How to find out with the AI Assistant why the AI took points in an assessment?

The AI Assistant compares the checklist criterion, the dialog, the knowledge base and the QA specialist's manual correction, and tells you who got it wrong: the AI, the criterion, the knowledge base or the human. With every verdict it gives verbatim quotes, shows where to look in the interface and suggests a draft fix. The assistant itself changes nothing.

When you need this

The AI took points for a criterion, but you think the operator did everything right.
QA specialists often correct AI assessments, and you need to understand why: the criterion wording, the knowledge base or the AI itself.
Before changing a criterion or the knowledge base you want an analysis with quotes, so you fix exactly what gets in the way.

Good to know

The verdict is the layer where a text change would have prevented the error. The assistant checks the layers in this order:
- Human — the AI decision follows directly from the criterion (and the knowledge base), and the manual correction relies on neither. No draft, only advice to review the correction.
- Criterion — the AI read the criterion the way its text allows: no topic boundary or exception, ambiguous wording. Draft — a change to the criterion text.
- Knowledge base — the rule in the knowledge base is wrong, outdated or missing. Draft — a text proposal for the document.
- AI — the needed rule is already written in the criterion or the knowledge base, and the AI contradicts it. Draft — a change to the criterion or to the evaluation system prompt.
No assessment error. If the QA specialist did not correct the criterion, and the AI took points by a rule written explicitly in the criterion (or in the knowledge base on the criterion's topic) and quoted the operator message that breaks it, the assistant answers "No assessment error: the penalty is justified" — it quotes the rule and the message, shows where to look and offers no draft.
When data is missing (no access, no dialog or knowledge base), the assistant writes "Verdict not given: missing …" and does not guess. A justified penalty is not missing data.
The knowledge base is read now. The assistant searches the knowledge base at the moment you ask, with the dataset's search settings. During the assessment the AI may have received other fragments: its queries are not logged. The answer carries a caveat about this.
If the criterion was changed after the assessment, the assistant warns you: the verdict is about the current criterion text.
A knowledge base search is a test query, the same as on the Search check screen. It changes nothing in the knowledge base, but it is saved in the knowledge base's test query history on the Fast Line Pro side and counted in the search metrics. The Search check screen does not show this history.
The assistant writes nothing. It does not change criteria, settings, the knowledge base or assessments and does not re-run the evaluation. You apply the draft fix yourself in the interface.
The assistant analyses up to 3 criteria at a time. It offers the rest as the next portion — write "continue".

Before you start

Your role has access to the AI Assistant module (how to use the AI Assistant).
Your role can view quality assessments: the Assessments section of the Quality Assurance module or the "View quality assessments" checkbox in the Operator Line rights. With the My Assessments section the assistant analyses only your own assessments. The Checklists section or the "Launch quality assessments" checkbox without viewing assessments open only checklists — that is not enough for diagnostics.
You have access to the assessment's dialog in the Operator Line (if your role is limited by skill groups or bots, the dialog may be unavailable).
You know the assessment number, the dialog number, or the checklist name and the period.

Step by step

1. Open the AI Assistant and ask

Click the "spark" icon button in the top bar.
Optionally type / and choose the "Quality Assurance — AI assessment diagnostics" skill. Without it the assistant also picks the skill from the question.
Ask in one of these ways:
- by assessment number: "Why did the AI take points in assessment 4055? Who is at fault and what should be fixed?";
- by criterion: "Analyse criterion 57 in assessment 4055";
- by dialog: "Analyse dialog 788610 — why the minus?";
- by manual corrections: "Show manual corrections of AI assessments for checklist 4 for 24–25.09 and tell me whose fault it is".
Send the request.

2. Read the answer

For each criterion the assistant answers in this order:

Verdict — Criterion 57 «Criterion name» — verdict: AI / Criterion / Knowledge base / Human and one sentence why. Or Criterion 57 «Criterion name» — no assessment error: the penalty is justified with the rule and the operator message quoted.
Quotes, verbatim: the criterion fragment, the AI comment, the QA specialist comment, the dialog message with its number, the knowledge base fragment with the document name.
Caveats, when they apply: the knowledge base was read now; the criterion was changed after the assessment.
Where to look — the path to the assessment "Quality Assurance → Assessments (or My Assessments) → #number" (a link only when it can be built from the address of the page where you opened the assistant), the checklist and criterion, the knowledge base document.
Draft fix — one or two sentences and a qa-fix-draft block with the exact text replacement. Below it — the line "Nothing was written". "Human" and "No assessment error" come without a draft.

If you asked about manual corrections for a period, the assistant takes the assessments where the QA specialist corrected at least one criterion — even when the total score did not change. The answer starts with lines like "Period: 28.09.2026 00:00 – 23:59 (Kyiv)" and "Reviewed assessments: 2, with manual corrections: 1, criteria analysed: 3". Without a period in the question the assistant takes the last 7 days, Kyiv time.

3. Apply the fix (if needed)

Open the checklist from the "Where to look" link and find the criterion.
Replace the fragment from the draft in the criterion text with the suggested one and save.
For a knowledge base fix, open the document from the link and change the fragment.

Note: saving a criterion updates the checklist's AI agent: its instructions are rebuilt from the evaluation system prompt and the criteria. Manual changes made directly in the agent's prompt are lost.

What happens next

The assistant only answers in the chat — no data on the platform changes. New rules apply to assessments launched after you save the criterion or the changed setting.

How to make sure it worked

The answer has a verdict, quotes, "Where to look" and, where needed, a draft fix.
The quotes match the text of the criterion, dialog and knowledge base behind the links.
After the session criteria and settings are unchanged, and no new assessments appeared.