aiResolutionRate in the Get agent analytics API. This guide walks through both, then shows how to break the number down and compare it fairly before and after a change.
If you want the definitions first, read Resolution vs deflection. The short version: Deflection rate is 100% - Human escalation rate and counts Waiting for Customer conversations as deflected, so it is always at least as high as AI resolution rate.
How do we measure AI support agent effectiveness accurately?
Accuracy starts with the definition. Deflection counts any conversation that didn’t reach a person, including ones where the customer gave up, so it overstates what the agent solved. Resolution counts only conversations the agent fully solved. For billing, Fini counts a resolution only when Fini fully solves the issue, no human ever touches the conversation, and the customer doesn’t reply for 72 hours. Then make the number trustworthy:- Measure each agent and channel separately, over full weeks.
- Read resolution next to human escalation rate, escalation reasons and CSAT, so a gain that hurts quality shows up.
- Compare equal windows before and after a change, and let waiting conversations settle.
- Open a sample of the conversations behind any number you report.
Before you start
- At least one agent deployed and handling real conversations. Analytics reads from live conversations, so a new agent needs a few days of traffic before the numbers mean much.
- Access to Analytics in the dashboard.
- For the API option: a workspace API key with
readscope.
Read it in Analytics
1
Pick the agent
Open Analytics and use the agent picker beside the page title. Every chart and table on the page scopes to that agent. In a multi-agent workspace, measure each agent separately.
2
Set the date range
The range defaults to the last 7 days. For a reporting number, use a full week or a full month so weekday and weekend traffic are both represented. The range applies to every KPI card, chart, and table below the header.
3
Set your filters
Decide what slice you are reporting on before you read any number. The pinned filters are Knowledge, Conversation status, Tags, Intent rule, Source, CSAT, and AI CSAT (when available). Channel, Sentiment, and Escalation reason live under + Filter.For an overall resolution rate, leave Conversation status unfiltered. Filtering it to one status makes that status 100% of the slice.Save a setup you will reuse with Views. A view stores filters only and keeps the current date range, so you can apply the same slice to any reporting window.
4
Read the KPI cards
The five cards show Conversation volume, Deflection rate, Human escalation rate, Average response time, and Average CSAT, each with a pill comparing the selected window to the preceding window of the same length.Note Conversation volume first. A rate built on a handful of conversations moves a lot from week to week.
5
Read the Conversation Status doughnut
Scroll to Distribution. The Conversation Status doughnut splits volume into Resolved by AI, Escalated to Human Team, and Waiting for Customer. The Resolved by AI share is your AI resolution rate for the window and filters you set.Check your reading: Deflection rate on the KPI card should equal the Resolved by AI share plus the Waiting for Customer share.
6
Check the trend
In the first trend chart, switch the toggle to Resolution rate. It plots the daily resolution rate alongside the Waiting for Customer share. Look for step changes and line them up with dates you changed prompts, knowledge, rules, or routing.
7
Find out why conversations escalated
In Escalations, the doughnut shows every conversation in the period, with Not escalated as the largest slice and escalated conversations split by reason (Knowledge, Action, User, and Policy families). Apply the Escalation reason filter to drill into one reason. The escalation reasons table says what to fix for each.
8
Break it down by intent rule and knowledge
Under Breakdowns, Intent rule breakdown shows AI Resolve Rate, Escalated Rate, and Volume for each Rulebook intent rule. Knowledge performance shows the same rates per folder, article, or source group, plus CSAT.Sort by volume and look for high-volume rows with a low AI Resolve Rate. Fix those first: they move the overall rate the most. The arrow on a row opens the conversations behind it in Inbox.
9
Spot-check the conversations
Open ten conversations behind any number you plan to report. Confirm the statuses look right in the conversation header and in Output Tag Selection inside AI Steps. A rate is only as good as the classifications behind it.
Useful slices
Read it through the API
Use the API when you want resolution rate in a BI tool, a weekly report, or an alerting job. Get agent analytics returns the full summary for one agent and window; Get agent analytics section returns one section such assummary.
1
Find the agent ID
Call List agents and copy the
botId of the agent you want. API paths use bot for what the dashboard calls an agent.2
Request the window
Pass
startEpoch and endEpoch as Unix epoch timestamps (the reference examples use milliseconds). Add the same filters you use in the dashboard: source, channel, ruleIds, tagIds, escalationReasonTagIds, and so on.3
Read the summary
summary holds the current window and comparisonSummary the comparison window. The fields you need:Deflection rate is
1 - humanEscalationRate. For daily series, read resolutionRateChartData (daily totals, resolved counts, and AI resolution rate) and conversationVolumeChartData (daily resolved, escalated, and waiting counts). For per-rule numbers, read ruleAnalytics.Compare before and after a change
A before-and-after comparison is only fair if the two windows differ in the change you made and nothing else.1
Write down the change and its time
Record exactly what you published (a prompt edit, new articles, a new intent rule, a routing change) and when. If you rolled out by brand or channel, note which ones.
2
Pick equal windows on either side
Use windows of the same length and, ideally, the same weekdays, for example the 14 days before and the 14 days after. The KPI change pills compare against the immediately preceding window of the same length, so setting the range to the “after” window gives you the comparison automatically.
3
Hold the slice constant
Apply the same saved View to both windows. If the change only affects one intent rule or one channel, filter to that slice so unrelated traffic doesn’t dilute the result.
4
Let waiting conversations settle
Conversations from the last day or two of the “after” window may still be Waiting for Customer. Give the window a few days after it closes before you read the final number, and treat early readings as provisional.
5
Compare resolution, escalation, and quality together
Compare AI resolution rate, Human escalation rate, the Waiting for Customer share, Average response time, and Average CSAT. A resolution gain that comes with a CSAT drop or a growing waiting share is not a clean win. Check the escalation reasons that shrank to confirm the change did what you intended.
6
Confirm with conversations and tests
Open a sample of conversations from the “after” window in Inbox and run the relevant Test Suite collection. Analytics tells you the rate moved; Inbox and Test Suite tell you why.
Watch for changes in traffic mix. A product launch, an incident, or a marketing campaign can shift which intents customers bring, and resolution rate moves with the mix. Compare the Intent rule breakdown and Knowledge performance volumes across both windows before crediting a change.
If the rate drops
Work through the same signals in order, from the cheapest check to the deepest: The escalation reasons table lists what to fix for each reason, and Resolution vs deflection explains what a growing waiting share usually means.Related
Resolution vs deflection
The exact definitions and a worked example.
How accuracy is measured
Pair resolution with accuracy signals you control.
Analytics
Full reference for every control, chart, and table.
Get agent analytics
Request parameters and response fields.