Working with data and tables
~ min read
30-second summary
- Three use cases where AI saves hours: cleaning inconsistent data, Excel or Sheets formulas you can’t remember, small analyses on extracts.
- The prompt shape changes: you pass a data extract (or a handful of sample rows) plus a precise question, not a page of context.
- Privacy is the main constraint: a CSV with customers, amounts, addresses is a company document. Anonymize or extract only the columns you need before uploading to a public chat.
- AI for ad-hoc, BI for recurring: if the same analysis comes back every month, the right place is a dedicated tool (Power BI, Looker, your CRM’s dashboard). AI earns its keep when the calculation is one-off.
- On numbers, always verify a sample: AI can miss a sum or a rounding and hand it back with the same confidence as a correct answer.
You have an Excel table with the wrong columns, customer names written fifteen different ways, dates half in European format and half American, and a formula you can’t remember. Without AI you’re in for an hour of copy-paste by hand. With AI, ten minutes.
The time saved is real, but there’s one catch you can’t skip: a work spreadsheet isn’t your journal. It has customers, suppliers, contracts, revenue. Before pasting it into a public chat, take thirty seconds to decide what leaves the company and what doesn’t.
This lesson is the practical toolkit: three concrete use cases, a privacy section that’s the most important piece, and a rule for when to stop asking AI and open the right tool.
Three concrete use cases
Section titled “Three concrete use cases”The three below cover 80% of what a non-developer knowledge worker does with a spreadsheet. They’re everyday moves, none of them requires code.
1. Cleaning inconsistent data
Section titled “1. Cleaning inconsistent data”The most common case. A “Customer” column with entries like “Smith John”, “J. Smith”, “smith j”, “John Smith Ltd”. Dates written as “12/03/2026”, “2026-03-12”, “March 12”. Duplicates with different capitalization. By hand it’s an afternoon; with AI it’s five minutes.
The prompt works if you give it an extract (ten to twenty representative rows) plus an explicit normalization rule:
I’m pasting ten rows from the “Customer” column of our CRM. Normalize them to “First Last” format (e.g., “John Smith”). If there’s a business suffix like “Ltd” or “Inc”, put it after a comma. Give me back the clean table plus one line with any open doubts.
The two pieces that make the difference: the rule (“First Last”, “suffix after the comma”) and the explicit ask for doubts. Without that, AI normalizes even where it should ask you how to read an ambiguous case (“J. Smith” is John or James?).
The same logic applies to dates, with one note on locale: if you work in European format, state “DD/MM/YYYY format” or “YYYY-MM-DD format” explicitly. AI tends to default to the US format (MM/DD/YYYY), and if your data is mixed it can read “03/04/2026” as “April” instead of “March”. One line of constraint fixes the problem.
What if the rows are five hundred, not ten? Two paths. The
first, practical one, is the sample, rule, extend pattern:
you pass ten to twenty representative rows, work out the
cleanup rule with AI, then apply it to the rest in Excel with
a formula (find and replace, PROPER, MID) or a macro.
The second is uploading the full CSV as an attachment: Claude and ChatGPT handle long PDFs and CSVs up to a few thousand rows well, as long as the file is within the platform’s limits. Between the two, the first is usually safer: it keeps you in control of how the rule gets applied, and it doesn’t dump five hundred rows of company data into a single AI call.
2. Excel or Sheets formulas you can’t remember
Section titled “2. Excel or Sheets formulas you can’t remember”“How do I count how many times a value appears in a column?” “How do I sum only the January rows?” “How do I compare two columns and find the differences?” AI gives you the exact formula and a one-line explanation.
What to pass here is which tool you use and which language it’s installed in. Both change the syntax:
- Excel in most European languages uses local function names
(e.g., Italian
SOMMA.SE, FrenchSOMME.SI). - Excel in English (and Google Sheets, which is always English
internally) uses
SUMIF,COUNTIF,VLOOKUP. - The parameter separator is
;in European locales,,in English ones. - The decimal separator is
,in most of Europe,.in English.
Say it to AI in the prompt: “English Excel” or “Google Sheets” or “Italian Excel”. If you don’t specify, sometimes you get a formula you can’t paste as-is and have to adapt by hand.
The real value is the explanation: always ask “explain what each part of the formula does” the first time. Next time you’ll write it yourself or adapt it to a different case. Without the explanation, you’ve solved today but learned nothing.
If you paste the formula and Excel answers with #NAME?,
#VALUE!, or #REF!, don’t start from scratch. Copy the
error as-is into the chat and ask “this is the error I’m
getting, what’s wrong?”. Nine times out of ten it’s one of
three things: wrong separator (, where it wants ;),
function in the wrong language (SUMIF in a non-English
Excel), or a cell reference that doesn’t exist on your sheet.
Pasting the error back into the chat is faster than guessing
the fix.
3. Small analyses on extracts
Section titled “3. Small analyses on extracts”“On these 50 orders, what’s the average value per customer? How many customers spend above the average? Is there a clear seasonality?” AI reads, computes, and answers.
Two precautions. First, pass the extract in a form AI reads well: a pasted CSV, or an attachment (everything from Photos, images, and files applies). Screenshotted tables work, but a text CSV is more reliable because AI reads it character by character, with no risk of mistaking one digit for another in a blurry photo.
Second, verify a sample. Ask AI for the calculation and then re-check a few rows by hand or with an Excel formula. A handy rule: 5% of the rows, minimum three. If the dataset is 50 rows you re-check three; if it’s 500, you re-check twenty-five; if it’s five thousand, the use case is probably wrong, and the section AI for ad-hoc, BI for recurring covers when to switch to a dedicated tool.
Pick rows that cover three profiles: one in the middle, one at the extremes (highest or lowest value), one with empty or oddly formatted fields. Not because AI is often wrong, but because when it is, it’s wrong with the same confidence as when it’s right, and on numbers a digit shifted by one position is an error you can’t catch at a glance. The pattern is the same as in When to trust it and when not to.
Privacy: the main constraint
Section titled “Privacy: the main constraint”A CSV with customer names, amounts, addresses, revenue is a company document. Even if it sits on your laptop and isn’t inside a CRM, even if you’re handling it on your own, it’s company data. Before uploading it to a public chat (ChatGPT Free/Plus, Claude Free/Pro, Gemini standard), three moves apply, in this order.
- Ask whether a policy exists. In many mid-size and large companies, IT or legal has already decided what can be uploaded and where. If there’s a guideline, follow it and save yourself the rest of the thinking. If there isn’t (or if you’re a freelancer and the policy is you), the handy test is this: “would I be fine with this data showing up on a newspaper tomorrow?”. If the answer is no, the two moves below aren’t optional.
- Extract only the columns you need. If you want to know how much customers spent on average, you don’t need their address, their tax ID, or their phone number. Delete those columns from the extract before uploading. Less data uploaded, less risk.
- Anonymize the names. Replace “John Smith” with “Customer 1”, “Beta Inc” with “Customer 2”, and keep the mapping table in a separate sheet. When AI returns the analysis, you map it back to the real names. It’s the same move as in Photos, images, and files, applied to tabular data.
These three moves don’t cover everything. Contracts with confidentiality clauses, health data, special category data (GDPR art. 9) are a notch higher: anonymization isn’t enough, you need a dedicated company channel or no public AI at all. A later lesson in this module, Company data and privacy, covers that level.
AI for ad-hoc, BI for recurring
Section titled “AI for ad-hoc, BI for recurring”AI is a copilot for the one-off case: you have a sheet tonight, you need to clean it up, you paste and go. If instead the same analysis comes back every month (a dashboard, a quarterly report, a weekly KPI for the team), AI is the wrong tool.
Recurring analysis is what Business Intelligence tools are for (Power BI, Looker, Tableau), or the native dashboard inside your CRM or ERP. There’s a structural reason for this: once it’s set up, the dashboard refreshes itself, it’s reproducible, and it’s shareable with the team. With AI you start from scratch every time, you risk a slightly different answer than last month, and no one can trace how you got a number.
Practical rule: if you’ll ask the same question again within a month, ask a tool other than AI. If it’s a question that’s born today and dies tonight, AI is fine.
A concrete example
Section titled “A concrete example”You have a five-row extract from a Customer column in the CRM, written inconsistently. You want them normalized to First Last format to import into a clean Excel sheet.
What AI got right: uppercase/lowercase normalized, First Last order respected, business suffix separated out. Better still: it didn’t decide on its own that “J. Smith” was “John Smith”, it flagged the doubt. The “plus one line with open doubts” constraint in the prompt is what triggered that behavior.
Without that constraint, AI would have probably written “John Smith” everywhere, merging rows that might have been different customers, and you wouldn’t have noticed.
What NOT to do
Section titled “What NOT to do”Don’t trust numbers without verification. On small analyses, always re-check two or three rows by hand. AI can miss an average, a rounding, or miscount rows that meet a condition. The verification pattern is the one from When to trust it and when not to, applied to numbers.
Don’t upload entire datasets of sensitive data. If you have a CSV of ten thousand rows with customers and amounts, you don’t need to upload all of it to ask for a formula. Extract ten representative, anonymized rows, and pass those. For the same reasons as in Privacy: the main constraint.
Don’t use AI as an analyst for important business decisions. Financial planning, investment appraisal, customer analysis for a strategic decision: those need a real analyst (or you, with AI as a targeted helper), not AI as the author of the conclusion. The perimeter is the one from Things NOT to do: where not to delegate the decision, regardless of how good the answer might be.
What’s next
Section titled “What’s next”The next lesson, Research for work, shifts the axis: from your own data (a sheet in your hand) to external sources (articles, reports, websites) when you need to reach a synthesis for a professional deliverable. It’s the place where AI speeds you up the most and also hallucinates the most, and method is what separates the two.