# Reading Conversation Insights

**Dashboard → Conversation Insights** *(sign-in required)* is built from the analysis that runs on every conversation when it closes: the summary prompt under **Settings (the gear icon) → Analytics → Analytics & Insights** plus the platform's own fields (outcome, intents, topics, promises, unanswered questions) and the fields of any process pack you have turned on. This article is about how the cards count, so a figure can be read with the right question in mind.

## Which conversations a card counts

Every card on the tab reads **every conversation on every channel** — web chat, SMS threads, calls and email threads — that is closed and billable (a call: one that connected), in the dashboard's date range, handled by a production agent (drafts and test agents are left out). Since October 2026 the tab counts all four channels; before that it counted web chat only, so figures for earlier dates changed when it did: those conversations were always analyzed, they were simply never counted here.

The line above the cards says what the range holds — *N conversations · chat · SMS · calls · email · N handled by your team* — and every number on it is a click that narrows the tab to that part.

**The date** is when the conversation **started** (a chat, a text thread, a call), cut at midnight in your workspace's time zone. An email thread is dated by its **last message**, the same way the Agents page dates it, because a thread can stay open for weeks. A card that reads a field your prompt or a pack added shows *Fields available since <date>* when the range starts before that field existed; conversations from before it were never asked for the field.

## Narrowing by channel, or by who handled it

The filter row above the cards narrows **every card, the line and the drill-downs together**:

- **Channel** — All channels, or any mix of Chat, SMS, Calls and Email.
- **Handled by** — Anyone, or **AI agents** (the agent handled it alone), **Your team** (a person on your team handled it: a call dialed from a contact's page, a text sent from inside a thread, an email sent from the CRM), or **Both** (the agent and a person).

The selection is in the page address, so a filtered view can be shared as a link, and your browser remembers your last choice.

A call your team dialed has no transcript and no analysis: it counts on the line and under *Your team*, but no card reads anything from it. Who handled a conversation is written from the record — a send by a user, a call a user placed, a colleague's answer relayed into a chat — never from the agent's wording.

## The cards and what they divide by

| Card | What it counts | Divides by |
| --- | --- | --- |
| **How conversations ended** | The outcome of each analyzed conversation: resolved, likely resolved, escalated, incomplete, needs review. | Conversations with an outcome. *Likely resolved* is its own slice here. |
| **What customers came to do** | Every intent recorded on a conversation. One conversation can carry several (inquire and book, for example). | Intent mentions, not conversations. |
| **Opening topic to closing topic** | Where each conversation started and ended, by topic. Links under three conversations are not drawn, and the footnote says how many were left out. | Conversations with a topic path. |
| **Why conversations stalled** | The drop-off reason on conversations that ended incomplete or escalated. | Those conversations only. |
| **Promises made**, **Unanswered questions** | Conversations where the agent made a commitment, or a question went unanswered. | Analyzed conversations in the range. |
| **Customer effort by outcome** | Effort (low, medium, high) against outcome. Cells under three conversations show `<3`. | Conversations with an effort score. |
| **Process pack stat tiles** (Booking requests, Cancellation requests, …) | Conversations where the pack's request flag is true. | **The conversations the pack evaluated** — the pack writes its flag on every analysis once it is on, so conversations analyzed before the pack was enabled, and never back-filled, are not in the denominator. The tile says *of N conversations the pack evaluated*. |
| **Booking rate by service** (Booking pack) | The share of booking requests that ended booked or rescheduled, per service and action. | Booking requests in each cell. |
| **From conversation to booked job** (Workiz) | The block at the top of the tab: a stream — conversations → got a price → asked to book → got a job → done — with the job step read from the Workiz record, not the transcript, a computed headline, the drops between steps, and one sentence naming the worst-converting step and its reasons. See [Counting the jobs your agents book in Workiz](/help/integrations/workiz-bookings). | Each step divides by the previous one. |

Cards about the **page a chat started on** or the **Google Ads click** behind it only ever apply to web chat; with other channels in the range, those conversations sit in no bucket and the card says how many it could read.

## Drilling into a card

Clicking a bar, a slice or a cell lists the conversations behind it, with a **Channel** column. A chat opens in place; a text thread, a call or an email thread opens its own page with the **Open** link. An email thread that is not in a group inbox has no page yet and says so.

## Two figures that look alike and differ

**Resolved on an agent's Conversations page vs. the donut here.** The **Resolved** tile on an agent's Conversations page counts *resolved* **and** *likely resolved* together, over every analyzed conversation of that agent whatever its status, and the tile's hint says how many were likely resolved. The **How conversations ended** donut shows *likely resolved* as its own slice, over closed, billable conversations of all production agents on every channel. The two are different questions, and will not reconcile; neither is wrong.

**Calls on the Overview tab vs. calls here.** The Overview tab's **Calls** card counts calls your AI agents handled. This tab also counts the calls your team dialed, under *Your team*. The difference is the team's own calls.

**"Book" under What customers came to do vs. bookings.** An intent is what the customer came to do; it says nothing about whether a job was booked. For bookings, read the Booking pack's cards (what the analysis found) or the Workiz card (what the agents wrote).

## When a card reads "verified"

With Workiz connected, a conversation whose own booking tool created a job carries that job, and its analysis is written as *booked · verified* regardless of what the AI read in the transcript. A conversation the analysis calls booked with no job behind it is marked *unverified*. The Booking pack's cards count the analysis either way; the Workiz card counts only the jobs.

## Back-filling history for a pack

**Analytics & Insights → Processes → Back-fill recent history** re-analyzes conversations closed since a date that lack a pack's fields, on every channel: the estimate reads *N chats · N texts · N calls · N emails*, and each channel is queued its own way and processed gradually by the background sweep.