Definition
True Resolution Rate (TRR) measures the percentage of AI-handled service desk tickets where the user's problem was actually solved, not tickets that auto-closed due to timeout, not tickets the user abandoned, not tickets that prompted a follow-up contact within the recurrence window. TRR is the denominator correction for FCR in an AI-augmented service desk environment. TRR is Pilot-Ready. Every data element is in your ITSM system today. The work is the recurrence query and the cross-channel matching logic, not new instrumentation. In Plain Language of all the tickets the AI chatbot marked resolved, count how many did not result in a reopen, a duplicate ticket on the same topic, a channel switch to a human agent, or an auto-close without user confirmation: within 14 days. That's your true resolution count. Divide by total AI-resolved tickets.
The Measurement Gap
The service desk metrics looked excellent. Ticket deflection rate: 68%. First contact resolution (FCR): 71%. Average handle time for human agents down 34% year-over-year. The AI chatbot was, by every available measure, transforming the service desk operation.
Then someone ran a 90-day recurrence analysis.
Of the tickets the AI chatbot marked as "resolved," 38% had at least one of these follow-on events within 14 days: the user reopened the ticket, the user opened a new ticket on the same topic, the user called the helpdesk directly, or the ticket auto-closed with no user confirmation after the system timeout. The deflection metric was, in practice, 38% theater: users who gave up, worked around the problem, or came back through a different channel.
FCR counted every chatbot closure as a first-contact resolution. The metric had no mechanism to look forward in time and check whether the resolution held. A ticket that auto-closes after the user doesn't respond within 48 hours is indistinguishable from a ticket the user actually confirmed resolved: both increment the FCR counter identically.
The service desk manager had suspected the numbers were inflated. But there was no metric to prove it, so the deflection rate stood unchallenged in quarterly reviews, and the AI chatbot's scope expanded to more ticket categories.
This failure mode is structural, not accidental. AI service desk systems are optimized for deflection: removing tickets from human queues. Deflection is easy to measure: did a human agent handle it or not? Resolution is hard to measure: did the user's problem actually get solved? The systems get rewarded on the metric they're optimized for, and the metric they're not optimized for goes unmeasured.
Legacy metrics built for human-staffed service desks don't account for a system that can close tickets without user confirmation, can respond to every message instantly, and can generate plausible-sounding responses that users accept without verification. FCR in the pre-AI context assumed a human interaction where the user had some ability to evaluate the quality of the answer before accepting it. AI chatbots can close tickets far faster than a user can evaluate whether the answer was right, and in many platforms, they'll auto-close if the user doesn't respond to dispute it.
The deflection numbers say 68% of tickets never reached a human. What they don't say is what fraction of those users actually had their problem solved versus found a workaround, gave up, or quietly reopened the problem through a different channel.
The measurement gap: there is no metric that tells you whether AI-deflected service desk tickets were actually resolved or merely abandoned, which means deflection rates overstate actual resolution quality by an unknown factor, a factor this KPI is designed to measure.
Formula
True Resolution Rate (TRR) measures the percentage of AI-handled service desk tickets where the user's problem was actually solved, not tickets that auto-closed due to timeout, not tickets the user abandoned, not tickets that prompted a follow-up contact within the recurrence window. TRR is the denominator correction for FCR in an AI-augmented service desk environment. TRR is Pilot-Ready. Every data element is in your ITSM system today. The work is the recurrence query and the cross-channel matching logic, not new instrumentation.
In Plain Language
of all the tickets the AI chatbot marked resolved, count how many did not result in a reopen, a duplicate ticket on the same topic, a channel switch to a human agent, or an auto-close without user confirmation: within 14 days. That's your true resolution count. Divide by total AI-resolved tickets.
Formula
TRR = (AI_resolved − Reopened − Duplicate_followup − Channel_switch − Auto_closed_no_confirm) / AI_resolved × 100
Where:
- AI_resolved = tickets marked resolved by the AI system without human agent involvement
- Reopened = subset reopened by the user within N days
- Duplicate_followup = new tickets from the same user on the same topic within N days
- Channel_switch = same user contacts a human agent on the same topic within N days (phone, email, walk-up)
- Auto_closed_no_confirm = tickets closed by system timeout without explicit user confirmation of resolution
-
N = recurrence window, recommended starting value: 14 daysWorked ExampleAn enterprise IT service desk uses an AI chatbot to handle password resets, VPN access issues, software installation requests, and common account problems. In Q2, the AI chatbot marked 2,400 tickets as resolved without human agent involvement. Reviewing the 14-day recurrence window:
-
1,380 tickets: no follow-on event. No reopen, no duplicate, no channel switch, no auto-close without confirmation. True resolution.
- 312 tickets: user reopened the ticket within 14 days with "this didn't work" or a clarification request
- 288 tickets: user opened a new ticket on the same topic within 14 days (same user, same category keyword)
- 156 tickets: same user called the helpdesk on the same topic within 14 days (matched via user ID and topic keyword)
- 264 tickets: auto-closed by system timeout with no user response confirming resolution
TRR = 1,380 / 2,400 × 100 = 57.5%
The deflection metric reported 2,400 tickets handled without a human agent. FCR counted these as 2,400 first-contact resolutions. TRR says 1,020 of those (42.5%) were incomplete resolutions that generated follow-on work or left users without a solution. The service desk manager's suspicion was correct; now it's a number.
Sign Convention
Higher TRR is better. The gap between FCR and TRR is the "deflection illusion": the fraction of AI-resolved tickets that FCR counts as successes but that didn't actually solve the user's problem. Tracking this gap over time reveals whether AI chatbot tuning is improving genuine resolution quality or just making the deflection metrics look better.
Edge Cases
User confirms resolution but the fix was a workaround: TRR counts this as resolved, because from the user's perspective, their immediate problem is addressed. If you want to distinguish workarounds from root-cause solutions, track that separately: TRR measures the user outcome, not the solution type. Same user, different topic: topic matching is required to avoid false positives. A user who opens a new ticket on a genuinely different topic within 14 days should not count as a TRR failure. Match on topic category or keyword similarity, not just user identity. Channel-switch detection requires that phone or walk-up interactions are logged in the same ITSM system with user attribution. If they aren't, channel-switch detection becomes "Not Captured", and TRR will overstate resolution quality by missing those failures. Minimum sample: 100 AI-resolved tickets per measurement period. Cross-references: Upstream, TRR is fed by KQDI (Knowledge Quality Decay Index, GMI-IT-008). If the AI is resolving tickets from a knowledge base that contains workaround patches rather than root-cause fixes, TRR will decline as knowledge quality degrades: the AI gives confident answers that don't actually work. Downstream, TRR feeds service desk efficiency metrics, a declining TRR means human agents are absorbing follow-on work that deflection metrics are not attributing to the AI system's incomplete resolutions. Maturity: Pilot-Ready. All data elements are in standard ITSM platforms. The derived elements (duplicate detection, channel-switch matching) require query logic, not new data collection. Most service desk platforms have topic categorization that makes same-topic matching feasible without building a custom NLP classifier.
Proposed formula, pending validation pilots.