Claude for Excel "does not inherit custom data retention settings your organization might have set" and that "Activity is not included in Enterprise audit logs." It also warns that clearing local chat history "removes only the copy in your browser. It does not delete the messages or attached files that Anthropic keeps."
Training use splits cleanly along consumer and business lines. OpenAI says it "may use content submitted to ChatGPT, DALL·E, and our other services for individuals to improve model performance," and that "Content may include files that are uploaded," while stating it does "not use content submitted by customers to our business offerings such as our API and ChatGPT Enterprise." Microsoft's position on Python in Excel is the clearest single statement any vendor makes here: "Your data is your data. It's not shared with any third parties, nor used for AI training by Copilot."
One more constraint worth planning around: Python in Excel is metered. Microsoft documents a two-tier compute model where premium compute gives "faster Python formula calculation times" up to a monthly allowance, after which calculations fall back to standard speed until the allowance resets.
The Numbers That Do Not Exist
If you read enough comparison articles you will find accuracy percentages, row limits and confidence figures for these tools. None of them come from the vendors, because the vendors do not publish any.
- No accuracy rate. Not one of OpenAI, Anthropic, Microsoft or Google publishes a numeric accuracy or error rate for analyzing a user's own file. They say results can be inaccurate, and they stop there.
- No row limits. OpenAI gives a byte figure hedged as "depending on the size of each row." Anthropic, Microsoft and Google state no row count at all. A row cap in a blog post was invented somewhere.
- No documented failure modes. No vendor describes what happens when an analysis is wrong — no error taxonomy, no confidence signal, no promise that the tool will tell you when it cannot answer.
- No per-plan analysis quotas. OpenAI confirms free-tier access with "separate usage limits" and publishes no numbers. The gap between "included" and "how much" is undocumented everywhere.
- No stated quality ceiling for dataset size. Upload limits tell you what a tool will accept, not what it can analyze well. No vendor bridges that gap.
That last absence is the practical one. The size at which these tools stop being trustworthy is something you have to discover on your own data, because nobody will tell you.
Measuring the right things beats measuring faster. Get a free audit and see where your analytics and search data are pointing you wrong.
Final Thoughts
AI for data analysis is real, useful and narrower than it is sold as. The capability that holds up is code execution over your own file: computed, inspectable, and genuinely faster than doing it by hand. The capability that does not hold up is unverified numeric judgment, and every vendor in this category says so in its own documentation.
Choose on access rather than on feature lists. ChatGPT has the broadest free on-ramp for ad hoc file analysis. Claude is the strongest at producing a finished spreadsheet or model on paid plans. Copilot in Excel fits teams already standardized on Microsoft 365 with the right license. Power BI Copilot needs organizational Fabric or Premium capacity, which is a procurement decision, not a toggle. Gemini in Sheets fits Workspace shops willing to accept vaguer documentation.
And check the vendor page before you trust any specific feature. A function that appeared in a year of roundups stopped working in September 2026, and the roundups have not caught up.
If you want to see the workflow applied end to end on one tool, our walkthrough of trend analysis with ChatGPT covers the prompts, the failure modes and the verification steps in detail.
Want this done for you?
SEO Stuff gets businesses cited by ChatGPT, Gemini, Perplexity and Claude, and ranking in Google. Start with the free audit, a call, or the package.
See where you stand
A free, manual SEO + AI search audit of your site, with a report within 48 hours. No credit card, no sales call.
Get my free auditTalk it through
A free 15-minute call with the founder about whether AI search is a meaningful opportunity for your business.
Free strategy callHave it done
The Done-For-You Package: audit, 10 pages of content, 3 DR50+ placements, dashboard. $999 once, delivered in 21 business days.
See the package

