Skip to content

Understand Insights and Explanations

AI insights in MiaRec typically include two parts:

1) a value (score/category/date/text), and
2) an explanation (short rationale referencing what happened in the conversation).

The value powers dashboards and search. The explanation helps humans quickly validate and act on the result.

Common insight types

Scores (numeric)

Examples: - CSAT 1–5 - NPS (detractor/passive/promoter or 0–10, depending on your configuration) - QA score (0–100)

How to interpret: - Scores are useful for trends and filtering. - They’re not “facts”—they are inferred from the conversation content.

Categories (single-select)

Examples: - Sentiment: Positive / Neutral / Negative - Outcome: Resolved / Unresolved - Lead Stage: Discovery / Negotiation / Closed

How to interpret: - Category labels come from your organization’s definitions. - If the category doesn’t match your expectation, check whether the transcript has enough evidence.

Multiple labels (multi-select)

Examples: - Topics: Billing, Cancellation, Product issue - Sales objections: Price, Competitor, Timing

How to interpret: - Multi-label insights can reflect multiple segments of the conversation. - Explanations should clarify why each label was selected (depending on configuration).

Extracted entities (text/date/amount)

Examples: - Reservation start date - Deal dollar amount - Competitor name - Next action text

How to interpret: - Entity extraction depends heavily on transcript/thread quality. - Confirm with the transcript if precision matters.

How to read explanations

A good explanation should: - be short (usually 1–3 sentences), - cite specific moments or statements (even informally), - connect evidence to the value.

When you review an insight: 1. Read the explanation. 2. Find the referenced moment in the transcript/thread. 3. Decide whether the value is “good enough” for the intended use (coaching, QA, reporting).

Tip: Explanations are designed to reduce “black box” frustration. Use them first—before escalating a “wrong score” issue.

Handling ambiguity

Sometimes the conversation doesn’t contain enough evidence to confidently assign a value: - the call is too short, - the customer is neutral/polite but not clearly satisfied, - the issue is unresolved but promised for later.

In these cases, your organization may: - assign a neutral score/category, - use an “Unknown / Not enough evidence” bucket (if configured), - rely more on explanation than on the value.

What to do if you disagree with an insight

  • If the explanation is clearly inconsistent with the transcript/thread, capture:
  • the conversation link/ID,
  • the insight value,
  • the part of the transcript that contradicts it.
  • Share it with your admin team. They may tune:
  • prompt definitions,
  • filters (exclude short calls),
  • or thresholds/buckets.