Findings
Recurring friction, knowledge gaps, customer feedback, product recommendations and sales opportunities - and how to work through them.
Shown in Companion Dashboard only - this view isn't currently available on the org-wide Dashboard.
Below the heatmap the report lists findings in five sections, in the order they appear on the tab. Every section works the same way, so this page covers all of them. Each one also has a card at the top of the tab that counts it and scrolls straight to it.
The five sections
| Section | What it collects | Who usually acts on it |
|---|---|---|
| Recurring Friction | Difficulties customers expressed with the product itself - a fault they reported, something not working, confusion about how it works, a limitation they hit. Every reported fault lands here, including one the customer only described to raise a support ticket. Not complaints about the chat. | Product / UX |
| Knowledge Gaps | Specific product questions inside the companion's domain that it did not answer, and what it was missing - plus answers customers said were wrong in their feedback. Out-of-scope requests it correctly declined, and requests about the chat itself (switching language, reaching support), are never gaps. Each one marks its conversation unresolved. | Whoever owns the manual or knowledge base |
| Customer Feedback | Themes in what customers wrote when rating a conversation. | Product / support |
| Product Recommendations | Changes to the product that customers asked for, or clearly wished for after hitting a limitation. Never suggestions the analysis made about the companion itself. | Product |
| Sales Opportunities | Accessories, upgrades, compatible products or pricing that customers showed interest in buying. Never sales tactics the analysis would suggest. | Sales |
Customer Feedback is the one section not derived from the transcript. Every other finding is a model's reading of what happened; these are themes pulled from what the customer themselves typed in the feedback box. Short or empty feedback ("ok", "thanks") produces nothing - only feedback with substance becomes a finding, and specific praise counts, since it tells you what to keep.
Two more kinds of finding appear elsewhere on the tab rather than as their own section:
- What customers ask - the product questions themselves, ranked by how many conversations raised them, in the card above the heatmap. Voice checks, language requests and small talk are left out.
- Product areas - the emergent feature taxonomy, shown as the heatmap.
One finding, one title
Findings are written in English whatever language the customer used, so the same question asked in Hindi and in English is counted as one finding - the source cards keep the customer's original words. Titles name the product and the specific thing about it ("Battery life of the Sony A7 III", not "Product details"), and questions stay questions. Two findings in the same section never share a title: a conversation that would produce a title already in use is merged into that finding.
The columns
| Column | What it shows |
|---|---|
| Finding | The settled wording, covering every conversation merged into it, with a one-line summary beneath. |
| Chats | How many conversations raised it, scoped to the selected date range. Findings seen in fewer than three are tagged Emerging. |
| Trend | A sparkline, which turns red when the finding is rising - so a growing problem is visible without reading the number. |
| Severity | A dot, from grey (none) through amber to red (critical). |
| Status | Your workflow state - see below. |
| Priority | A single ranked score. |
How Priority is calculated
Priority combines five things, so the top of the list is what's worth fixing rather than merely what's most frequent:
- How often it comes up (volume)
- How severe it is when it does
- How often those conversations go unresolved
- How badly customers rated the conversations that raised it - a finding people scored 1-2 stars outranks an equally common one they scored 4-5. Findings nobody rated are unaffected, never penalised.
- How recently it's been happening - an issue that stopped months ago sinks
Sorting by Chats instead gives you pure frequency when that's what you want.
Working through findings
Click a finding to open it. The panel shows the settled summary and every source conversation as a card with who was chatting, when, and the quote from that transcript.
Where the customer rated that conversation or left written feedback, it appears on the source card too - their own score and words next to the model's summary of what happened.
Click a source card to open the full transcript in a new tab.
If any of a finding's conversations were rated, the panel also shows the average rating for that finding and the spread of stars behind it.
Status
Each finding carries a workflow state you set: New, Acknowledged, In Progress, Resolved or Dismissed. It's shared - your whole organization sees the same state - and it drives the "still open" count on that section's card at the top of the tab, so keeping it current keeps those numbers meaningful.
Nothing sets status automatically. A finding that keeps recurring after you mark it resolved will keep gaining conversations and climbing the priority list again, which is the signal that the fix didn't land.
How to use this
- Treat Recurring Friction as the product backlog: high count plus high severity plus a rising trend is a genuine problem.
- Work Knowledge Gaps top-down. Each one is a specific question your content doesn't answer, with the evidence attached - the shortest path to a better resolution rate.
- Read Customer Feedback when you want the customer's own words rather than an inference - it is the shortest path to why a number moved.
- Hand Sales Opportunities to whoever owns revenue. Open one to see Where It Comes From - interest concentrated in one country is a regional play.
- Weigh Emerging findings lightly. One conversation is an anecdote; act once the count grows.
- Read two or three source conversations before acting. The count tells you it matters; the transcript tells you why.
Last updated on