Analyze Business Data Conversationally — Without Shipping a Wrong Number

Workflow area: Data & Evidence · Tier: Core · Hands-on time: ~35 minutes

The wrong-query-plausible-chart failure mode is the most dangerous artifact in the BA toolkit: the AI misreads a column, silently drops rows with null dates, or averages an average — and hands you a confident, well-formatted chart that survives every glance and no audit. The fix is not to abandon conversational analysis; it is to treat every AI-produced number as a hypothesis until you have inspected how it was computed and reproduced at least one figure by an independent route.

When to use — and when not to

• Use it for exploratory questions on operational extractsticket volumes, cycle times, order values — where you would otherwise spend an hour building pivot tables before you even know which question matters.
• Use it when a stakeholder claim needs a fast evidence check"volumes doubled since March" takes two minutes to test conversationally, and the verification habits make the answer defensible.
• Use it to draft the analysis you will then reproducethe AI finds the interesting cut; you rebuild the headline number by hand before it enters a deliverable.
• Skip it for numbers going into contracts, business cases, or regulatory filings without a full manual rebuild — conversational analysis is a drafting tool, not a system of record.
• Skip it when you cannot see the generated logica tool that shows only the answer, never the formula, query, or code behind it, gives you nothing to inspect and nothing to trust.
• Skip it on data you are not licensed to movea production customer extract does not belong in a consumer chatbot because the pivot was annoying.

The recipe

1. 1.

   Stage a dataset you understand. Export the extract to a working copy and spend five minutes on orientation: row count, one row's meaning, the date column's format and range, and which columns hold codes versus labels. Note the row count — it is your first verification anchor. Remove columns you do not need, especially anything personal.

2. 2.

   Ask the question with the verification contract attached. In Claude or ChatGPT (with the file attached), Excel Copilot, or Gemini in Sheets, use this prompt:

   Copy

   You are assisting a business analyst. Dataset: [ONE LINE — e.g.
   "support tickets Jan-Jun 2026, one row per ticket, 8,412 rows"].
   
   Question: [e.g. "Has average resolution time worsened over the
   last three months, and if so, for which teams?"]
   
   Rules:
   1. Before answering, state: which columns you used, which rows
      you excluded and why, and the exact date range applied.
   2. Show the computation — the formula, code, or step list —
      not just the result.
   3. Report the row count feeding each figure.
   4. If a column is ambiguous (e.g. two date fields), STOP and
      ask which to use. Do not guess.
   5. Flag any data problems you notice: nulls, duplicates,
      format inconsistencies, outliers.
   6. Distinguish "the data shows X" from "X might explain it" —
      label any causal suggestion as HYPOTHESIS.
   

3. 3.

   Inspect the logic before you read the answer. Good output leads with its working:

   Copy

   COLUMNS USED: created_at, resolved_at, team
   EXCLUDED: 214 rows with null resolved_at (open tickets);
   3 rows where resolved_at < created_at (data error, flagged)
   DATE RANGE: 2026-01-01 to 2026-06-30 on created_at
   ROWS IN SCOPE: 8,195 of 8,412
   
   RESULT: Mean resolution 41.2h in Q2 vs 29.8h in Q1 (+38%).
   By team: Tier-2 stable (31h→33h); Tier-1 worsened (28h→46h).
   HYPOTHESIS: Tier-1's mix shifted — hardware tickets rose from
   12% to 31% of their volume; hardware tickets average 3x longer.
   Verify before concluding performance declined.
   

   Read the exclusions and date range as hard questions: are open tickets supposed to be excluded from this measure? Is created_at the right anchor, or does the business count from first response?

4. 4.

   Reproduce one number independently. Take the headline figure and rebuild it by a different route: a hand-built pivot table, a spreadsheet formula on the raw column, or a filtered row count you can eyeball. If your number and the AI's disagree, the discrepancy is the finding — chase it before anything else. Do this for at least one figure per deliverable, always the one with the most consequence.

5. 5.

   Interrogate the pattern before you narrate it. Averages hide mix shifts, seasonality, and instrumentation changes. Ask the follow-up explicitly: "Could this trend be explained by a change in mix, a seasonal pattern, or a change in how the data is recorded? Test each against the data." In the sample above, that question converts "Tier-1 got slower" — a performance accusation — into "Tier-1's workload changed," a completely different requirement.

6. 6.

   Write findings with their audit trail. Every number in the memo carries its scope in-line ("mean resolution, closed tickets only, Jan–Jun, n=8,195") plus which figure was independently reproduced. Anything the AI proposed causally stays labeled HYPOTHESIS until a data owner or SME confirms it.

Make it yours

• Excel Copilot / Gemini in Sheets for spreadsheet-native work: the answer lands next to the data and the generated formulas are inspectable in-cell — apply the same contract prompt, and keep a "verification" tab holding your independent rebuilds.
• Claude or ChatGPT data analysis for bigger extracts: file upload plus generated Python handles hundreds of thousands of rows; always expand and read the code cell — the excluded-rows logic lives there.
• A standing analysis Project per engagement: store the dataset dictionary and the rules prompt in project instructions so every question you ask all month inherits the same verification contract.
• Free path: Gemini in Sheets on a consumer account plus a hand-built pivot for reproduction covers every habit in this playbook — the discipline, not the license, is the tool.

Quality check

• The logic was inspectedyou read the columns used, exclusions, and date range for every reported figure, not just the answer.
• One number reproduced independentlyby a different route, matching within rounding; the memo names which one.
• Row counts reconcilerows in scope plus exclusions equals the total you noted at staging.
• Mix, seasonality, and instrumentation were testednot just mentioned — before any trend was narrated as real.
• Causal language is earnedthe deliverable separates what the data shows from hypotheses awaiting SME confirmation.
• Scope rides with every numberpopulation, filters, date range, and n are stated where the figure appears, so a reviewer can challenge it without a meeting.