Skip to content

Intelligence ​

Intelligence is what your conversations say about your product: the bugs customers hit, the customers at risk of leaving, the gaps in your docs, and what people keep asking for. The analyses that look for problems keep a list of what they find. Each issue becomes a finding: the problem in plain words, the conversations behind it, how severe it looks, and whether anyone has decided what to do about it. When a later run reports the same issue, Overdesk adds the new conversations to the finding it already has instead of listing it again. A run recognises the issue by its conversations, when they largely match those of the finding's latest report and the two names share some words, or by a name close to the finding's. Sharing a conversation alone is not enough, because one conversation can raise two problems. An issue a later run reports under a new name, with mostly different conversations, is listed as new, and you can join the two (see below).

One entry per issue ​

Several analyses can report one problem. A bug after a release can come from Post-release regressions, from Recurring bugs, and from Frustration signals as the anger it caused. The Intelligence page shows those findings as one issue when they come from different analyses, share at least half the conversations they cite between them, and name the problem in the same words. Sharing conversations alone is not enough, because one conversation can raise two problems.

  • The issue is named after the report of the problem itself, so a regression names it rather than the frustration it caused, unless the frustration was rated more severe.
  • It is as severe as its most severe report, and it counts every conversation its reports cite.
  • Reported by lists each finding in the issue, with the analysis that reported it, its own summary and its conversations, and says how many times when one analysis reported the issue more than once. Nothing is merged or deleted: each finding keeps its own history.
  • On the same conversations links to other issues that share half their conversations with this one but name something else. They stay separate entries.

When the page shows one issue as two entries, or two issues as one, anyone who can decide on an issue can put it right:

  • Same issue as another entry?, in an issue, joins it with the entry you choose, each holding the reports the page shows under the status you are viewing. From then on they are one entry, with every report and conversation they hold. A later report under a new name is a new entry, which you can join too.
  • Not the same issue, under one of an entry's reports, moves that report into an entry of its own, and it stays apart however many conversations the two share later.

Each change records who made it, and one undoes the other.

Where findings appear ​

  • Analytics has a What needs a decision panel: the 8 most severe issues nobody has closed yet, one row per issue, with a line counting how many issues there are, how many are new this week and how many are getting worse. See Analytics.
  • Intelligence, under Overview in the sidebar, lists all of them, Critical and High first and the rest under Lower priority. Filter by status, or by what the issue is about: Bugs & regressions, Customers at risk, Knowledge gaps or Requests & ideas. Choose an issue to see which analyses reported it and its conversations, each one a link into your help desk. A link to one finding, from an alert, an email or your tracker, opens the issue it belongs to with that finding's report expanded, and links to the old Findings page still work.
  • The Weekly summary, the Weekly insights email and its report open with the same count and the most severe findings, and the Executive summary starts its top issues from them. See Report types.
  • A new High or Critical finding from an analysis that runs every day, or a finding that came back, raises an alert. See Alerts for exactly when.
  • Ask Overdesk in Slack can list them. See Ask Overdesk.
  • The finding webhook sends each finding to your own endpoint as it opens, comes back or changes status, for your tooling or an AI agent to act on. See Developers.

Findings come from Post-release regressions, Recurring bugs, Frustration signals, Churn signals, Documentation gaps, Most repeated questions, Feature requests, Automation candidates and Out-of-date docs. Recurring themes and Per-product breakdown keep none. One more finding comes from the mail analyses set aside, described under Mail that is not support. See Analyses.

What a finding shows ​

  • Severity: Critical, High, Medium or Low, as the analysis rated it. While a finding is open, a later run that rates it more severe moves it up, and a milder rating never moves it down.
  • New marks a finding first reported in the last 7 days.
  • Getting worse marks an older finding whose new conversations this week outnumber last week's by at least 2, and by at least a quarter. A one-conversation wobble is not a trend.
  • Which analysis found it, when it was first and last seen, and how many runs have reported it.

An issue shows New only when every finding in it is new, and Getting worse when any of them is.

A finding keeps the name it was first given, so it stays recognisable while each run describes it in its own words. Its summary follows the latest run.

Confirmed defects ​

A Product investigation can open a finding of its own, a confirmed defect. It opens one when an investigation has confirmed a bug, found the last version that works and the first that breaks, and read the cause in the code. Overdesk then raises an alert called "Confirmed defect" with the finding's name. The alert gives the versions and where the cause is, says whether a tested workaround or a suggested patch is on the finding, and counts the conversations that are confirmed and the ones that are only suspected. It links to the finding.

The finding shows a Diagnosis above who reported it. A part with nothing to say is left out.

  • Affected versions: the last version that works and the first that breaks.
  • Cause: the file and line the defect is in.
  • Workaround: the code, collapsed until you open it, then the versions it was tested on and when to remove it.
  • Suggested patch and Reproduce: each is code, collapsed until you open it.
  • Impact estimate: the investigation's own estimate of who is affected. It is an estimate, never a count of conversations.
  • Proposed reply: a reply written for the customers the defect affects, with a Copy button.

Every conversation on a confirmed defect is marked confirmed or suspected, and every count says which. A conversation is confirmed when the investigation reproduced the defect on it, or when the customer said the fix worked. It is suspected when it looks like the same defect and nothing more. Overdesk keeps matching newer conversations to the defect and marks each one the same way. The finding says how many are confirmed and how many are suspected, and never gives a bare total.

The proposed reply is for your team to copy and send from your help desk. It is shown exactly as it was written and is not attached to a draft. It may hold a placeholder in double square brackets where a customer's own detail belongs, so fill in every placeholder before you send it. An issue opened from a confirmed defect carries the same sections, without the reply. It is written from the diagnosis, so no investigation brief is offered for a confirmed defect. See Opening an issue from a finding.

What customers at risk hold ​

A finding from Churn signals or Frustration signals names customers who may leave. When you have connected Freemius or Stripe, Overdesk looks those customers up there after the analysis has sent its report, the same lookup a draft makes, and the finding then shows What these customers hold:

  • One line on what is at stake, counted per customer rather than per conversation: how many were checked, how many hold a license or subscription, how many licenses expire within 30 days, how many subscriptions renew within 30 days or are set to cancel, how many hold a lifetime license, how many have lapsed, and what they have paid in total.
  • Under it, each conversation's customer: their licenses and subscriptions, when each expires, renews or ends, and what they have paid through each product or account.

A few things it is careful about:

  • A Freemius license shows when it expires. Freemius has no renewal date, so a license never shows as renewing.
  • The total is shown only when it is known in full for every customer, in one currency. When two Freemius products report the same total for one customer, it may be one payment listed under both, so the total is left out.
  • Could not check names a product that could not be read at the time. It never counts as the customer holding nothing there.
  • No licence or subscription under this email means every product answered and none knew that address. A purchase made under another address is invisible to the lookup.

Each run looks up at most 25 customers, starting with those never checked and the most severe findings. The rest are looked up when a later run cites their conversations again, and until then the finding counts them as not checked. A customer checked in the last 6 days is not looked up again, unless you have connected a product since or one of their products could not be read. The lookup starts once the analysis has finished and sent its report, so it never delays or fails the analysis. A customer you have erased is never looked up again. It is part of customer context, on the Pro and Scale plans and the free trial, and it goes nowhere else: the finding webhook, issues opened from a finding, the AI model and your reports do not carry it.

Deciding on an issue ​

Choose an issue and decide what happens to it. The decision applies to each finding in the issue that can take it, and the button says how many that is when it is not all of them. If some could not be updated, Overdesk says how many were, and trying again finishes the rest. Agents, admins and the owner can decide; everyone on the team can read. An API key can read findings but not change them, so every decision carries the name of the person who made it.

  • Acknowledge: you have seen it and someone is on it. It stays on the list.
  • Resolve: it is fixed. It leaves the list, and opens again only if an analysis later finds it in a conversation that started after you resolved it. The conversations you already saw do not reopen it.
  • Dismiss: it is not worth tracking. It leaves the list and stays closed unless an analysis later rates it at least 2 severity levels higher than when you dismissed it, for example Low to High. A Regression, Recurring bug or Unusual pattern rated High or Critical also opens again when an analysis ties it to a conversation the customer wrote in after you dismissed it, and raises a Dismissed finding is back alert (see Alerts). Conversations it had already cited count only if the customer wrote again, and dismissing it again starts from then.
  • Reopen puts a resolved or dismissed finding back on the list.

The finding shows who last changed it and when. When the change says Overdesk, a run reopened it or it was resolved on its own.

Opening an issue from a finding ​

When GitHub, Linear or Jira is connected, an issue has an Engineering card with Open an issue. It files the finding that already has a tracker issue, or else a confirmed defect if the issue has one, or else the one the issue is named after. Overdesk fills in the finding's title, its summary, how severe it is, a link back to the finding, and its conversations, each one a link into your help desk. When one of the issue's reports is a recurring bug, a regression or an unusual pattern, and none of them is a confirmed defect, you can ask for an investigation brief instead, written from that report's latest conversations. You edit it, choose where it goes and confirm, exactly as when you open an issue from a conversation. Each finding can have one issue on each tracker. See Creating issues from a conversation.

Findings Overdesk resolves on its own ​

When an analysis runs and lists what it found, any of its open findings that it has not reported for more than 14 days are resolved. Only a run that listed its findings can do this: a run that failed, or whose list could not be read, says nothing about whether an issue is still there, so it resolves nothing. Acknowledged findings are left for you to close.

Mail that is not support ​

Analyses set aside mail that is not support before they read a window. What your team still reads by hand, delivery failure reports, automatic replies and notices such as a website's new-user sign-ups, becomes one Not support finding: how many there were, of how many conversations, and up to 20 of them to open. It is always Low, so it never outranks a customer's problem.

Mail already in your help desk's spam folder, or from a sender on Excluded senders, is left out of the finding because it is already dealt with. To make the finding go away, filter that mail in your help desk or add its senders to Excluded senders in Settings → Privacy & Data. It then resolves on its own like any other finding.

When nothing needs a decision ​

An empty list is only good news if the analyses looked recently. The What needs a decision panel says nothing needs a decision only when no scheduled analysis is late, and either a scheduled analysis has listed its issues since it was last due or one listed them in the last 48 hours. It then names each analysis with when it last checked. An analysis is late when it was due more than 3 hours ago and has not run since: a weekly analysis that stopped is named even while a daily one ran this morning. After an older check it says nothing needed a decision then, which says nothing about today, and before any analysis has listed its issues it says exactly that. In every case it names each analysis whose last run failed or did not list its issues, because whatever that analysis would have found is missing from the list.

Privacy ​

A finding describes the issue and points to its conversations by number. A finding never quotes a customer, and a name the analysis wrote into it shows as (a customer). An investigation brief, written only when you ask for one, puts what customers reported in Overdesk's own words and quotes only the error messages and log lines they sent. What a churn or frustration finding keeps about its customers' accounts holds no email address, name or license key. Erasing a customer's data removes their conversations from every finding, with what was kept about their account, and a finding left with no conversations is deleted. See Erasure and deletion requests.

Plans ​

The Intelligence page is on every plan. On Assist, findings come from Documentation gaps, Most repeated questions and Automation candidates, the analyses Assist includes; Trial, Pro and Scale get them from every analysis that keeps findings.