Support metrics have a reputation problem. Teams track a lot of them, but VP-level conversations tend to orbit around five: first response time, CSAT, resolution rate, average handle time, and escalation rate. Everything else is operational detail. These five are the ones that show up in board slides, in quarterly reviews, and in the conversations that happen when the support function is under scrutiny.
What's changed over the past few years is that AI resolution has started to affect each of these metrics in ways that aren't always intuitive. Some of the effects are positive and obvious. Others are counterintuitive — and if you don't understand the mechanism, you can misread what your dashboard is actually telling you.
First Response Time (FRT): the easy win, and its trap
FRT is the number AI ticketing discussions always start with, because it's the most visually dramatic effect. An AI agent that resolves a WISMO ticket in 20 seconds produces an FRT that makes the whole queue look fast. Blended FRT — across AI-resolved and human-handled tickets together — drops significantly. You see it in dashboards immediately after deployment.
The trap is treating blended FRT as the whole story. When you segment FRT into AI-resolved tickets and human-handled tickets separately, you often find that human-handled FRT hasn't changed — or has gotten slightly worse, because the AI is absorbing the easy tickets and leaving agents with a denser queue of complex, time-consuming issues. A VP who sees overall FRT improve 40% and concludes "support is operating better" may be missing that agents are actually under more pressure on the work that remains.
Track segmented FRT from day one. AI-resolved FRT tells you whether the auto-resolution system is working. Human-handled FRT tells you whether your agents are coping with what's left. The goal is for both to improve — or at minimum, for human-handled FRT not to degrade as ticket mix shifts toward complexity.
CSAT: the metric that punishes bad handoffs most severely
CSAT on AI-resolved tickets tends to run high for simple transactions — order status, password reset, refund confirmation — when the resolution is fast and accurate. Customers who got their answer in 15 seconds don't have much to complain about. Industry-realistic CSAT for well-configured AI resolution on tier-1 tickets sits in the 4.4–4.8 range out of 5.
The CSAT risk in an AI-integrated support stack is concentrated at the handoff: tickets that the AI attempted but couldn't resolve, transferred to a human agent without adequate context. If an agent picks up a ticket and has to ask the customer to re-explain their issue — because the AI's summary was incomplete or the conversation transcript wasn't forwarded — CSAT drops sharply. Customers who've already explained themselves once and have to do it again are measurably more frustrated than customers who went straight to a human.
This means CSAT measurement needs a new segment: AI-attempted-then-escalated tickets. These are the highest-risk group in your stack. Track them separately and watch them closely after any change to your AI resolution logic or confidence thresholds. If you tighten your thresholds (meaning more tickets escalate instead of being auto-resolved), your escalation CSAT should improve because better-prepared handoffs carry more context. If it doesn't, the problem is in your handoff design, not your threshold.
Resolution Rate: the denominator problem
Resolution rate — closed tickets divided by total inbound tickets in a period — looks better with AI, but the denominator needs scrutiny. If your AI is resolving tickets that shouldn't be resolved (low-confidence classifications that happened to produce a reply the customer didn't immediately reject), you're inflating resolved ticket counts while setting up future reopens.
The more useful metric is first-contact resolution rate (FCR) on human-handled tickets, combined with reopen rate on AI-resolved tickets. FCR on complex tickets tells you whether your agents are actually solving problems or just passing them along. Reopen rate on AI-resolved tickets tells you whether the AI is closing tickets because the problem was actually fixed or because the customer gave up temporarily.
A healthy reopen rate on AI-resolved tickets is roughly comparable to or lower than reopen rate on equivalent human-handled tickets. If AI-resolved tickets are reopening at 2–3x the rate of human-resolved tickets in the same category, your confidence thresholds are too aggressive — the AI is resolving tickets it's not actually equipped to handle.
Average Handle Time (AHT): the composition shift
AHT typically rises when you introduce AI resolution, even when the support operation is improving. This is the composition shift: AI absorbs the 3-minute tickets, leaving agents with the 12-minute ones. Your average goes up not because agents are slower, but because the distribution of what they're handling has shifted toward complexity.
We're not saying rising AHT after AI deployment is a problem — we're saying it's often evidence that the system is working correctly. If easy tickets are no longer in the agent queue, agent AHT should rise. The right frame for AHT in an AI-integrated stack is: are agents resolving the complex tickets they're now receiving efficiently, and is the overall cost-per-resolution improving even as per-ticket handle time goes up?
A consumer app support team running 800 agent-handled tickets per day before AI deployment, averaging 6 minutes AHT, might shift to 300 agent-handled tickets per day at 10 minutes AHT after deployment — with the other 500 tickets auto-resolved in under 30 seconds. Total agent-minutes drop from 4,800 to 3,000. AHT went up. Cost went down. Both facts are true simultaneously.
Escalation Rate: the design signal nobody watches closely enough
Escalation rate — the percentage of tickets that require routing from tier-1 to a specialist or higher-level agent — is the metric that most directly reflects the quality of your triage logic and your AI's confidence calibration. Too low an escalation rate suggests the AI is holding tickets it should be passing up. Too high suggests your tier-1 resolution capability, human or AI, isn't sufficient for your ticket mix.
What's underappreciated about escalation rate is how it maps to agent morale. Agents who receive escalated tickets without context — just a frustrated customer and a thread with no summary — are spending their highest-effort time on the worst-prepared handoffs. Over time, this shows up in attrition and performance degradation. Escalation rate is technically a routing metric, but it's also a leading indicator of agent experience quality.
After deploying AI triage, track escalation rate by original ticket source: AI-initiated vs. human-initiated (agent-opened) conversations. AI-initiated escalations should carry a context card that includes conversation summary, customer history, and a recommended action. If your AI-initiated escalation rate is acceptable but agent experience on those tickets is still poor, the problem is in context quality, not volume.
Reading these metrics together
Each of these five metrics tells you something, but the real signal comes from reading them together. A support operation that's genuinely improving after AI deployment should show: FRT improving across both AI-resolved and human-handled segments, CSAT holding or improving (with escalation-segment CSAT specifically tracked), resolution rate holding or improving alongside reopen-rate stability, AHT rising modestly but total agent cost per resolution falling, and escalation rate stable or declining as AI classification improves over time.
If you're seeing FRT improvement but CSAT decline and AHT spike simultaneously, that's usually a sign that AI is resolving tickets at high volume but with poor handoff quality — agents are handling escalations blind, taking longer, and customers are rating the overall experience poorly despite the fast initial reply. The fix is in the handoff design, not the classifier.
Metrics in isolation mislead. Metrics in combination reveal the actual shape of what's happening in your support operation.