# Session 2 Handout: Ask and Verify

<p class="meta">FAST-11 AI-Assisted Analytics and Business Intelligence · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

Plain-language questions turn a spreadsheet or dashboard into a conversation. The answers come back fast and confident, and sometimes wrong: the wrong column, a hidden filter, an average of averages. The habit this session builds is simple: **every AI answer is a claim until you check it.**

## Key concepts

### 1. Ask specific questions

Name the **measure**, the **time period**, and any **filter** or grouping.

| Vague | Specific |
| --- | --- |
| "Which machine is worst?" | "Total scrap cost by machine, July 1 to September 30, 2026" |
| "Why are no-shows high?" | "No-show rate by weekday, last 26 weeks" |

Ask the tool to **show the calculation** (columns, filters, aggregation). Tools: Copilot in Excel, Copilot for Power BI (Power BI Q&A retires in February 2027), Tableau AI and Pulse, ChatGPT data analysis, Gemini in Sheets.

### 2. Verify with pivots and formulas

- Rebuild the answer with a **pivot table**, **SUMIFS**, **COUNTIFS**, or **AVERAGEIFS**.
- Compare **row counts** and **date ranges** first.
- Record match or mismatch on W2; **resolve every mismatch** before reporting.

### 3. Where AI analysis goes wrong

| Failure | Example | Check |
| --- | --- | --- |
| Wrong column or measure | Uses quantity instead of cost | Read the calculation |
| Hidden filter or dropped rows | Excludes 24 blank-reason rows | Compare row counts |
| Average of averages | Averages monthly rates | Weight by volume |
| Correlation as cause | "Rose because of the new coolant" | Is the cause in the data? |
| Confident wording | "Most scrap is from VF-4" (41%) | Compare words with numbers |

## Worked example (fictional)

Copilot reported Q3 scrap cost of **$14,920**. SUMIFS gave **$16,102**. Row counts: 388 vs 412. The 24 missing rows all had a blank reason, so the AI's view had filtered them out. The correct total is $16,102; the cause goes on W2.

## Checklist

- [ ] Ten questions with measure, period, and filter.
- [ ] Calculation shown for each.
- [ ] Three or more verified independently (one share, one change over time).
- [ ] Row counts and dates compared.
- [ ] Every mismatch explained.

## Do and don't

| Do | Don't |
| --- | --- |
| Verify with a different method | Ask the same question again as a "check" |
| Explain mismatches | Average the two numbers |
| Report patterns | Report causes the data can't show |