# What should happen between a business question and an AI-generated number?

By Sean Everett · October 8, 2026

Make the interpretation, metric definition, permitted data and calculation inspectable, then check the numerical result and its explanation against the intended question.

Topics: Analytical AI · Business intelligence · Calculation controls

## Translate the question into explicit analytical choices

Before calculating, identify the measure, population, period, grouping and comparison the user intends. A request about average customer activity might mean activity per account, per active account or per individual. If plausible interpretations would change the decision, ask a focused question or show the chosen definition.

Snowflake’s semantic-view documentation separates business entities, relationships, dimensions and metrics from their physical storage. That provides a useful design pattern: map familiar terms to maintained definitions instead of inventing a new formula for each conversation. Preserve the selected definition and its version with the result.

Check joins and aggregation at the required level. Combining an account table with many activity rows can multiply values that belong only once per account. Make these choices reviewable alongside units, exclusions and the meaning of missing values.

## Execute the calculation inside a bounded data workflow

Use a query or analytical engine for the calculation, with access controls enforced by the connected system. Restrict the workflow to the data and operations needed for the question. Keep retrieval of definitions, selection of records and execution of arithmetic distinguishable in the product’s diagnostics.

The Spider 2.0 research benchmark evaluates enterprise text-to-SQL workflows that can require database metadata, dialect documentation and multiple operations. It provides a reason to test realistic analytical tasks rather than isolated query examples. Its published benchmark results are specific to the configurations evaluated.

A hypothetical completion rate needs an eligible population, a completed count and a period. Return those ingredients with the calculated rate. Treat an empty eligible population as an unresolved or undefined result under the metric’s rules, rather than silently manufacturing a zero.

## Verify the result and what the explanation claims

Snowflake’s Cortex Analyst evaluation documentation describes executing generated queries and comparing their results with verified queries. Adopt the underlying check for your own system: use independently reviewed question-and-calculation pairs, and keep evaluation answers from serving as runtime examples for those same tests.

Test cases that expose wrong interpretations even when a query runs: different denominators, repeated records, boundary dates, missing categories and restricted access. Check the reference calculation too. Matching an incorrect reference would preserve the same analytical mistake rather than establish a valid answer.

Review the explanation separately. It should carry the population, period, units, data limitations and assumptions needed to use the number. A computed association alone cannot explain its cause, and a sample estimate needs uncertainty appropriate to its design. Escalate consequential statistical interpretation to a qualified reviewer and preserve unanswered questions.

## Related reading

- [Snowflake: Semantic views and business definitions](https://docs.snowflake.com/en/user-guide/views-semantic/overview)
- [Spider 2.0: Enterprise analytical workflows](https://arxiv.org/abs/2411.07763)
- [Snowflake: Executed-query evaluation](https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-analyst-evaluations)
- [Testing an AI feature’s actual job](https://evergences.com/insights/ai-behavior-evaluation/)
- [Related owned product: Jetpack Intel — custom internal AI workflows](https://jetpackintel.com/)

Source: https://evergences.com/insights/ai-analytical-calculation-boundaries/