Interview Analysis

After an interview is completed, InsightAgent automatically generates AI-powered analysis including an executive summary and key quotes extracted from the conversation.

Overview

Interview analysis provides:

• Executive Summary - A concise paragraph summarizing the main topics discussed, key insights, and notable opinions shared by the expert
• Key Quotes - Notable verbatim quotes from the expert with context explaining their significance (agent quotes are excluded to focus on expert insights)

Analysis is generated automatically when an interview completes, or can be manually triggered for completed interviews.

Viewing Analysis

Analysis appears in the Analysis tab when viewing an interview. The tab displays:

1. Summary Section - The AI-generated executive summary
2. Key Quotes Section - Notable quotes with:
   • The exact quote text from the expert
   • Context explaining why the quote is significant
   • Clickable timestamp to jump to that point in the recording

Quote Navigation

Click on a quote's timestamp to:

• Seek the audio player to that moment
• Highlight the corresponding entry in the transcript

Interview Quality & Agent Conduct scores

Each completed interview receives two complementary scores, each from 0–100 and always accompanied by plain-language reasons — never a bare number:

• Interview Quality — how much usable insight the conversation produced against that interview's own objectives. The outcome: given what the agent was instructed to accomplish, how much substantive expert information did the conversation actually yield.
• Agent Conduct — how well the AI interviewer ran the conversation, given the expert in front of it: asking sharp, on-objective questions, probing thin answers, following up, and handling deflections or confidentiality gracefully. This is deliberately independent of how forthcoming the expert was — an interviewer that probes skilfully but meets a guarded or unavailable expert still scores high; one that under-probes a willing expert scores low.

Why two scores. A single number conflates two different things. An expert who declines to share details, or ends the call early, can cap the quality (insight gathered) even when the agent did everything right. Reading the two together tells you which is which — "the interview could have gone deeper" versus "a good interview that a guarded expert limited" — so you know whether to refine the interview or simply accept the outcome. Both scores are derived from each agent's own objectives, so they work across interview types with no configuration.

Score bands

Both scores use the same 0–100 bands:

• 0–30 — Quality: refused, cut short, or little usable insight. Conduct: drifted off-objective, failed to follow up, or ended prematurely.
• 31–60 — partially achieved, with notable gaps.
• 61–100 — substantially achieved; for conduct, the interviewer probed well, adapted, and stayed on objective.

Where they appear

• Interview detail → Analysis tab — the Interview Quality score with its "Why this score" reasons (collapsed by default; expand for detail), and alongside it the Agent Conduct score and reasons.
• Interviews list — an Interview Quality column, plus a quality filter on the Completed view (for example, pull all interviews scoring under 50 in one view).
• Billing → usage detail — both scores per interview, including in the CSV export, so you can review quality and conduct across a period in one view.
• Post-interview webhook — both scores and their reasons are sent in the completion webhook for every scored interview. See the API Reference.

Reading the two scores together

The gap between the two scores is the actionable signal. Read them as a pair:

Interview Quality	Agent Conduct	What it means
High	High	A strong interview — substantive answers, well run.
Low	High	The agent ran a good interview, but a guarded expert or an early exit limited the insight. Usually nothing to fix on the agent.
High	Low	Useful information still came out despite the agent under-probing or drifting — the agent's instructions likely have room to improve.
Low	Low	Both the insight gathered and how the conversation was run fell short — review the agent's instructions, and for complex flows consider splitting the work across multiple agents.

A recurring Low-Quality / High-Conduct pattern usually points at expert sourcing or scheduling rather than the agent. A Low-Conduct pattern points at the agent's prompt.

Incomplete and ended-early calls

Some interviews end before they finish — the expert drops off, hangs up, or the call is cut short. These calls correctly receive a low Interview Quality score, because little insight was gathered, and the detail view and interviews list mark them with an "Ended early" badge next to the score.

The badge is the signal that a low score reflects an incomplete interview, not a poorly performing agent — read it as "the call didn't finish," not "the agent failed." Agent Conduct on these calls is evaluated only against the portion that actually took place, so a well-run conversation that the expert ended early can still earn a solid conduct score while quality stays low.

Multi-step interviews

When an interview is handled by more than one agent — for example a short screening step that hands off to the main interview — the detail view shows a per-agent breakdown for each score: each part is evaluated against the agent that handled it. For Interview Quality, the overall weights the substantive interview so screening or hand-off steps don't inflate or deflate it. Single-agent interviews show only the overall scores.

Each score reflects only what the conversation actually covered: when an agent follows one branch of a multi-path script, it is judged on the path it took, not on steps that belonged to branches it never entered.

Not scored

An interview shows Not scored when there isn't enough to evaluate — for example no usable transcript, or the agent's objectives can't be determined. A score is never fabricated.

Regenerating Analysis

If you need to regenerate the analysis (for example, after the AI models have been improved), click the Regenerate button in the Summary section. This will:

• Generate a new summary
• Extract fresh quotes
• Replace the existing analysis

Account Settings

You can control which analysis features are enabled for your account:

1. Go to Settings in the sidebar
2. Under Post-Interview Analysis, toggle:
   • Summary Generation - Enable/disable automatic summary generation
   • Quote Extraction - Enable/disable automatic quote extraction

These settings apply to all new interviews. Existing interviews retain their generated analysis.

How It Works

Summary Generation

The AI analyzes the complete transcript and generates a factual summary based only on what was actually discussed. For brief or incomplete interviews, the summary will honestly reflect the limited content available.

Quote Extraction

Quotes are extracted directly from the expert's responses with strict validation:

• Only expert quotes are included (agent questions/statements are filtered out)
• Only exact verbatim quotes are included
• Each quote must appear word-for-word in the transcript
• Timestamps are verified against the actual transcript
• Hallucinated or paraphrased quotes are automatically rejected

API Access

Analysis data is included in the interview response:

{
  "id": "interview-uuid",
  "status": "COMPLETED",
  "summary": "The interview covered topics including...",
  "quotes": [
    {
      "quote": "Exact quote from the transcript",
      "speaker": "expert",
      "context": "This quote highlights the expert's view on...",
      "timestamp_secs": 145
    }
  ],
  "analysisGeneratedAt": "2026-01-09T10:00:00.000Z"
}

See the API Reference for complete details.

Tips

• Short interviews: If an interview is cut short, the analysis will note that the content is limited rather than fabricating details
• Multiple sessions: Analysis is generated from the primary (active) session's transcript
• Regeneration: Use regeneration sparingly - it consumes AI processing resources