Provide the minimum explanation needed for the decision

A result such as East declined eight percent is unusable without revenue definition, period, refund treatment, and scope. Raw SQL exposes implementation and still does not explain business meaning.

Use a middle layer: business-readable plan nodes linked to controlled technical evidence. Increase detail with risk and role rather than showing the same explanation everywhere.

Represent a query as stable plan nodes

Useful nodes include ResolveMetric, ResolveScope, SelectSource, JoinEntity, Filter, Aggregate, Compare, Validate, and Present. Each retains input and output grain, state, description, and evidence ID.

These nodes remain stable across database dialect changes while semantic changes create a new explicit version. They are also testable, unlike free-form explanatory prose.

Explain interpretation first

Show the chosen metric, dimension, date role, comparison basis, currency, and defaults. State how month, customer, and complete data are defined.

Identify which prior message supplied an inherited constraint. Ambiguity still requires clarification; an explanation cannot legitimize an unconfirmed guess.

Explain joins through grain and cardinality

A business user can understand that order lines link to products and customers while header value is counted once. They do not need the JOIN syntax.

Store expected cardinality, unmatched share, duplicate protection, and history choice. Successful execution is not evidence that the join preserved meaning.

Expose filters, exclusions, and cutoff

List status, test-account, cancellation, refund, organization scope, watermark, and missing partitions. Exclusions often explain more than formulas.

For multiple sources, use the common complete cutoff. Mark fallback batches, caching, and samples alongside the result rather than in a distant footnote.

Show formula meaning and units

Gross margin rate should display gross profit over net revenue. Year comparison needs calendar, comparable population, and zero-denominator rules. Weighted averages and shares need their aggregation grain.

Layer the display: readable definition first, governed expression and lineage for authorized reviewers. Never expose credentials, hosts, or restricted literal values.

Mark approximation and inference

Top-N truncation, sampling, approximate distinct counts, timeout fallback, and cache change result properties. Label exactness, coverage, error where known, trigger, and re-run option.

Separate facts, association, and generated recommendation. Fluent text must not turn simultaneous movement into causality.

Test the explanation itself

For golden questions, compare displayed nodes with actual execution. Shared IDs should prove that the refund filter described to the user was applied in the query.

Use layered accessible display, keyboard navigation, and text risk cues. Exports should carry the same explanation version.

Reject misleading explanation patterns

Fixed boilerplate, stale plan versions, exposed chain-of-thought, raw schema leaks, hidden defaults, and unmarked approximation are failure modes. Internal model reasoning is not audit evidence.

Acceptance requires every number to map to a node, complete scope and watermark, non-leaking authorization, stable explanations after physical rewrites, and business reviewers who can detect a wrong interpretation.

AskTable.ai boundary

Pilot high-risk operational or finance questions, define minimum plan fields, link them to execution evidence, and test comprehension before expanding.

AskTable.ai may be evaluated as the natural-language entry point. This article does not establish that every plan node, SQL evidence, approximation marker, or export explanation is exposed by the current product.

Public references

Ready to help your team start?

Talk through a real scenario and see how AskTable.ai can fit your business.

Book a demo