Ticket Deflection: What the Number Covers

A published rate cannot, on its own, identify the best AI ticket deflection software. A deflection rate is one event, counted over one population, within one window, and each vendor sets all three. Zendesk sorts its AI agent outcomes into assisted escalations, contained resolutions and verified resolutions. Freshworks counts queries resolved without human help, negative feedback or a ticket. Ravenna classifies tickets as Resolved, Assisted, Escalated or Not Applicable. Serval counts tickets resolved without human intervention. Before setting two rates side by side, align the definitions and the observation windows, then examine the request mix, eligibility and sample size; report eligible requests as a share of all incoming requests; a request too recent for its recontact window to have closed is not yet assessable. Before setting one vendor’s figure beside another’s, ask five things: which event counted as deflected or resolved, out of which requests, over which window, what happens to a request that comes back, and whether the tickets behind the number can be opened.

What does each vendor count as deflected or resolved?

Published definitions differ in the event they count.

Vendor	What the published definition counts
Zendesk, automated resolution tiers (introduced May 18, 2026)	Assisted escalation: the AI agent contributed before a human completed the resolution. Contained resolution: the agent answered and the customer asked for nothing more (no clarification, feedback or human), but a check of the conversation by a large language model at its end did not confirm the resolution. Verified resolution: the same, with the check passed.
Freshworks, Freddy AI Agent (Classic) overview report, modified April 7, 2025	Ticket Deflection Rate: the percentage of queries or support requests resolved by Freddy AI Agent without human help, negative feedback, or the employee creating a ticket. The page is filed under the Classic agent and does not say whether the definition applies outside it.
Ravenna, AI Outcome	Per ticket: Resolved (the agent handled the ticket end-to-end), Assisted (a human used or built on the agent’s work), Escalated (the agent’s output was unusable), or Not Applicable (spam, bounced email, monitoring alerts).
Serval, analytics	AI resolved: tickets resolved without human intervention. Separate categories for AI assisted (Serval ran workflows, a human resolved the ticket), Unassisted, and Resolved outside Serval.
Atomicwork, in its Ammex Corp customer story	Deflection: “the percentage of conversations where the Atom could answer without involving a human agent or creating a ticket”.
Atlassian, AI for service management	No formula on the page; Rovo is described as able to “deflect a meaningful share of routine requests before they reach an agent”.

The same conversation can count under one definition and not another: at Zendesk, one in which the customer asked for nothing more but which failed the verification is Contained, not Verified, and the page says neither assisted escalations nor contained resolutions count against the resolution allowance. The unit varies too. Freshworks and Atomicwork count queries or conversations, Ravenna and Serval count tickets, and Freshworks defines a conversation as the exchanges within the same context, greetings and acknowledgements excluded.

Four questions apply to any rate a vendor presents.

• 01

  What event was counted

  An answer the employee accepted, a conversation closed without a ticket, a ticket resolved without a human, or an action carried out in another system. Each is a different event, with its own definition.

• 02

  Out of which population

  Users, conversations, tickets or requests. Whether imported, spam or unclassified items sit in the denominator changes the percentage without changing the queue.

• 03

  Over which window

  A verdict taken two hours after the first message and one taken after thirty days do not describe the same outcome. A window that is not stated leaves two figures without a common basis.

• 04

  What the saved work was worth

  A flat per-ticket estimate is a model, not a measurement. Handling time by request type, taken from your own tickets, is the figure a cost argument can rest on.

Out of which population, and over which window?

The denominator moves even inside one product. Ravenna’s agent-impact dashboard counts only classified tickets, leaving out those marked Not Applicable or Not Computed. Ravenna’s “Percent of” widgets, by contrast, divide by the full filtered population unless an “Exclude empty from total” toggle is turned on, and a tooltip such as “Out of 142 tickets” shows the denominator a widget actually used. Serval keeps tickets resolved without its involvement, imported tickets and Silent mode tickets included, in a category of their own.

A customer page can shift units within itself. Freshworks’ Seagate story reports a “Freddy AI Agent ticket deflection rate” of 32%, then writes that the agent deflects 32% “of tickets” in one passage and “of incoming requests” in another.

The window decides what a verdict can see. Zendesk considers a conversation ended 72 hours after the first email, 2 hours after the first message in messaging by default (extendable to 72), and at hangup for voice, and runs its verification then. Ravenna’s Thinkific case study states its figure for “the 30-day reporting window discussed during the interview”. Ravenna’s documentation also notes that classifier updates reclassify previously closed tickets, so historical AI Outcome values can shift: a trend holds only if the definition behind it is dated.

Does answering count as deflecting, and does acting?

Either an answer or an executed action can avoid a ticket. The work saved depends on the request and its confirmed outcome, not on which of the two happened. Report them as two figures: they fail in different ways. An answer can be unusable: Ravenna classifies a ticket as Escalated only with evidence that the agent’s output could not be used, and as Assisted when the record does not show whether a human used it. An action can stop at a person: Atlassian’s Request Resolver carries out a resolution plan across connected tools, after an agent approves each plan in supervised mode, or directly for qualifying requests in autonomous mode. Where the approval sits changes the count. Ravenna does not treat approving an automated step as substantive human involvement; Serval counts a ticket on which it ran workflows but a human resolved it as AI assisted, not AI resolved.

Atomicwork’s Ammex Corp story shows a customer tracking the two separately: deflection, measured over conversations, handled FAQ-driven questions, while what still reached the help desk was “action-oriented”, and the team’s focus “shifted from just deflection to automation”.

Knowledge gaps appear on the answering side: Freshworks reports Unanswered conversations, where the knowledge sources lacked the information or the query was not understood, and the top unanswered topics. Closing them is documentation work, covered in training an agent on internal documentation. Access changes for joiners and leavers, a request type that ends in an action, are covered in automating onboarding and offboarding.

How should a published customer figure be read?

A customer figure is one organization’s account of its own queue, published by the vendor: read it for its unit, population and window, and align definitions and windows before setting it beside another. Two figures that look alike can count different events. Ravenna’s Thinkific case study reports that 88.7% of requests were “triaged or handled by AI” over a 30-day window, and adds: “That figure includes requests escalated to the team after triage.” Moveworks’ Broadcom story reports 88% of support issues resolved autonomously as of 2025, “up from approximately 10% at launch”; it gives a date, not a window. One figure includes escalations, the other does not say.

Siit’s customer story on Monzo, published February 19, 2025, reports “60% of inbound support requests automated” and quotes Monzo’s Tech Ops Support Lead, Ashley Brien: “Siit helps us automate and solve 60% of our inbound support requests.” The story attributes the result to article suggestions drawn from Monzo’s Notion knowledge base. It states no window and no denominator beyond “inbound support requests”. Like every figure in this section, it is one customer’s account reported by the vendor, not evidence that another team would see the same share.

What does a rate leave out?

The cost of what left the queue. A rate gives each request the same weight. In an illustrative case, a password question and a long access request each count once, so a rate can rise while hours spent stay flat. Dashboards that show savings model them: Serval computes time saved as 10 minutes per AI-resolved ticket and 5 minutes per workflow run, at a default labor rate of 100 dollars per hour that the admin can change, and bases these estimates on the typical time human agents need for similar tasks.

Whether the problem stayed solved. A conversation that ended without a ticket can mean an answer was used, an answer was unusable, or the employee stopped asking. Freshworks does not count a query with negative feedback as resolved, and counts a conversation as helpful when the employee clicked thumbs-up or completed a suggested service request form without creating a ticket. Zendesk decides its tiers when the conversation is considered ended. Neither the Freshworks nor the Serval page read for this guide states a rule for a request that comes back after it was counted.

Which indicator can a team define for itself?

A team can fix its own definition, so that two periods, or two tools trialled on the same queue, are measured the same way. This convention is proposed here, not a market standard:

• Formula. Eligible requests resolved without human intervention and without re-contact within a chosen window, divided by eligible requests in the same period.
• Eligibility and exclusions. Decide in advance which requests count, leaving out spam, monitoring alerts, imported tickets and requests the agent is not configured to handle, and say so next to the figure.
• Window. Choose the re-contact window before measuring, in days, and report it with the rate.
• Sample. Read a sample of closed conversations by hand each period, its size set beforehand and reported, to catch unusable answers.
• Abandonment and reopening. Count conversations that ended with no ticket and no confirmation separately, and count reopened tickets against the resolution.
• Effort saved. Take handling time by request type from your own tickets rather than a flat per-ticket estimate, and report it next to the rate.
• Separate figures. Report requests answered from documentation, requests closed by an action in another system, and tickets never created as separate figures, each with its definition.

What such an agent is meant to change in an IT team’s work is set out in what an AI service desk agent is, and is not. A deflection rate describes a queue, not a headcount.

Which vendors publish a deflection or resolution metric definition and reporting period?

Listed here: vendors of an AI service desk agent that publish, on one of their own pages, the definition of their deflection or resolution metric and its reporting period. This does not establish an outcome-verification or recontact window. The list is alphabetical and does not compare the rates.

• Freshworks: its Freddy AI Agent (Classic) overview report defines the Ticket Deflection Rate as the share of queries or support requests “resolved by Freddy AI Agent without human help, negative feedback, or employee creating a ticket”, reviewed “over a selected time period”.
• Ravenna: its analytics documentation defines its Resolved metric as the “percentage of classified tickets the AI agent handled end-to-end”, read over the dashboard’s date range.
• Serval: its analytics documentation counts “tickets resolved by Serval AI without human intervention” and tracks their ratio to total tickets over a time range of one week, one month, three months, all time or a custom period.
• Siit: its Dashboards documentation defines the deflection rate as “questions your agents resolved without reaching an admin”, available “in weekly or monthly views”.
• Zendesk: its AI agents dashboard documentation, for a feature in an early access program, defines the automated resolution rate as “the percentage of customer issues that were fully resolved by an AI agent, without any human agent intervention”, with reports filtered by conversation start date.