Frame the question before the prompt

AI makes it easy to ask broad questions because it can return polished answers to almost anything. The danger is that a fluent summary can conceal an undefined goal. Before uploading a table or writing a prompt, state the decision the analysis will inform.

“Analyze customer churn” is a topic. “Decide whether to fund a retention campaign for customers in their first 90 days” is a decision. The second statement suggests a population, a time window, a possible comparison, and an action. It also makes clear why the analysis matters.

Use this sequence:

1. Decision: What choice, action, or understanding should this analysis support?
2. Population: Which people, transactions, products, or events count?
3. Measure: What quantity represents the outcome?
4. Comparison: Compared with what baseline, group, or period?
5. Time window: Which dates belong in the analysis?
6. Evidence standard: What result would be strong enough to change the decision?

If the requester cannot name a decision, produce an exploratory brief instead of pretending the work is conclusive. Exploration is useful, but it should be labeled as exploration.

Business words often hide multiple formulas. “Growth” might mean total revenue growth, unit growth, customer growth, or growth after refunds. “Best customer” might mean highest lifetime revenue, highest margin, most recent activity, or strongest retention. “Average order” might include cancelled orders unless you explicitly exclude them.

Write operational definitions that another person could reproduce. For example:

Monthly net revenue equals completed order value minus refunds, grouped by the order completion month, excluding test accounts.

That sentence identifies a status rule, subtraction, date field, grouping level, and exclusion. It is far safer than asking AI, “Show monthly revenue.”

Ask the model to list ambiguous terms before calculating:

My decision is whether to expand the weekend support schedule.
The table contains ticket_id, created_at, resolved_at, channel,
priority, and customer_segment.

Before analyzing, identify terms that need definitions, including
"weekend demand," "response time," and "resolved." Propose options,
but do not select definitions without my approval.

This prompt gives AI a useful role: exposing choices rather than silently making them.

The unit of analysis is what one row represents after preparation. It could be one order, customer, day, support ticket, or store-month. Mixing units creates plausible but incorrect results.

Suppose an orders table has one row per product line. A single order containing three products appears three times. Counting rows as orders inflates the order count. Joining that table to one-row-per-customer data can also repeat customer attributes. Before calculating, complete this sentence: “One row represents one ______.”

Then identify the unique key. If no key exists, decide whether a combination such as order_id + product_id is unique. Ask AI to test the claim, but verify the result in the table.

A compact question contract prevents the analysis from drifting:

• Decision: Allocate the next campaign budget.
• Question: Which product category had the largest year-over-year change in net revenue?
• Population: Completed, non-test orders in active regions.
• Measure: Sales minus refunds, in USD.
• Comparison: January–June this year versus the same months last year.
• Breakdown: Product category; show region as a diagnostic.
• Deliverable: Calculation table, one chart, three evidence-based observations.
• Limits: Do not infer causes; flag missing months and category changes.

This contract is also a review artifact. A stakeholder can correct the question before you spend time cleaning or charting.

Tables can usually support statements about what changed. They do not automatically reveal why it changed. If conversion fell when a price changed, the timing is evidence of association, not proof that price caused the decline. Promotions, seasonality, traffic mix, product availability, and tracking changes may also matter.

Label the analytical task:

• Descriptive: What happened?
• Diagnostic: Where and when did it happen?
• Predictive: What is likely next?
• Causal: What happened because of a specific intervention?

AI can help structure all four, but causal claims need an appropriate design, such as a randomized test or a carefully justified comparison. A chart alone does not create causality.

Analysis can expand forever. Before starting, define what “enough” means. You might stop when the primary measure is verified, the key comparison is stable under reasonable definitions, major data-quality concerns are documented, and the result answers the named decision.

Also define what would block a conclusion: missing historical periods, an unknown denominator, inconsistent status values, or too few observations. “Insufficient evidence” is a valid result. Instructing AI to say what is missing produces better work than demanding an insight from every table.