Understanding the question
What the question actually asks, and what it commits you to. Trick questions and implicit commitments are identified here.
The anatomy of a defensible answer.
An AI-generated answer is only worth something if you can say where it came from, how solid it is, and who accepted it. Here are the six things that travel with every answer we produce, and the controls that produce them.
What is a defensible answer?
A defensible answer is one its author can stand behind in front of a third party: it cites the document, page and version that evidence it, carries a confidence score, and was approved by a named person on a known date. It can be replayed later to reconstruct how it was produced.
Six elements
What travels with every answer
Taken from a real information security requirement: the question asks how the organisation determines the internal and external issues relevant to its management system.
- 01
The answer text
Written in the language and register the recipient expects, with no vague restatement: what is done, by whom, how often.
- 02
The source: document, page, version
Not "per our security policy" but the file name, its version and the exact pages that carry the claim.
- 03Confidence score for this answer, out of 100.
The score, 0 to 100
A measure of how solidly the cited evidence supports the answer, not a measure of how confident the model feels.
- 04
The explanation of the score
Analysis, improvements identified, recommendation: the reviewer knows what to fix, not merely that something is wrong.
- 05
The approver and timestamp
A name and a date. The workflow blocks the sending of an unapproved answer: accountability stays human and personal.
- 06
Replayability
Months later you can reconstruct the question, the sources in the version they had then, the score and the approval.
The control
Five layers, including seven anti-hallucination checks
Every answer passes through the five layers in order. An answer that fails a layer is not sent: it is corrected, or declared incomplete.
Search in your repository
Only your indexed documents, with their versions. No external source, no general knowledge from the model.
Drafting by domain agents
The relevant function and sector agents write the answer and tie it to its evidence, sentence by sentence.
Seven anti-hallucination checks
A second team of agents checks the first team’s output: source exists, faithful to the passage, version in force, no internal contradiction, requirement fully covered, consistent with history, commitment detected.
Named human approval
The last layer is not automatic. A person accepts the answer, and their name stays attached to what is sent.
The objection, head on
“A confidence score means the AI can be wrong?”
Yes. And that is exactly why it is shown.
Any system that produces text can be wrong, including a human under deadline pressure on a Friday evening. The useful question is not "does your AI ever get it wrong?" but "will you see the error before you sign?".
The score answers that question. It concentrates human review on the fragile answers instead of spreading it thinly across one hundred and eighty lines. An answer scored 95 against a policy in force is reviewed in ten seconds; an answer scored 40 is treated as a gap.
An AI that never doubts cannot be audited: nothing in its output distinguishes what is solidly evidenced from what is merely plausible. That is precisely the risk a general-purpose assistant carries when it is used on a document that binds the organisation.
The score, in plain terms
What raises it, what lowers it
It rises when
the source is the version in force; the cited passage answers the question directly; several independent documents agree; the answer was already approved by a named person in an equivalent context; the figure is dated and its method published.
It falls when
the source predates the latest known revision; the evidence is indirect; part of the requirement is uncovered; two of your documents contradict each other; the answer assumes a practice no document describes.
Frequently asked
The mechanism, in six questions
01
What does “replayable” mean?
That months later you can ask the system how an answer was produced: which question, which sources in which version, which score, which approver, on what date. It is the first thing an auditor or a lawyer asks for.
02
Does the model use knowledge from outside our documents?
To understand the question and the framework invoked, yes, that is the expertise trained into the agents. To assert a fact about your organisation, no: only your indexed documents count, and the claim cites its source.
03
Can human approval be switched off to go faster?
No. The fifth layer is the point of the system: without a name on the answer there is no accountability, and therefore nothing defensible. What the platform reduces is the volume to review, not the requirement to approve.
04
How many answers need reviewing in practice?
It depends on your documents, but the observed pattern is that most answers are approved on a quick read, with attention concentrated on low scores and detected commitments. On a 180-line questionnaire that typically means about an hour and a half.
05
What is an agent, concretely?
A component specialised in one domain, information security, public procurement, ESG reporting, a sector, with its own evidence criteria. There are 85: 72 function agents, 12 sector agents, and one librarian that indexes your documents.
06
Does the recipient see the scores?
No, unless you decide otherwise. Scores and annotations are internal review tools; the client, buyer or auditor receives the approved answers and, if requested, the documents cited.
Judge the mechanism on one of your real questions.
Bring the hardest question from your last questionnaire. We show the answer, its source, its score, and what would be missing to reach 100.