Sign in

Blog · AI and self-service analytics

How to ask an analytics assistant a question it can answer reliably

What separates a question an AI analytics assistant answers correctly from one it answers plausibly: naming the measure, the level, the period and the comparison, asking for the table before the sentence, and checking the citation. Six rewrites of common questions, and the four kinds of question an assistant should refuse.

The short answerA reliable question names four things: the measure, the level it is measured at, the period, and what it is compared to. 'How are we doing in the north?' has none of them. 'Coverage per rep in the north region, this quarter, against last quarter' has all four, and an assistant built on computed tables can answer it with a citation to the table. Ask for the table before the sentence, and treat a number with no table behind it as unverified.

An analytics assistant will answer any question you ask it. The difference between a correct answer and a fluent one is mostly in the question. This guide sets out the four things a reliable question names, six rewrites of questions people actually ask, and the questions an assistant should decline.

Four things a question names

Element Vague Named
Measure "how are we doing" coverage, share of wallet, dormant accounts
Level "the north" per rep in the north region
Period "lately" this quarter to date
Comparison none against last quarter, against the norm, against plan

A question with all four maps to a computed table. A question missing any of them makes the assistant choose, and it chooses silently.

Six rewrites

As asked As answerable
"Who should I call this week?" "Untouched accounts per rep, this week, ranked by prior-year revenue"
"Is the pipeline healthy?" "Pipeline coverage per rep by stage for this quarter, against the needed multiple"
"Which customers are we losing?" "Accounts with no orders in 90 days that ordered monthly last year, ranked by run rate"
"How is the forecast looking?" "Forecast bridge from last week to this week, deal by deal, for this quarter"
"Are we compliant?" "Contracted share by category this month, against the agreement"
"What changed?" "Measures that moved more than their stated threshold since last week, per region"

Every rewrite on the right is a table that exists. Every question on the left is an invitation to guess.

Ask for the table first

Ask for the table, then the sentence. A number in a sentence with no table behind it is a claim. The table is the evidence, and a good assistant produces it with the citation: rows, period, definition version. The sentence is the assistant's reading of the table, and it is the part that can be wrong.

Questions to refuse

An assistant built on computed tables should not do arithmetic in prose, and should say so:

  • "Add these up." The roll-up is a computed table with an identity check. The assistant points to it.
  • "What is the average?" Averages are computed once, with a stated method, and read.
  • "Extrapolate this." Run rate and trend are computed measures with stated windows.
  • "Estimate what we do not have." A missing wallet is estimated by a stated method in the table, not improvised in an answer.

A refusal of this kind is the assistant working correctly. The number it would have produced would have been plausible.

The identity, for the reader

Any number the assistant gives should be traceable:

the number = a cell in a computed table = rows in an upload, under a definition version

If the chain breaks, the number is not evidence.

Where it goes wrong

Level unstated. The assistant picks the company level and the reader wanted the rep.

Period unstated. Trailing twelve months when the reader meant this quarter.

Sentence without table. The fluent wrong answer.

Arithmetic in the answer. Every generated model can produce a sum that is wrong. Read; do not compute.

Every question, four elements

Covirage's assistant reads computed tables and cites them, and declines to compute. The self-service analytics solution describes the setup, and the arithmetic guide covers why the model reads rather than calculates.

Questions people ask

Why does the level matter so much?

Because the same measure means different things at different levels. Share of wallet for a customer, a rep, a region and the company are four numbers that roll up but are not interchangeable. An assistant asked for 'our share of wallet' has to guess the level, and it will.

What is a citation from an assistant?

A reference to the computed table the number came from: which rows, which period, which definition version. If the assistant cannot produce it, the number was generated rather than read, and it should not go in a deck.

Which questions should it refuse?

Anything that requires it to compute rather than read: 'add these up for me', 'what is the average of', 'extrapolate this trend'. A well-built assistant answers those by pointing to the tool that computes it, not by doing arithmetic in a sentence.