AI Performance
In the app: AI → Performance
Needs the 'View AI Config' permission.
AI Performance, titled "AI Dashboard" on the page, shows how the AI agent is doing: how many tickets it handled, how much of the replying it did, how often it handed over or was taken over, and what it used along the way. It has six tabs: "Overview", "Escalations", "Manual Takeovers", "Feedback", "Releases" and "Exports".
Store, period and refresh
The controls at the top right apply to the "Overview", "Escalations", "Manual Takeovers" and "Releases" tabs. "Feedback" has its own period buttons and no store filter, and "Exports" always covers the whole organization.
- The store picker appears when you have more than one store. "All stores" shows the stores you have access to together.
- "7d", "30d" and "90d" set the period: that many days back, up to and including today. Days are counted in UTC, not in your own timezone. Charts show one point per day; for "90d" they show one point per week. A custom range shows one point per day up to 14 days, per week up to 90 days and per month beyond that.
- "Custom" opens a start and end date. The range can be at most 366 days and the date fields do not offer a day after today. "Apply" uses it; "Cancel" keeps the current period.
- "Refresh" reloads the overview numbers and the tab you have open. The "Feedback" tab is not reloaded by it. The overview numbers also reload by themselves every 5 minutes; the other tabs do not.
The "Overview", "Feedback" and "Releases" tabs need the 'View AI Config' permission. "Escalations" and "Manual Takeovers" also need 'View Reports', and "Exports" needs 'General Settings'. On "Feedback", the "Error table" also needs 'View QA Reviews'; without it the table stays empty.
Your AI agent isn't set up yet
Instead of the tabs, the page shows "Your AI agent isn't set up yet" when the AI has not run in the selected period and every store in view has the agent switched off and has no active main prompt. The tabs come back as soon as the AI has run in the period, or a store in view has the agent switched on or has active instructions.
"What the AI agent still needs" lists three steps with a link each: "Give this store an AI agent" (choose a configuration), "Switch on your AI agent's instructions" (the instructions must be activated in the configuration the store uses, writing them is not enough) and "Set an AI agent name for this store". The first and third step show "Done" only when every store in view has it. Nothing is switched on by clicking a row. "Build an agent from a description" opens the Builder and "Configure it yourself" opens the AI settings.
Overview numbers
Each tile shows a figure for the selected store and period. Hover over the (i) for a short definition; the definitions below say exactly what is counted.
The ticket figures look at tickets that were created in the period, whatever happened to them later. A ticket created before the period and closed inside it is not in these figures. The example ticket and outbound tickets the customer never answered are left out. Deleted and merged tickets are counted.
- "Closed, AI-Handled": of the tickets created in the period that are closed now, the share where an AI-written reply went out to the customer, nobody pressed close, no person is assigned and the ticket was not handed to a team. A ticket a teammate closed is not counted. The line under the figure gives both counts.
- "Messages by AI": replies written by the AI as a share of all replies sent to customers in the period, by the AI and by teammates. A reply is placed by the day it was written. It includes replies the AI sent itself and drafts a teammate approved and sent. Older replies attributed to the AI are included even when their authorship cannot be verified; the tooltip gives both counts. A reply the chat split into several bubbles counts once. Answers from the chat widget's saved FAQ count in the total but not as AI replies.
- "Drafts Awaiting Approval": AI drafts on open tickets that are waiting for a teammate right now. It is not limited to the period. Drafts on closed or snoozed tickets are not counted.
- "Total Interactions": the number of times the AI ran on a ticket in the period, one per run, whether it replied, stayed silent or failed. It is not a number of conversations. Test runs are not counted: "Try it", the Playground and the "Test" tab of the Builder.
- "Avg Response Time": the average time a customer waited for an AI reply, from the customer's message to the moment the AI's reply was delivered. It includes waiting before the AI started and any delay before sending, so it is longer than the time the AI spent writing. Saved FAQ answers are left out. A burst of customer messages is timed from its first message. Only messages written inside the selected period are considered, so a reply to an earlier customer message can be left out. Replies still waiting for delivery are left out; where no delivery time was recorded, the reply’s creation time is used.
- "Failed Runs": failed AI runs as a share of all AI runs in the period (the "Total Interactions" figure). The figure turns red above 5%. It says nothing about whether a reply was correct.
- "Escalation Rate": the number of tickets the AI handed to a person in the period, divided by the number of tickets created in the period. The two groups are not the same tickets: a ticket created earlier and handed over now counts in the first number only. A ticket that was given back to the AI afterwards is no longer counted as handed over.
- "Confirmed Resolution Rate": closed, AI-handled tickets (as in the first tile) whose latest close was a resolve, as a share of the closed tickets created in the period. A ticket that was reopened since is not counted.
- "Resolution Time": the average time from the creation of the ticket to its close. The switch on the tile chooses between every closed ticket (the default) and only tickets that are resolved now. Your choice is remembered in your browser and is the same one the Reports overview uses.
- "Reopened After AI Resolve", marked "Test metric": of the tickets whose latest resolve came from the AI (the customer confirmed, or did not reply to the AI's follow-up), the share that is no longer resolved, or that needed another reply before it was resolved again.
- "Tickets Touched by AI": tickets where the AI sent a reply, a teammate approved its draft, the AI handed the ticket over or the AI decided to ignore the message, as a share of the tickets created in the period. A draft nobody approved does not count. An approved acknowledgment of a handover does not count by itself. Handovers caused by a technical failure, an exhausted return re-check or a failed refund payout are left out.
"About these numbers" under the tiles shows how many replies were saved FAQ answers. Open it to read what is excluded from the reply counts: internal notes, campaign sends, templated back-office emails, and deleted, merged and example tickets.
Overview charts
Under the tiles, charts show "Closed, AI-Handled", "Messages by AI", "Total Interactions", "Avg Response Time" and "Failed Runs" over time. "Absolute" and "Percentage" on the first two switch between counts and shares. Tickets are placed by the day they were created, replies by the day they were written and AI runs by the day they ran. Dashed vertical markers, labelled with the name of a configuration set, show the days a configuration set was assigned to a store, so you can see whether a switch moved the numbers.
- "AI Outages": the number of outage episodes that started in the period, split into "Recovered" and "Ongoing" by their state now.
- "Active Conversation States": where the AI stands on conversations now: "Idle", "Processing", "Awaiting Approval", "Escalated" and "Error". It follows the store picker but not the period.
- "Specialist agent invocations": tickets the main agent handed to a specialist in the period, against all tickets created in the period. "Absolute" and "Percentage" switch between the count and the share. Filter by specialist and by configuration set. Those filters narrow the handoffs, not the total tickets used for the share. Unanswered outbound tickets are left out of that total; example, deleted and merged tickets are included. Handoffs are placed by when they happened, so they can involve older tickets. The line under the chart also gives the number of handoffs, which is higher when a ticket went to a specialist more than once.
- "Tool Usage": the 20 tools the AI called most in the period, with "Count" (calls), "Avg Time" and "Errors" (failed calls). Handoffs to a specialist are not listed here.
Escalations
Needs the 'View Reports' permission.
An escalation is the AI handing a conversation to your team.
- "Escalation Reasons": the reasons, each with its count and share, and the total for the period. It counts handoffs, so a ticket the AI handed over twice counts twice. Handoffs without a reason are listed as "(uncategorized)".
- The chart "Escalations": tickets handed over per day. It counts a ticket once, on the day of its latest handoff, so its total can be lower than the total next to "Escalation Reasons".
- "Latest Escalations": the most recent handoffs, 10 at a time, with "Ticket" (a link to it), "Store", "Config Set", "Reason", "Escalated" and "Notes". The reason list filters the table, and "Load more" shows older ones.
- "Export CSV" downloads the escalations that match the store, period and reason filter, at most 5,000 rows. It is disabled when nothing matches. The table shows times in your own timezone; the file uses the organization's timezone and says which one.
The reasons themselves are managed under AI → Handoff reasons.
Manual Takeovers
Needs the 'View Reports' permission.
A manual takeover is a teammate taking over a ticket that the AI was working on, through the takeover form in the Inbox.
The tab shows the report only when "Require an explanation on manual takeovers" is on under Settings → Organization. When it is off, the tab says "Manual takeovers reporting is off" with a link to that page. Takeovers are recorded either way; one that was made without a reason is listed as "(no reason recorded)".
- "Manual Takeover Rate": of the tickets created in the period that the AI worked on, the share a teammate took over in the period.
- "Manual Takeovers": the number of takeovers in the period, and the chart of takeovers per day. A ticket that was taken over twice counts twice. The days on this chart follow the organization's timezone.
- "Top Reason": the reason given most often.
- "Takeover Reasons": all reasons with their count and share.
- "Latest Manual Takeovers": the most recent takeovers, 10 at a time, with "Ticket", "Store", "Reason", "Agent", "Taken over" and "Explanation". Filter by reason; "Load more" shows older ones.
- "Takeovers by Agent": the 20 teammates with the most takeovers, each with their number and share. The card is labelled "Admins only", but it is shown to everyone who can open this tab.
Feedback
The "Feedback" tab shows how AI replies were rated. It covers all stores you have access to. The store picker and period at the top of the page do not apply here.
- "7 days", "14 days" and "30 days" set how many days the tab covers, up to and including today. The default is 14 days. Days follow the organization's timezone.
- "AI Response Ratings": rated tickets per day, split into "Good", "Bad" and "Don't know", with a "Bad share" line and a "Past 7 days" summary. A ticket is placed on the day of its latest rating. Click a day to filter the table below to that day.
- "Show only AI QA reviewer scores": when on, only scores that reviewers gave on a ticket's AI QA tab are shown. These reviewers are people on your team. When off, each ticket has one final score: the worst of its takeover rating, its escalation rating and its AI QA score, where "Bad" is worse than "Good" and "Good" is worse than "Don't know". Only ratings first recorded in the chosen days are used. Editing an AI QA score changes its value but keeps its original date for this report.
- "Latest Ratings": the rated tickets, most recent first, 8 at a time, with "Ticket", "Score", "Source", "Rated by", "Rated", "Store" and "What did AI do wrong?". The score list filters the table. Deleted and merged tickets are left out.
- "Error table": the error types AI QA reviewers recorded on tickets, with a description, the last occurrence, the linked tickets and a status ("Occurring", "Fix in progress", "Resolved" or "Reoccurring"). Reviewers add errors from a ticket's AI QA tab. Changing a status needs the 'Manage QA Reviews' permission. Setting an error to resolved asks "Mark error as resolved?" and removes its ticket links from the table; each ticket's own AI QA summary still shows it as resolved. This table is not limited to the chosen days.
Releases
"Release History" lists the prompts that were made active in the selected period, newest first, at most 100. With a store selected it lists the prompts of the configurations that store uses. Each row shows the prompt's name and version, its type, who activated it and when. A green dot marks a version that is active now. Use it to see what changed in the AI's instructions and when.
Exports
Needs the 'General Settings' permission.
"Daily AI KPI's" writes one row per day to a Google Sheet.
- "Write AI KPI's to Google Sheets" switches the export on. It can only be switched on when a Google Sheet you can edit is connected.
- "Destination spreadsheet" chooses the spreadsheet. Squid creates a worksheet called "Daily AI KPI's" in it. "Manage Google Sheets" opens Settings → Integrations.
- "Save" stores the switch and the spreadsheet. "Sync now" writes the missing days at once; save your changes first.
- "Last synced" shows when a sync last finished, also when it had no new rows to write.
Each row holds the date, the tickets closed and AI-handled as a percentage and a count, the total tickets closed, the messages by AI as a percentage and a count, and the total outbound messages. Rows cover complete days only, up to yesterday in UTC. Every sync, by hand or the automatic one that runs once a day, adds the days of the last 30 that are not in the worksheet yet. A closed ticket is counted on the day it was last updated, so a day's row can differ from the dashboard chart, which groups tickets by the day they were created. The export covers the whole organization.