Decision
What decision should this analysis improve?
AI-Native Analytics · Module 09
AI-Assisted SQL & Analysis.
Use AI to compress the mechanical parts of analysis without outsourcing the question, metric definition or judgment.
Core idea
The quality of AI analysis depends less on a clever prompt than on the quality of the analytical context you provide.
01
Context
The AI wrote 90 lines of SQL in seconds.
“The query runs and the numbers look reasonable. Can we send the conclusion to leadership?”
Fast syntax generation is useful, but an AI system does not automatically know your metric contracts, grain, exclusions, late-arriving data rules or business decision.
A strong workflow gives AI a structured analytical brief first, then treats generated SQL as a draft that must earn trust.
02
Context bundle
Give the model the same context a good analyst would ask for.
Definitions
Provide metric contracts, exclusions and time logic.
Data model
State table grain, keys, relationships and known caveats.
Output contract
Specify the required grain, comparisons and validation checks.
Rule
Do not ask AI to “analyze the data” when you have not defined what a correct answer would look like.
03
Challenge
Turn a weak AI request into an analytical specification.
Exercise · 15–25 minutes
Rewrite the prompt.
- What decision is the stakeholder making?
- What metric definitions must be provided?
- What table grain and join risks should be stated?
- What output grain should the answer use?
- Which control queries should the AI provide?
- What conclusions should the AI explicitly avoid making?
04
AI assist
Ask for reasoning structure before code.
Strong analysis request
Business question:
Explain the week-over-week decline in completed orders.
Decision:
Determine which area should be investigated first; do not recommend spend changes yet.
Metric contracts:
- completed_order = distinct eligible order_id with final status completed
- reporting week = created_at week
- cancellation_rate = cancelled eligible orders / placed eligible orders
Data model:
orders = one row per order
customers = one row per customer
status_events = many rows per order
Before SQL:
1. Restate the question and required output grain.
2. Identify ambiguities and join risks.
3. Propose the analysis plan.
4. Write SQL in testable layers.
5. Provide control queries.
6. Separate observed findings from causal claims.05
Validation
The AI-assisted analysis checklist.
01
Did the prompt include the decision, not only the metric?
02
Were metric definitions and time attribution explicit?
03
Did the model state the expected grain before joining?
04
Did it identify unknowns rather than invent them?
05
Can each major number be reconciled to a simpler control?
06
Does the final language distinguish observation from cause?
06
Worked takeaway
AI should reduce typing, not analytical ownership.
Frame the decision and write metric definitions.
Provide schema context and request an explicit plan.
Validate SQL, totals, edge cases and conclusions.
Keep only claims supported by checked evidence.
Analyst habit
Treat AI-generated analysis like work from a very fast junior analyst: useful, inspectable, and never exempt from review.