Ask: answers that show their work

The problem with a natural-language answer is not that it is sometimes wrong. It is that a wrong one looks exactly like a right one.

Ask is Nexatron's question surface: a person asks in plain English or by voice, and the platform returns the answer together with the query it ran, the sources it read, and a confidence signal. The generated query is shown rather than hidden, so an answer can be checked by anyone who can read SQL instead of being taken on trust.

What it does

The query is part of the answerEvery response carries the SQL, SOQL or source request that produced it. Not in a debug panel, next to the number, where the person deciding something can read it or hand it to someone who can.
A confidence signal, and what moved itAnswers carry a score with its breakdown: whether a certified metric was used, how well the question matched the schema, whether validation passed. A low score is shown as a low score rather than smoothed away.
Governed and ungoverned are visibly differentAn answer built from a certified metric is labelled differently from one built from raw tables. Both can be useful; presenting them with identical authority is how an ad-hoc query becomes a board number.
Voice in, spoken summaries outAsk by voice and hear the answer read back, with the same governance and the same evidence as a typed question.
Follow-ups keep the thread"Now break that down by region" resolves against the previous answer rather than starting over, so a line of enquiry behaves like a conversation.

What it deliberately won’t do

These are specifications, not caveats. Each one is a case where returning something would have been easy and wrong.

It will refuse rather than approximateAsk for a dimension the connected user cannot see and it says which field it could not read. It does not substitute a similar column and it does not quietly drop the dimension from the breakdown.
A zero is reported as a zero, not as nothingAn empty result says the filters matched no rows, which is a different statement from the data not existing, and the difference has cost people money.

Test the claim, don’t take it

Ask it something you already know the answer to, then ask it something your data cannot answer. Those two questions tell you more than a feature list.

How we measure

What it does

The query is part of the answerEvery response carries the SQL, SOQL or source request that produced it. Not in a debug panel, next to the number, where the person deciding something can read it or hand it to someone who can.

A confidence signal, and what moved itAnswers carry a score with its breakdown: whether a certified metric was used, how well the question matched the schema, whether validation passed. A low score is shown as a low score rather than smoothed away.

Governed and ungoverned are visibly differentAn answer built from a certified metric is labelled differently from one built from raw tables. Both can be useful; presenting them with identical authority is how an ad-hoc query becomes a board number.

Voice in, spoken summaries outAsk by voice and hear the answer read back, with the same governance and the same evidence as a typed question.

Follow-ups keep the thread"Now break that down by region" resolves against the previous answer rather than starting over, so a line of enquiry behaves like a conversation.

What it deliberately won’t do

These are specifications, not caveats. Each one is a case where returning something would have been easy and wrong.

It will refuse rather than approximateAsk for a dimension the connected user cannot see and it says which field it could not read. It does not substitute a similar column and it does not quietly drop the dimension from the breakdown.

A zero is reported as a zero, not as nothingAn empty result says the filters matched no rows, which is a different statement from the data not existing, and the difference has cost people money.