Sign in

Blog · AI and self-service analytics

When the assistant says 'I cannot compute that': what a refusal should look like

An analytics assistant that reads computed tables and declines arithmetic will refuse some requests, and how it refuses decides whether the reader trusts it. This guide sets out the four requests a well-built assistant declines, add these up, average this, extrapolate that, estimate what is missing, what each refusal says, where it points instead, the tone that keeps the reader, and the refusals that are wrong because the table exists.

The short answerA well-built analytics assistant declines four kinds of request: to add or subtract figures in prose, to average or aggregate, to extrapolate a trend, and to estimate a value the tables do not hold. Each refusal names the reason, points to the computed table that holds the answer or the tool that would compute it, and offers the closest thing it can do: read the total, cite the average the table already has, show the run rate with its window, or list what would be needed. A refusal that only says no is a failure; one that says where the answer is, is the assistant working.

An assistant that reads and never computes will be asked to compute, several times a day. What it says next is the difference between a reader who trusts it and one who stops using it. This guide sets out the four refusals, what each says, where it points, and the refusals that are wrong.

The four refusals

Request Why declined Refusal says Points to
"Add these three regions up" Arithmetic in prose is unverifiable "I read totals rather than compute them" The roll-up table's total row
"What is the average deal size?" Aggregation belongs in a computed table "The average is in the pipeline table, computed on won deals" The table, with its window
"Extrapolate this to year end" Extrapolation is a stated method, not a sentence "Run rate at a stated window is computed; here it is with its range" The run rate table
"Estimate what customer 4471 spends elsewhere" Estimates use a stated method in the norm table "The wallet estimate is by the norm method, on this line" The share of wallet table, with the method

What a good refusal contains

  1. The reason, in one clause.
  2. The place the answer lives, named.
  3. The answer read from there, if it exists.
  4. What would be needed, if it does not.

"I do not add figures. The coverage table's region row shows North at 61 percent, from 138 covered of 226 assigned, definitions v2026.3."

Refusal, location, answer, provenance. One sentence.

The tone

Plain, brief, no apology longer than the answer. The reader asked a reasonable question; the assistant explains the design once and reads the number. An assistant that apologises for three sentences before pointing at the table has made the refusal about itself.

Refusals that are wrong

Request Wrong refusal Why wrong
"Coverage for the North?" "I cannot compute regional figures" The coverage table has a North row; read it
"Which rep has the most untouched revenue?" "I cannot rank" The untouched list is ranked; read row one
"Has share of wallet moved?" "I cannot compare periods" The movements page lists it with the threshold

The rule: read first. Refuse only what no table holds.

A worked exchange

Reader: What is total pipeline for Q4 across all reps? Assistant: I read totals rather than add them. The pipeline table's total row for Q4 shows $8.2m open, $5.1m weighted by the team's stage conversion, as at 12 September. Table 7, row 26. Reader: And if it grows at last quarter's rate? Assistant: I do not extrapolate in prose. The run rate table shows pipeline creation at $1.4m a week on a four-week window, range $1.1m to $1.7m. Four weeks of that would be $5.6m at the median, and the table's projection row shows it with the range. Table 7, row 30.

Two refusals, two answers, four citations.

Where it goes wrong

Refusal without location. "I cannot do that." The reader leaves.

Refusal instead of reading. The table had it.

Computing anyway, hedged. "Roughly $8m, but check." The worst of both.

Apology longer than the answer. The refusal is about the assistant.

Every refusal, an answer beside it

Covirage's assistant reads computed tables, cites them, declines the four kinds of computation, and points at the table or the tool every time. The self-service analytics solution describes the setup, and the asking questions guide covers the questions that never need a refusal.

Questions people ask

Why refuse simple arithmetic at all?

Because a language model asked for a sum produces one, and it is sometimes wrong in a way that looks right. The computed table has the total, checked against the identity. The assistant reading it is reliable; the assistant computing it is a guess with confidence. The refusal is the design, not a limitation.

Does the refusal frustrate the reader?

It does if it stops there. It does not if it points at the table and reads the answer from it, which is what the reader wanted. 'I do not add figures; the coverage table's total row shows 61 percent for the region' is a refusal and an answer in one sentence.

When is a refusal wrong?

When the table exists and the assistant did not look. A request for regional coverage that is refused as arithmetic, when the coverage table has a regional row, is the assistant failing to read. The rule is: read first; refuse only what no table holds.