Teste rápido
Guess first, no penalty. A table shows Brazil: 12 orders, $1,320. Canada: 240 orders, $26,400. India: 310 orders, $21,700. The AI's summary says "Brazil had the highest revenue per order." Is that right?
AI summaries of a table read well and sound sure. Sentences that sound sure are not evidence. Some claims are wrong arithmetic. Others are true numbers with too much weight on them, like "strong demand" from 12 orders.
So check the summary the way you check a query: every sentence against the numbers it rests on. It's faster if the AI shows its working. Ask it to put the numbers behind each sentence right after it, and to flag any claim that rests on a small group.
A worked example
A prompt (the message you send to the AI) that makes it show its working:
Summarize this table for a sales manager in 3 sentences.
After each sentence, list the exact numbers it rests on.
Flag any claim based on fewer than 50 orders.
Don't describe trends: the table covers one quarter only.Now check a summary of the table above:
Encontre os erros
Tap the two sentences the table doesn't support.
AI summary
0 de 2 encontrados
Preencha as lacunas
Fill the gaps.
Prompt
Summarize this table in 3 sentences. After each sentence, list . Flag any claim based on .
Try it
Seus passos, cerca de 1 minutos
- 1.Check each sentence of one real AI summary against its table. Add the summary check to your checklist.1 min
Your data-check checklist is now complete. Run it before any AI-assisted number leaves your desk:
1. Explain first: tables, filters, joins, what one row means.
2. Ask with the rules: tables, definitions, dates, exclusions.
Then read the assumptions list.
3. After each join: COUNT(*) vs COUNT(DISTINCT key).
4. Date edges: >= first day and < day after the last.
MIN and MAX show both edges. Name the time zone.
5. Summaries: each sentence against its numbers. Flag small groups.