1. Where did this come from?
Every figure in a decision should trace back to a source you recognise: a system, a document, a date. An answer that cannot show where it came from is a draft, however confident it sounds.
What good looks like: each number links to the record it was read from, so you can check it in one step.
2. How current is it?
A debtors' figure from an export that is a week old and one that is an hour old can point to different decisions. AI agents read whatever they are given. Unless they are set up to say how old it is, the answer will not tell you.
What good looks like: the answer states the date and time of the data behind it.
3. What could it not see?
Gaps are silent. A branch that reports late, a mailbox the agent cannot open, or a spreadsheet on someone's laptop will not show up as missing. It will not show up at all.
What good looks like: the answer lists what it did not read, as clearly as what it did.
4. Is it comparing like with like?
One supplier quotes including value-added tax (VAT), another excluding it. One team counts an active customer as anyone who bought in the last 12 months, another as anyone who bought in the last 3 months. The same word can carry two meanings, and a summary will blend them without a warning.
What good looks like: the terms behind each comparison are defined once, and every source uses them the same way.
5. Which parts are found, and which are guessed?
A useful answer keeps three things apart: what it found in the records, what it calculated from them, and what it inferred. The first two you can check. The third is judgement, and it should be labelled as judgement.
What good looks like: inferences are marked, with the reasoning beside them.
6. Who signs it?
An AI agent can prepare a decision. A person owns it. If nobody's name is attached to the outcome, the answer will be acted on by default, and nobody will be accountable when it is wrong.
What good looks like: the person who decides is named, and so is what the agent may do without them. As the agent's answers prove reliable on one kind of question, it earns more room on that question, and only that one.
What the six have in common
None of these questions is about how clever the model is. They are about sources, timing, gaps, definitions and ownership. That is good news, because those are design choices a business controls.
An AI platform built for decision quality answers the first five questions before you ask them. It shows its sources, dates its data, names its gaps, holds one set of definitions and marks its guesses. The sixth question stays with you, by design.
If you find yourself asking all six every time, the answers are not the problem. The way the work is set up is, and that can be fixed.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets.
