Skip to content

Overview ​

Analyses read a batch of your recent conversations and answer one question about them: what keeps coming up, what broke in the last release, what customers are asking you to build, who is about to leave. Each one comes back as a written read you can hand straight to product, documentation or engineering.

Reports tell you how the queue is doing. Analyses tell you what is inside it.

Run an analysis ​

  1. Open Analyses in the sidebar.
  2. Set the Time window: how far back to look, written like 7d, 2w or 3m. The default is 7d, and the window can reach back up to 12 months.
  3. Click Run analysis on the card you want.
  4. The result appears on the Reports page under Recent reports. Pick the analysis from the list above the table to see only its results.

Agents, admins and the owner can run analyses. Everyone on the team can read the results. If the same analysis is already running, Overdesk tells you so rather than starting another, and the result you get is the one already under way, with the window it started with.

What each analysis finds ​

AnalysisWhat it finds
Recurring themesThe 5 topics that came up most in the window, each with an estimated ticket count, what customers are running into, and one fix to consider.
Post-release regressionsNew or worsening problems that point to a recent release, each rated for severity. Says plainly when it finds none.
Documentation gapsTopics customers ask about 3 or more times that point to a missing or unclear page, with a suggested page title for each. It reads your conversations, not your docs, so treat each gap as a lead to check against them.
Most repeated questionsThe 10 questions customers repeat most, written the way they actually ask them, each with a suggested way to answer it once and for all.
Frustration signalsThe angry and disappointed conversations: what set customers off, the quotes that show it, and an overall temperature for the period.
Recurring bugsRepeated defect reports grouped into distinct bugs, with ticket numbers, symptoms and a severity rank. Ignores how-to questions and one-off mistakes.
Feature requestsDuplicate asks merged into single requests and ranked by demand, with the need behind each one rather than just the words customers used.
Churn signalsCancellations, refund requests, competitor mentions and threats to leave. Each one cites the ticket, quotes the signal and suggests a retention action.
Automation candidatesRepetitive, low-judgement tickets that a saved reply, a help article or a self-serve flow could handle, ranked by volume, with the approach to use for each.
Out-of-date docsThe AI drafts your team corrected before sending, read one by one for what changed and why. A corrected fact, step or offer counts, and a reworded reply does not. It names the help article or saved reply that said the wrong thing, with the change to make, and groups the corrections no document covered into pages to add. It lists only what a change to your docs would fix.
Per-product breakdownThe top 3 themes for each of your products, grouping conversations by the product triage gave them. Conversations triage has not labelled yet are counted rather than listed, and a run with none labelled is marked skipped and says why.

How much each run reads ​

Each run reads up to 100 conversations from the window, across every help desk you have connected. Default mailbox ID and Default view ID (Zendesk only) in Settings → General narrow analyses the same way they narrow reports.

Mail that is not support is set aside before the 100 are counted, so an analysis reads your customers rather than your mail servers: anything in your help desk's spam folder, senders on Excluded senders, delivery failure reports, automatic replies, and notices such as a website's new-user sign-ups. Each report says how many it set aside.

Analyses that start within 20 minutes of each other over the same window share one read of your help desk, so a schedule that runs several at once asks your help desk once. A run inside those 20 minutes does not see a conversation that arrived after the first one read the window.

Post-release regressions, Frustration signals, Recurring bugs and Churn signals read each conversation's full message thread, except the notes Overdesk posted and the replies it sent itself. The others work from each conversation's subject and preview, apart from Out-of-date docs. A conversation where Overdesk sent a reply itself is read without its preview, because the preview can be that reply.

Out-of-date docs reads the drafts your team changed before sending instead: the conversations graded ai-edited, and those graded ai-bad where your team wrote a reply of its own (see How Overdesk grades its drafts). It reads up to 60 from the window, newest first, and the window counts back from when each draft was graded, which happens after the conversation closes. Each draft is read with your team's reply, the customer's message it answered, what was said earlier in the conversation, and the help articles and saved replies it was written from. Default mailbox ID narrows it and mail that is not support is left out, as for every analysis; Default view ID (Zendesk only) does not narrow it.

Where results go ​

Every finished analysis is saved alongside your reports and kept for as long as your plan keeps reports. It is also sent to each notification destination that has Analyses switched on in Settings → Notifications. See Notifications.

Findings ​

Every analysis except Recurring themes and Per-product breakdown also keeps each problem it finds as a finding, which your team can acknowledge, resolve or dismiss. A later run that reports the same problem adds to the finding rather than repeating it. The most severe open ones appear on Analytics, and all of them on the Intelligence page, one entry per issue however many analyses reported it. The mail set aside as not support keeps one finding of its own. See Intelligence.

Analyses on a schedule ​

Overdesk schedules the analyses for you once your Anthropic API key is saved. Post-release regressions, Frustration signals and Churn signals run every day at 06:10 over the last 3 days, and the rest run on Mondays at 06:20 over the last 7 days, in your workspace time zone. Each waits for its first scheduled time rather than running when it is created. See What runs automatically.

Out-of-date docs is the one analysis Overdesk does not schedule for you. Run it from Analyses, or add a schedule for it. A Time window of 30d gives it a month of corrections to read.

You can pause any of them, move them to another time, or give one a different Time window in Schedules: click New schedule, pick the analysis under its own heading in What should run, and set a Time window that matches the cadence. A monthly schedule with a one-week window reads only the last week.

Plans and cost ​

Trial, Pro and Scale include every analysis. Assist includes Documentation gaps, Most repeated questions and Automation candidates; on the Analyses page the others carry a Pro tag instead of Run analysis. A scheduled run the plan no longer includes, for example after moving to Assist, is recorded as skipped with the reason.

Analyses run on your own Anthropic API key, so an analysis started without a saved key fails and raises an alert. Out-of-date docs makes one AI call for each corrected draft it reads, then one for its report. See Add your Anthropic API key and Plans and pricing.