TL;DR
Paste the table plus the decision it informs into the prompt below. Ask the model for a 60-80 word paragraph naming the single dominant pattern, three action bullets, one “watch out” line, and one data-quality caveat. Put the digest above the table. If the table lives in a file, upload the .csv or .xlsx to ChatGPT or Claude and let it run Python instead of pasting cells. Always re-check any total with a calculator: models confidently restate sums that do not add up.
The task
Your report has a dense table — say 12 columns and 30 rows of numbers — that the reader will glance at and skip. You want a short paragraph above it stating what the table actually shows, plus two or three bullets telling the reader what to do next. The reader is busy. If the takeaway is not on the page, they will not derive it themselves.
When AI helps — and when it does not
AI is genuinely good at finding the dominant pattern in a table (the row or column that drives most of the variance) and stating it in plain language. It is poor at knowing what matters to your reader, which depends on the decision the table is meant to inform. Feed the model both the data and the decision, or you get a generic summary that restates the obvious.
What to feed the AI
- The table data — a markdown table or pasted spreadsheet for small tables; a
.csv/.xlsxupload for anything above ~50 rows - Reader role and the decision they need to make (e.g. “PM deciding whether to ship feature X”)
- What the columns mean: units, definitions, baselines
- What “good” and “bad” look like in your business context
- Caveats: known anomalies, missing weeks, segments that should be ignored
- Length target: one paragraph, three bullets, or longer if the table is critical
Copy-ready prompt
Explain this table to a busy reader.
Reader role and decision: [role, decision]
Column definitions and units: [list]
"Good" vs "bad" in our context: [one line]
Caveats / anomalies: [list]
Length target: 1 paragraph + 3 bullets
Table:
"""
[paste]
"""
Return:
1. A 60-80 word paragraph: what the table shows, in business terms, with the single dominant pattern named.
2. Three bullets answering "so what do I do with this": each is an action, not a restatement.
3. One "watch out" line: the most likely wrong conclusion someone might draw.
4. A one-sentence caveat naming any data-quality issue.
Do not re-list cell values. Do not write "as you can see" — the reader cannot, which is why I'm asking.For technical readers, run a follow-up: “Now write a version for an analyst — surface the statistical caveats and the columns that need normalisation.”
Which tool for which table
For a table under ~50 rows, paste it as markdown into any chat — the model reads it directly and no file handling is involved. For larger files, upload the spreadsheet so the model runs Python over the real numbers (a sandbox commonly called Advanced Data Analysis or, in Claude, a code Artifact) instead of eyeballing pasted text. Verified limits and tiers as of June 2026:
| Tool (June 2026) | Best for | File upload limit | Code execution | Entry price |
|---|---|---|---|---|
| ChatGPT (GPT-5.5) | Pasted tables + heavy file analysis | ~50 MB per file; ~80 files / 3 hrs on Plus | Yes — Python sandbox (Advanced Data Analysis) on paid tiers | Plus $20/mo |
| Claude (Sonnet 4.6 / Opus 4.7) | Narrative digests, large 1M-token context | 30 MB per file, up to 20 files per chat | Yes — Python rendered as a live Artifact | Pro $20/mo |
| Gemini 3.1 Pro in Sheets | Tables already living in Google Sheets | Reads the workbook directly via Workspace | Yes — in-cell =AI() + sidebar analysis | Google AI Pro $19.99/mo |
Notes: ChatGPT’s Free tier (with GPT-5.5) can read a pasted table but the Python sandbox is a paid feature. Claude Sonnet 4.6 and Opus 4.7 both carry a 1M-token context window, so a very wide table fits without chunking. Gemini’s in-Sheets analysis rolled out to Google AI Pro and Ultra subscribers from late April 2026.
Recommended output structure
Paragraph above the table, a “Takeaways” bullet list below it, an optional “watch out” callout, and a “data caveats” footnote. Put the AI-written digest above the table — many readers stop at the digest, which is fine.
How to check the output is usable
- The paragraph names one dominant pattern, not five competing observations
- Each bullet is an action-verb sentence (“Cut SKU group C next quarter”), not a restatement
- The “watch out” line is a real misreading risk, not boilerplate (“correlation is not causation”)
- The caveat names the actual data issue, not a generic disclaimer
- A reader who only reads the digest still gets the correct decision direction
- Every total or percentage in the digest survives a manual re-check
Common mistakes
- Re-listing cell values in prose (“Q1 was 12.3, Q2 was 13.1…”) — the single most common AI failure on tables
- No “so what” bullets: a summary without action is just a longer table
- Letting AI promote noise to a trend — ask for a confidence note
- Skipping caveats: your reader will be quoted on the number, missing data and all
- Burying the digest under the table: most readers do not scroll back up
- Trusting a sum the model typed instead of one the Python sandbox computed
FAQ
- What if the table has 50+ columns? Ask the model to identify the five columns driving the decision and write the digest only on those. Push the rest into an appendix.
- Can AI catch arithmetic errors? Sometimes. When you paste a table as text, the model often re-types totals from memory and gets them wrong. Upload the file so it runs Python on the real numbers, then still spot-check one total by hand.
- Should I paste the table or upload the file? Paste anything under ~50 rows; upload
.csv/.xlsxabove that. ChatGPT accepts files up to ~50 MB; Claude up to 30 MB per file (20 files per chat) as of June 2026. - What about charts instead of tables? See chart takeaway: same logic, different prompt.
For the official capabilities and current limits, see OpenAI’s data-analysis help page and Anthropic’s Claude-for-Excel guide.
Related
- Chart takeaway: same digest pattern for charts instead of tables
- Spreadsheet error diagnosis prompts: when the table itself looks wrong
- Chart & table explanation: combined explainer workflow
- AI table explainer tutorial: full workflow including prep
- ChatGPT spreadsheet help: tool walk-through inside ChatGPT
Tags: #Data analysis #Workflow #Research