How to Use AI for Data Analysis at Work (2026): Clean, Query, Chart and Check

Short answer: write down the business question first, upload a clean export (CSV or XLSX) to an assistant that can run code, ask it to profile and clean the data, then ask for the analysis and a chart. The step most people skip is the last one: check every number against the source file before it goes in a deck. Claude Pro costs $17 a month billed annually and Google AI Pro $19.99 a month (prices read October 30, 2026).
The AI part is the easy part. A chat assistant will happily produce a confident chart from a file with duplicated rows, mixed date formats and a total line counted as data. The workflow below is built around that risk: the model does the tedious work, and you keep the judgment calls and the final check.
The 5-step workflow at a glance
| Step | What you do | What the AI does | What to watch |
|---|---|---|---|
| 1. Frame | Write the decision and the metric | Turns a vague ask into testable questions | Vague prompts get vague charts |
| 2. Upload | Export a clean CSV or XLSX, strip what is not needed | Reads the file and describes its columns | File limits, company data rules |
| 3. Clean and profile | Approve each cleaning rule | Finds blanks, duplicates, outliers, type errors | Silent row drops |
| 4. Analyse and chart | Ask one question at a time | Runs code, builds tables and charts | Wrong grouping or date range |
| 5. Verify | Recompute 2-3 key numbers yourself | Shows the code and filters it used | Numbers that do not match the source |
Dupple data: Toolradar, our software directory, tracks 226 business intelligence tools as of October 30, 2026, and the median cheapest paid plan across the 59 that publish a price is $50 a month (Toolradar business intelligence data). A general assistant with code execution costs well under half that, which is why most teams start there and only move to a BI tool once a question repeats every week.
Step 1: How do I frame the question before using AI?
Start with the decision, not the dataset. "Analyse our sales data" invites a tour of every column; "Which three regions missed Q3 target, and was it fewer deals or smaller deals?" gives the model something it can answer and you something you can check.
Write three things before you open a chat: the decision someone will make, the one metric that drives it, and the time window. Then ask the model to sharpen it.
I need to decide [decision]. My data covers [what, which period]. Turn this into 3-5 specific questions the data can answer, and for each, name the columns you would need. Do not analyse anything yet.
That prompt does two jobs. It forces a plan you can review, and it tells you early when the file is missing a column you need, before you spend an hour on charts.
Step 2: How do I upload company data safely?
Export only what the question needs. Drop names, emails and free-text notes unless the analysis depends on them, and replace customer names with IDs. A smaller file is also easier for the model to read in full.
Each assistant has its own limits, and they matter once your exports get large:
| Assistant | Plan prices (US) | File limits read on the vendor's help pages | Spreadsheet handling |
|---|---|---|---|
| Claude | Free $0; Pro $17/mo billed annually ($200 a year) or $20 monthly; Team Standard $20/seat/mo billed annually or $25 monthly | Up to 20 files per chat, 500MB per file in chat; project files 30MB each | CSV and XLSX; XLSX needs code execution and file creation turned on |
| Gemini | Free $0; Google AI Plus $4.99/mo; Google AI Pro $19.99/mo; Gemini Enterprise Business from $21/seat/mo | Up to 10 files per prompt, 100MB per file (2GB for video) | Charts from uploaded spreadsheets in the web app, editable chart type and labels |
| ChatGPT | Check the current price on OpenAI's site | Check the current limits on OpenAI's site | See our ChatGPT for Excel guide |
Read on Claude pricing, Claude's file upload help, Google AI plans and Gemini's file upload help on October 30, 2026.
Privacy is the bigger question for work data. On Claude, consumer plans (Free, Pro, Max) let you opt out of model training, while Team plans do not train on your content by default. Google's Gemini Apps privacy notice says human reviewers read some chats and asks users not to enter confidential information; work and school accounts can fall under different data terms (Gemini Apps Privacy Hub). If the file holds customer or financial data, use your company's business plan or ask IT first.
Once the file is in, ask the model to describe it before doing anything:
Describe this file: number of rows and columns, each column's type, the date range, and anything that looks like a header, total or note row rather than data.
Step 3: How do I clean and profile the data with AI?
Ask for a data profile, then approve the cleaning rules one by one. A profile lists blanks per column, duplicate rows, values outside a sane range, and columns stored as text that should be numbers or dates.
Profile the data: missing values per column, duplicate rows, outliers, and type problems. Propose a cleaning rule for each issue and wait for my approval before changing anything. Report how many rows each rule would remove.
The "how many rows" line is the safeguard. Models clean aggressively when left alone, and a rule that silently drops 400 rows with a blank region can change your answer. Ask for the cleaned file back as a download so the version you analysed is on record.
Typical fixes worth approving: unify date formats, trim spaces from category names ("West" and "West " count as two regions), remove subtotal rows and convert currency text such as "$1,200" to numbers. Typical fixes to question: filling blanks with averages, and deleting outliers that might be your biggest customer.
Step 4: How do I get the analysis and charts?
Ask one question per prompt and ask the model to show its work. Assistants that run code (Claude with code execution, Gemini's chart feature in the web app) compute on the actual file instead of guessing from a sample, which is the difference between analysis and a plausible paragraph.
Using the cleaned data, answer question 1. Show the code or formula you used, the filters applied and the row count behind each number. Then make one chart that answers it, with a title that states the finding.
Three prompts cover most work requests:
- Trend: "Plot monthly [metric] for the last 12 months and flag months more than 20% above or below the average."
- Breakdown: "Split [metric] by [segment] and rank segments by change versus the prior period."
- Driver: "For segments that missed target, was it volume or average value? Show both side by side."
Ask for the chart title to state the finding ("West missed Q3 target on deal count, not deal size"). A chart titled "Sales by region" pushes the interpretation back onto your reader.
Step 5: How do I check the AI's numbers?
Pick the two or three numbers that will appear in your slide or email and recompute them yourself, in the spreadsheet, with a pivot table or a SUM over a filter. If one does not match, ask the model which filter it used; the gap is almost always a date boundary, a duplicate or a total row.
List every number in your answer with the exact filter and row count behind it, so I can reproduce each in a spreadsheet.
Never forward a number you have not reproduced once. That rule costs five minutes and protects the credibility of the whole analysis. It also catches the rarer case where the model wrote a figure in prose that is not in its own output table.
Should I use a chat assistant, Excel's built-in AI or a BI tool?
Use the route that matches how often the question comes back.
- One-off questions on an export: a chat assistant with code execution. Fastest to start, nothing to set up.
- Data that lives in a workbook you maintain: the AI built into your spreadsheet. Our guide to Copilot in Excel covers which Microsoft 365 plans include it, and the best AI for spreadsheets guide compares the Google Sheets and Excel options, with add-ons in the best AI spreadsheet tools list.
- The same dashboard every week for several people: a BI tool connected to the source. See the best AI business intelligence tools.
For a wider comparison of the assistants themselves, our best AI for data analysis and best AI data analysis tools guides rank them, and the best AI data analyst tools list covers assistants built for analysts.
How much does AI data analysis cost for a team?
For one analyst, Claude Pro on annual billing costs $200 a year, versus $240 a year paying $20 monthly, by our arithmetic. Google AI Pro at $19.99 a month comes to $239.88 a year. Before paying, see whether a trial covers you: our guide on getting Claude Pro for free lists the current routes.
For a 5-person team that needs business data terms, Claude Team Standard costs $1,200 a year on annual billing (5 seats x $20 x 12) and $1,500 a year on monthly billing, by our arithmetic. Gemini Enterprise Business starts at $21 a seat a month, or $1,260 a year for 5 seats on that starting price. Five seats of a BI tool at the $50 median cheapest plan would cost $3,000 a year, though BI pricing often counts viewers and creators differently, so check the vendor's own unit before comparing.
Common mistakes to avoid
- Uploading the raw export. Totals, notes and blank rows end up treated as data.
- Asking five questions in one prompt. The model answers all of them shallowly and you cannot trace any number.
- Letting the model clean without approval. Rows disappear and nobody knows how many.
- Trusting a chart without its filter. A chart that hides its date range or segment filter invites your reader to the wrong conclusion.
- Pasting confidential data into a personal account. Use the business plan your company pays for.
FAQ
Can AI analyse an Excel file?
Yes. Claude reads XLSX and CSV files once code execution and file creation are turned on, with up to 20 files per chat. Gemini accepts spreadsheets up to 100MB per file and can build charts from them in the web app. If your data lives in Excel already, the AI built into the spreadsheet is the other route.
Which AI is best for data analysis at work?
Pick an assistant that runs code on your file rather than one that only reads it, and pick the one your company already licenses, since data terms matter more than small feature gaps. Our best AI for data analysis guide compares the options.
Is it safe to upload company data to AI?
It depends on the plan. Claude Team does not train on your content by default, and Claude's consumer plans let you opt out. Google says human reviewers read some Gemini Apps chats and asks users not to share confidential information. Check your company's AI policy before uploading customer or financial data.
How do I stop AI from making up numbers?
Ask it to run code on the file, show the filter and row count behind every number, and recompute the two or three numbers that matter yourself. A number you cannot reproduce in a spreadsheet does not go in the report.
How much does an AI assistant for data analysis cost?
Claude Pro costs $17 a month billed annually or $20 billed monthly, and Google AI Pro costs $19.99 a month. Team plans start at $20 a seat a month on Claude Team Standard (billed annually) and $21 on Gemini Enterprise Business. Prices were read October 30, 2026.
Do I need to know SQL or Python?
No. The assistant writes and runs the code. Reading it helps, though: asking the model to explain each step in plain English is the quickest way to spot a wrong filter.
How we checked
Plan prices and file limits were read on Anthropic's and Google's own pricing and help pages on October 30, 2026; ChatGPT's limits are left to OpenAI's site. Team totals are our arithmetic on printed prices. We do not run a test lab, so this page describes a workflow and the vendors' stated limits, not timed results. See how we rate tools; we re-check prices about every 60 days.
Cite this: Dupple, "How to Use AI for Data Analysis at Work (2026): Clean, Query, Chart and Check", November 2026.
Want to get better at AI for everyday work? Start the Dupple X trial for practical AI courses and tool guides.
Some offers on this page may be paid placements or contain affiliate links.