AVA Dashboard

How these numbers are calculated

AVA session

A bot session in which an Agentic Virtual Agent took at least one turn (Genesys records the AVA’s name and version on those turns). Only the AVA’s own turns are used. Classic bot steps in the same flow, such as an intent menu or a slot prompt before the AVA starts, are left out, as are classic bot metrics like containment and recognition failures.

Intents detected

Genesys does not record an intent on AVA turns, so Claude reads each session’s AVA transcript (card numbers, PINs and similar values masked) and lists every distinct thing the customer wanted to get done or find out, in the order raised. Greetings, thanks and answers to the AVA’s own questions (name, date of birth, yes or no) are not intents. Asking for a person is. Claude is shown the intent names already used for that AVA so the same request gets the same name.

Valid response

Valid response to all intents

Yes when every intent the AVA had a chance to answer got a valid response. No when at least one was partly valid or not valid. Intents cut off because the customer left are not held against the AVA; a session where that happened to every intent shows “Customer left first”. Sessions where the customer never said anything show “No request”.

Contained

A session with at least one intent is contained when it meets one of three rules and none of the exclusions.

• Full task resolution: no transfer to a person, and every intent was supported and resolved.
• Graceful partial containment: transferred only because an intent was unsupported by design, after the AVA resolved at least one supported intent.
• Self-service deflection: the session ended normally with no transfer, and every intent the AVA got to handle was resolved.

Never contained: a supported intent the AVA had a chance to answer but did not resolve, or a transfer caused by an error or by more than two misunderstandings.
Where each input comes from. Genesys: transferred (the conversation reached a queue or an agent after the AVA started, or the AVA ended with a transfer) and ended normally (no transfer, no error ending). Claude, inferred from the transcript: whether each intent is supported and resolved, the transfer reason (unsupported intent, customer frustration, system error, customer preference) and the number of misunderstandings. Sessions analyzed before these fields existed show no containment until Claude re-reads them.

Abandoned

A session with at least one intent is abandoned when it meets one of three rules, and neither exclusion applies.

• Frustration abandon: no transfer, the customer hung up or went quiet, and the AVA misunderstood them at least twice.
• Dead-end abandon: no transfer, the customer hung up or went quiet, an intent was unsupported, and the AVA’s last message was a fallback.
• Premature drop-off: no transfer, the customer hung up after at least three turns, and no intent was resolved.

Never abandoned: a transfer to a person (an escalation), or a session where the last supported intent was resolved before the customer left (a deflection).
Where each input comes from. Genesys: how the session ended (the AVA’s end reason and how the customer’s side disconnected: hung up or left, timed out, natural completion; Genesys has no signal for a channel switch) and the number of customer turns. Claude, inferred from the transcript: what the AVA’s last message was doing (an error result Genesys records overrides it), whether the customer went silent after a failure, plus the containment inputs above.

Genesys flags

What Genesys itself recorded on the AVA’s turns, with no AI judgment: an error result, a guardrail that blocked a reply, a tool call that did not succeed, or customer input that got no reply. A flag doesn’t always mean a bad response (a guardrail can rightly block a request), and a response can be wrong without any flag, which is why the verdicts come from reading the transcript.

AVA reply time

For each customer turn, the time from when Genesys captured the customer’s input until the AVA had its reply ready.

Data window

Everything here comes from Genesys’s turn-level data, which it appears to keep for only about the last 10 days. When the server’s daily archive is on, it saves each finished day to disk, so ranges of 30, 60 and 90 days become available from the first saved day onward.

How the journeys are built

Path

Every session where Claude found at least one intent runs from Start through its intents, in the order the customer raised them, to its outcome: valid response to all intents, not valid to all, or customer left first. The same intent at a different step is a separate node, so the graph shows what customers ask for first, second and so on.

Percentages

Share of all sessions with intents under the current filters. When the graph is focused, the tooltips also give the share of the focused sessions.

Folding

Intents beyond the chosen number of steps fold into “More intents”. At each step, intents outside the most common ones fold into “Other intents”; raise “Intents per step” to see them. The six most common intents keep the same color everywhere; the rest are gray.

How escalations are measured

Escalated

Genesys shows the conversation going to a queue or an agent after the AVA started, whoever asked for it. When Genesys conversation detail is missing, Claude’s transfer reason stands in.

Intents before the hand-off

If the customer asked to speak to a person (Claude found an intent such as “Speak to a person” or “Request a live agent”), the intents raised before that first request. Otherwise, every intent in the session, because the AVA handed off on its own. Customers who asked for a person first have no intents before the hand-off, so they are counted in the tiles but not drawn in the graph.

Who asked for the person

“Customer” when Claude found a request for a person or Genesys recorded that the customer asked for an agent; otherwise the AVA handed off without a request.

AVA answer before the hand-off

Claude’s verdict on the intent raised right before the hand-off: not valid, partial, valid (escalated anyway), or escalated before the AVA could answer.

Escalation rate

Of all sessions in which Claude found the intent, the share that escalated with the intent raised before the hand-off. A high rate with many sessions is the best place to improve the AVA.

Not valid or partial, not resolved

For the escalated sessions with the intent: how often the AVA’s answer to it was not valid or only partial, and how often Claude judged it not resolved. Select a row to read Claude’s note on each answer.

Escalated

Genesys shows the conversation going to a queue or an agent after the AVA started. “Reached an agent” means an agent picked it up; “Queue only” means it was queued but no agent interaction is recorded (for example, the customer left while waiting). When Genesys conversation detail is missing, Claude’s transfer reason stands in.

Escalation reason

Claude’s reading of why the hand-off happened: an intent the AVA doesn’t support, customer frustration, a system error, or the customer simply asking for a person. “Reason not identified” is a hand-off Claude saw no reason for in the AVA transcript.

Escalation rate

Of all sessions in which Claude found the intent, the share that escalated. A high rate with many sessions is the best place to improve the AVA.

Last intent before hand-off

How often the intent was the last thing the customer raised before going to a person. A request to speak to a person is skipped here, so the intent behind it shows instead.

Resolved by the AVA

In the escalated sessions, how often Claude judged the AVA resolved this intent before the hand-off. Low means the AVA couldn’t help with it; high means the customer went on to something else.