Conversations and Quality
Two pages help you keep your agent sharp: Conversations shows every chat with a visitor, and Reports shows how well it is doing overall. Used together, they tell you what people ask, where your agent shines, and where your Knowledge Base needs a little more.
Conversations
Open Conversations in the sidebar to see every visitor chat. Filter by status (Active, Finished, Handed to a person), date, or keyword, or show only the ones you have flagged.
The Conversations page showing the outcome metrics strip, the filter chips, the CSV export button, and the conversation list with outcomes
Click any conversation to open it. You will see when it started, how it ended (Resolved, Handed to a person, Finished, Customer left, or Active - Finished means the visitor left after your agent had answered; Customer left means they left with a question still open), the visitor's email if they shared one, their city and timezone, and the full back-and-forth. If your agent searched your Knowledge Base or used a tool, you can expand it to see what it looked for. Your own chats with Jarvis while building the agent are not listed here - only conversations with the agent are.
As you review, you can:
- Flag a conversation that needs your team's attention
- Rate it 1-5 stars to track quality over time
- Add a note for internal context (visitors never see notes)
Reports
Reports gives you the big picture. Open Reports in the sidebar, right under Conversations, or choose See all reports under the numbers on your Home page. Pick the time range (7, 30, or 90 days) with the selector at the top right; it opens on 30 days.
The Reports page showing the headline metric tiles, the time range picker, the Trends by week panel with CSV export, and the By topic breakdown
From top to bottom the page shows the headline numbers, Conversations to review, Customer feedback, Trends by week, By topic, and two charts (Resolution rate over time and Response times). Hover any headline number to see what it counts and the window it covers, such as "in the last 30 days".
The headline numbers:
- Resolution rate - how often your agent resolves a chat window conversation on its own (resolved conversations out of all that finished). Above 70% is strong; below 50% usually means content gaps.
- Handover rate - how often it hands the conversation to a person. Some is healthy and expected.
- Messages per resolved chat - how many customer messages it takes to resolve a chat. Fewer means quicker answers.
- Average response time - how long a visitor waits for the first reply.
- Helpful answers - how visitors rated the answers with the thumbs in the chat, counted per conversation and shown as "94% positive · 17 conversations". A conversation counts as positive when its ratings are mostly thumbs-up.
- Conversations - how many chats your agent had in the range.
With very few conversations the tiles show "Collecting data" instead of a percentage - a 0% or 100% computed from three chats would be noise, not truth. Rates appear once about ten conversations have finished; satisfaction appears once five conversations have been rated.
Customer feedback
Customer feedback shows what your customers said when the chat asked them (see Rating the Chat and NPS):
- Customer rating - the share of customers who rated a chat 4 or 5, with the average out of 5, and how many were satisfied (4 or 5), neutral (3) and unsatisfied (1 or 2).
- Would-recommend score (NPS) - customers answer how likely they are to recommend you, from 0 to 10; promoters (9 or 10) minus detractors (0 to 6), from -100 to 100, with the count of each. NPS is asked of returning customers, at most once every 90 days, so it fills more slowly than the chat rating.
- What customers said - the comments customers left with a rating, at any score, newest first. View opens the conversation. A card number or similar detail a customer types is replaced before the comment is stored.
Until there are enough answers to mean something, each number shows Collecting with the count so far: 10 chat ratings, or 30 NPS answers. If every agent has both questions switched off, the card says ratings are off and links to where to turn them back on.
Seeing the conversations behind a number
On the Pro plan and above, the Resolution rate, Handover rate, and Conversations tiles are clickable - each one opens the Conversations page pre-filtered to exactly the chats that number was computed from, so "78% resolved" is something you can inspect, not just read. The Trends by week panel below the tiles charts conversations, resolution rate, and satisfaction week over week, and the 90-day range unlocks with it. On other plans these show as locked previews with an upgrade path - never hidden, never fake numbers.
Topics (Pro)
The By topic panel on the Reports page breaks conversations down by what they were about - Billing, Technical, Account, Shipping, Returns, or General - with each topic's volume, resolution rate, and handovers. New conversations are categorized automatically overnight, so a topic's first numbers appear the day after the traffic. It is the quickest way to spot patterns like "billing questions are handed to a person twice as often as everything else."
An Unclassified row collects conversations from before topic tracking started (older history is never re-processed) plus the rare chat that could not be categorized - it always stays in the totals, so the numbers you see add up. When one topic's resolution rate falls well below your average, the dashboard raises a suggestion pointing you at this view.
Exporting to CSV (Pro)
Three surfaces export as CSV files you can open in Excel or Google Sheets, or forward as-is:
- Conversations - the download button next to the refresh icon exports the list with your current filters applied (outcome, dates, search, and so on), including each conversation's topic. Very large result sets export the most recent 5,000 rows and the page tells you when that cap applied - narrow the date range to get the rest.
- Trends by week and By topic - the download button on each panel exports exactly the numbers the panel shows.
Exports contain the same data you can already see on the page - nothing more.
Conversations to review
The Conversations to review list sits right under the headline numbers. It shows chats worth a second look, newest first, each with a View link that opens the conversation. These are helpful nudges, not failures:
- May have given a wrong answer - your agent may have said something your content does not clearly support. Usually a sign the topic needs better coverage.
- Customer seemed frustrated - signals the visitor was not fully satisfied (for example, repeated questions).
- Went in circles - your agent gave similar answers more than once, often because it could not tell follow-ups apart.
- Reached the message limit - the chat hit the most messages your agent allows in one conversation.
When there is nothing to look at, the list says "Nothing to review in this period."
Chats to review are most common when you are new. Each one points at where your Knowledge Base can improve - as you add content in response, fewer chats land here.
Close the Gaps
The fastest way to raise your resolution rate is to fix what visitors ask about but your content does not cover. Go to Knowledge Base > Unanswered questions (also called gaps or missing answers) to see those questions, ranked by how often they come up.
The Unanswered questions tab listing the questions still to do, with their status filter and the Draft the answer action
For each question, Draft the answer has the AI write an article starting point - a structured outline with the question phrased the way customers ask it and clear placeholders for the details only you know. The draft lands in Teach > Suggested answers, where you complete and approve it before it enters your Knowledge Base; nothing is published automatically. You can also mark a question Answered (you covered it yourself) or Ignored (not worth covering); questions still waiting are marked To do.
To keep the list reviewable, DeskClone records at most a few unanswered questions per conversation and pauses recording entirely past 200 that are still to do - the page tells you when that happens and recording resumes as you answer or ignore them.
For teams that want to go deeper, DeskClone also records which tools your agent used, how confident it was, why conversations were handed to a person, and any guardrail activity - useful once you have real traffic and want to understand how your agent reaches its answers.
The Improvement Loop
Quality grows with a simple weekly habit:
- Skim conversations and the ones under Conversations to review for patterns
- Answer the top unanswered questions
- Tweak the personality if the tone is off
The first few weeks move the needle the most. Keep at it and your agent gets noticeably better at your specific visitors and topics.
You don't even have to remember to check: on paid plans, DeskClone emails you a weekly business summary of these numbers - volume, resolutions vs handovers to a person, and the trend against the week before (Pro and above add multi-week trends, open unanswered questions, and a "what customers asked about" topics list). Turn it on or off under Settings > Emails from DeskClone (Workspace Settings).
From the Chat
Jarvis can review conversations for you. Ask it for the flagged ones, those with a given outcome, or those in a date range. It answers with a count, at most 20, and a link to Conversations; it can also read one in full. What visitors wrote is shown as quoted text, never taken as instructions. Its own chats with you are not listed here, as on the page.
It can flag a conversation, add your note, or rate it 1 to 5, as you can on the page. Asking for "no rating" or "no note" clears it. Each change is in Jarvis's list of changes with its own Undo, which puts back each field that has not changed since.
If a visitor asks you to erase their data, Jarvis can do it from a card. See Erasing a visitor's data.
What's Next
To invite team members and set who can see and do what, head to Team and Permissions.
Was this page helpful?