All writing

AI in a data estate

When the right answer is "I cannot tell"

The hardest behaviour to build into an automated analyst is the refusal to produce a number it cannot support.

1 minute read

Every system that reports figures faces the same temptation at the same point: a value is unavailable, and a zero is available. Zero renders well. It sums, it averages, it charts. It is also a measurement — a claim that the thing was measured and found to be nothing — and it is false.

The three states, not two

  • Measured, and it is a number.
  • Measured, and it is nothing — a real finding, and different from the first only in value.
  • Not measurable, with a reason. Not a number at all.

Collapsing the third into the second is how a dashboard ends up with an average computed over invented zeroes, and how a conclusion gets drawn from a denominator that was never true. Collapsing the second into the third is less dangerous but still wrong: it hides a real answer behind an apology.

The reason has to be specific

"Not available" is almost as bad as a zero, because it gives the reader nothing to do. "This platform bills per cluster rather than per query, so one tool's share of a shared cluster cannot be separated" is a different kind of statement: it is a finding about the platform, it tells the reader what would change the answer, and it is frequently the most useful line on the page because it is the next piece of work.

A figure you cannot take is often more informative than one you can, provided you say why.

Run this against your own estate

Gecko reads metadata and query history — never the rows in your tables — and tells you what it found.

Get started