A recorded conversation is only useful if someone reads it, and nobody reads transcripts. Call insights turn a call's transcript into a structured, skimmable analysis on the call record itself — what was discussed, how it went, and what should happen next. Insights are available where your organization has telephony configured and transcription producing text for the call; no transcript, no analysis worth having.
What the analysis contains
The insights panel on a call record presents:
- Key discussion points — a short prose summary of what the call was about.
- Sentiment — a positive / neutral / negative read of the conversation, shown as a tag.
- What went well and Areas for improvement — coaching notes, each with a severity level and, for improvement items, a concrete suggestion. This is the part aimed at the agent and their supervisor rather than at the record.
- Next best actions — suggested follow-ups grounded in the conversation.
- Key points — tag-style highlights for fast scanning.
- A suggested disposition — advisory only; the agent's own choice is what lands on the call. The analysis never writes the call's fields.
The same analysis surfaces in three places: the in-call panel while the conversation is live, the call record afterwards, and the call-details view of the supervisor console.
How it is produced and stored
Analysis runs on demand and is cached. Opening a call's insights panel fetches the stored analysis if one exists, or generates one from the transcript; Refresh forces a regeneration — useful when the transcript arrived after the first look. The result is stored with the call's transcript record, so it survives reloads and is generated once, not on every view.
The transcript is the single source of truth for the analysis. The generated text also uses what the platform already knows — the matched contact and account names, the call's direction — so summaries read like they were written by someone who knew who was on the line.
One honesty mechanism worth knowing: where no AI model is configured for call analysis, the panel can still produce a simpler rule-based summary, and it says so with an explicit "generated without an AI model" note. You will never mistake the fallback for the real thing.
What insights are not
Insights are a reading of the call, not a change to it. They do not edit the call record, do not set the disposition, do not create follow-up tasks by themselves, and are not an approval or workflow step. If a suggested next action deserves to exist as work, an agent creates the task — the same human-in-the-loop stance the platform's other AI surfaces take (see Asking Sense about your data for the general principles).
Common questions
Why does a call have no insights?
Because it has no transcript — the analysis is built from the transcript, and calls without one (transcription not enabled, the call never connected, or the text simply hasn't arrived yet) have nothing to analyze. If the transcript arrived after you first opened the panel, use Refresh to generate against it.
Is the sentiment or coaching visible to the customer?
No. Insights live on the call record and the supervisor console inside your organization — internal surfaces under your organization's permissions. Nothing about the analysis is sent outward.
Can I trust the suggested disposition?
Treat it as a proposal. It is derived from the same transcript as the summary and is often right, but the agent who was on the call picks the disposition that gets saved, and reporting is built on the agent's choice — precisely so a model's misread can't quietly skew your complaint or sale counts.