Analyzing conversations

Analyzing conversations well means reading two layers together:

Quantitative — the what. How much is happening: the trend in number of conversations, number of messages, and number of action calls.
Qualitative — the why behind the what. The intent breakdown (action vs. information), which questions and actions come up most, where the action gaps and knowledge gaps are, and what drives positive and negative sentiment.

The quantitative layer tells you that something changed. The qualitative layer tells you why, and what to do about it. Read them in that order, then watch a handful of the actual conversations to confirm what you concluded.

flowchart LR
  A["1. Quantitative<br/>the what"] --> B["2. Qualitative<br/>the why"]
  B --> C["3. Watch<br/>session replay"]
  C --> D["Act<br/>knowledge or action"]
  style A fill:#e8ecff,stroke:#1842ef,color:#0a0e1a
  style B fill:#fef3c7,stroke:#d97706,color:#0a0e1a
  style C fill:#dbe2ff,stroke:#1842ef,color:#0a0e1a
  style D fill:#d1fae5,stroke:#059669,color:#0a0e1a

Part 1: Quantitative — the what

Path: Analytics → Dashboard → Conversations

Start with volume. These three numbers tell you how much work the agent is doing and how hard users are working to get it.

Measure	Widget	What it tells you
Number of conversations	Conversations	Overall engagement — how much the agent is being used.
Number of messages	User Messages	Total volume of what users send the agent.
Action calls	Action Calls	How often the agent does real work, not just answers.

Read them as a set, not in isolation:

Conversations rising is engagement growing — but on its own it’s ambiguous, because retries after a failure also raise the count. Pair it with the qualitative layer to know which.
Messages rising faster than conversations means users are working harder for the same result. That usually points to a knowledge or clarity problem, not a capability one.
Action calls rising is the agent resolving by doing. Action calls growing alongside steady engagement is the healthiest pattern on the dashboard.

Each widget supports the time-range selector and an Agent filter, so you can see whether a trend is fleet-wide or concentrated in one agent.

Part 2: Qualitative — the why behind the what

The numbers above don’t tell you what users are asking or why a conversation went well or badly. That lives in the signal tags on each conversation — surfaced in aggregate on Signals frequency and User intent, and readable one by one in Agent Studio → Conversations, where Agent Knowledge Gap and Agent Action Gap are among the filters you can apply.

Work through the qualitative layer in this order: the intent breakdown, then the top asked questions and top-used actions, then the action gaps and knowledge gaps, then what drives positive and negative sentiment. Each step narrows the next.

Intent breakdown — action vs. information

User intent classifies each conversation into what the user was trying to do:

Information Request — asking for information.
Action Request — wanting the agent to do something.
Reported Issue — reporting a problem.

The breakdown tells you what the agent is really for. A rising Action Request share means users increasingly want the agent to act, not just answer — which raises the value of closing action gaps. A rising Reported Issue line is usually a product problem worth investigating, not an agent one.

Top-used actions

Next, the action side: see which capabilities carry the load. In Reports, add the Action Called metric and break it down by Action key — a stacked bar of which actions are called most over time.

The top actions are what users actually rely on the agent to do. Keep them healthy — a regression there hurts more than anywhere else.
An action with zero or near-zero usage is either broken, undiscoverable by the agent, or solving a problem users don’t have. Check its wiring before assuming the last one.

Action gaps

The inverse of top-used actions: what users ask the agent to do that it can’t. Every such conversation is tagged Agent Action Gap, ranked in aggregate on Signals frequency.

A rising Action Gap share means users want capabilities that don’t exist yet. Filter Conversations by the tag, cluster the requests by verb, and a repeated verb is a candidate action. Before building, check whether the action already exists and the agent just isn’t calling it — that’s an instruction fix, not a build. A repeated theme is a roadmap item; fifty one-off requests are a signal about scope, not a backlog.

Knowledge gaps

The same pattern for questions: where the agent lacked the content to answer, the conversation is tagged Agent Knowledge Gap.

A rising Knowledge Gap share is a reading list for the Knowledge Base — filter Conversations by the tag, cluster the questions, and check with View Analysis whether the answer was missing, stale, or simply not retrieved. Missing content needs writing; stale content needs updating; unretrieved content is a retrieval problem where another article won’t help — see Knowledge validation.

What drives positive sentiment

Ask the embedded Foldspace agent on the Conversations page to surface the topics correlated with the positive signals — “What topics get the most thumbs up?” or “What topics correlate with positive sentiment?” It reads the same tags and topics across every conversation, so the answer is measured, not anecdotal.

The topics it surfaces are the questions and tasks the agent handles well — the ground you want to protect as you change things, and the pattern to extend to adjacent topics.

What drives negative sentiment

Ask the inverse — “What topics get the most thumbs down?” or “What topics correlate with negative sentiment?” — to get the topics dragging sentiment down.

Then read a few of the underlying conversations behind each topic. Filter by User Feedback = thumbs down or User Sentiment = Negative, and on a rated reply read the message before it — the rating lands on the reply, but the cause is usually the turn before. What you find routes the fix: a content problem goes to the Knowledge Base, a capability problem is an action gap, and a product problem the agent reported accurately belongs on the product backlog, not in the agent.

Part 3: Watch the conversations behind the numbers

The tags and topics tell you what to look at; watching the actual conversation tells you why it happened. For a conversation you’ve flagged, Session replay shows exactly what the user did on screen, with the conversation events overlaid as a timeline.

Reach for it when the transcript alone doesn’t explain the outcome:

A conversation that escalated — watch what the user was doing right before they asked, to see what the agent should have handled.
A thumbs-down whose cause isn’t obvious from the text — the screen often shows the real intent behind an ambiguous message.
A highly positive conversation on a valuable task — worth understanding so you can extend the pattern.

Turn it into a worklist

Every pass should end with something to ship, sorted by how often it comes up and how much it costs when it does:

Knowledge gaps → add or refresh a Knowledge Base article. Usually the cheapest fix, and one good article can close a long tail of related questions.
Action gaps → build or fix an action. More effort, but it converts conversations the agent can’t resolve today into ones it can.

Then re-check the tag’s share next period. If it dropped, the fix worked; if it held steady, go back and read the new ones.

Conversational analytics overview — the widgets used in Part 1 and Part 2.
Conversations (HITL) — the qualitative surface for reading and filtering conversations.
Session replay — watching the conversations behind the numbers.
Knowledge validation and The action lifecycle — where the fixes get made.