Most "customer service KPI" lists hand you 21 metrics in a flat row and leave you to work out which ones matter. That was fine when every ticket went to a person. Now an AI agent handles a slice of the volume, and two things happen at once: the list is missing the metrics that tell you whether the AI is actually working, and half the old ones mean something different than they used to. Here's the grouped version, with formulas, what good tends to look like, and the AI-era metrics the standard guides skip.

What Customer Service Metrics Actually Tell You

Sort your metrics into two kinds before you argue about which to track. Activity metrics tell you how much and how fast: tickets handled, time to first reply. Outcome metrics tell you whether the customer's problem actually got solved. You want both, but when they disagree, outcome wins.

One mistake quietly wrecks more support dashboards than any other: counting a deflected conversation as a resolved one. Deflection means no human touched the ticket. Resolution means the customer's problem is fixed. Those are different things, and an AI agent widens the gap, because it's easy to deflect a question with a confident answer that solved nothing. Hold that line and most of the rest of this follows.

The guide uses four groups: speed, quality and outcome, volume and cost, and the AI-automation metrics most lists forget. Track at least one from each. Reading a single group on its own is how teams end up fast, cheap, and bad at the actual job.

Speed Metrics

Speed metrics are the easiest to measure and the easiest to over-trust. They tell you about the experience of waiting, not the experience of being helped.

First response time. How long until the customer gets a first meaningful reply, not an autoresponder. Measure it as the median time from ticket created to first genuine response. Use the median, not the average, because a handful of overnight tickets will drag the mean somewhere useless. What good looks like depends entirely on channel: live chat and messaging customers expect a reply in under a minute or two, while email buys you hours. The trap is optimizing this alone. A fast first reply that doesn't move the issue forward feels responsive and helps no one.

Average resolution time. How long from open to actually closed. This matters more than first response time, and it's the one leaders most often under-watch. Segment it two ways or it lies to you: by issue type (a password reset should resolve in seconds, a billing dispute in days, and averaging them hides both) and by whether a human or the AI closed it. If your resolution time drops but reopens climb, you sped up closing tickets, not solving problems.

Average handle time. How long an agent actively spends per issue: total handle time divided by number of interactions. It's genuinely useful for staffing and capacity planning. It's genuinely dangerous as a performance target. Push it down and agents rush, close tickets that reopen tomorrow, and the number looks great while the customer experience gets worse. Treat handle time as a capacity input, never a goal on its own.

Quality and Outcome Metrics

These are the metrics that actually track the job: was the problem solved, and how did it feel to get there.

First contact resolution (FCR). The share of issues solved in a single interaction with no follow-up: issues resolved on first contact divided by total issues. It's the workhorse outcome metric because customers feel it directly. Nobody wants a second conversation about the same problem. Set your target from your own baseline rather than a benchmark, and always read it next to the reopen rate, because "closed on first contact" and "solved on first contact" are not the same claim.

Customer satisfaction (CSAT). A post-interaction "how did we do," usually the percentage of positive responses out of all responses. Read it knowing who actually answers: delighted and furious customers respond far more than the indifferent middle, so a 92% CSAT built on a 15% response rate is a weaker signal than it looks. Track the response rate right beside the score, and treat a low response rate as its own warning.

Customer effort score (CES). How hard the customer had to work to get helped, usually captured with a single "how easy was it" question on a 1-to-7 scale. Effort predicts repeat business and loyalty better than delight does for support interactions, which is why it earns its own line rather than being folded into CSAT.

Net promoter score (NPS). A relationship-level "would you recommend us," calculated as the percentage of promoters minus the percentage of detractors. It's a company metric, not a support-interaction one. Useful as context, but don't hang a support team's performance on a number that moves with pricing, product, and everything else.

Reopen and repeat-contact rate. How often a "resolved" ticket comes back, or the same customer returns with the same problem: reopened tickets divided by resolved tickets. This is the lie detector for first contact resolution and for any resolution claim, human or AI. A high resolution number sitting on top of a high reopen rate means you are closing tickets, not solving problems.

Volume and Cost Metrics

Ticket volume and channel mix. The denominator under everything else. Watch the trend and the shift across channels (chat, email, voice, social), not the absolute count on any given day. A volume spike with flat resolution time usually means automation is absorbing it. The same spike with rising resolution time means you're underwater and need to act.

Cost per contact versus cost per resolution. Cost per contact divides total support cost by number of contacts. Cost per resolution divides it by problems actually solved. Once AI carries real volume, cost per resolution is the honest number, because cost per contact quietly rewards deflection whether or not anything got fixed. A quick example: if you handle 10,000 conversations for 50,000 dollars but only 7,000 were actually resolved, your cost per contact looks like 5 dollars while your true cost per resolution is closer to 7. The gap is the work you paid for that didn't land. This matters twice over if your AI customer service software prices per resolution, because then it is also your invoice, and a tool that charges every time it succeeds scales your bill in lockstep with your success. If you are putting a number on the savings, the enterprise ROI math is worth running before you commit.

The AI-Era Metrics Most Guides Skip

Here's what the standard lists leave out. The big ones, Zendesk's 21 KPIs and Intercom's five, cover none of these, and once an AI agent handles any of your support, they matter more than half the classics.

Deflection rate versus resolution rate. Deflection counts conversations that never reached a human. Resolution counts problems actually solved. Track both, and report resolution as the headline. A deflection number on its own tells you what you avoided, not what you delivered, and it is the single easiest metric to fool yourself with.

Automated resolution rate (containment). The share of conversations the AI resolved end to end, with no human involved and no follow-up: AI-resolved conversations divided by total conversations. This is the headline metric of AI support, and the definition is exactly where vendors get slippery. Insist on "resolved," meaning the customer's issue was closed and stayed closed, not "contained," meaning the conversation simply ended without a handoff. A customer who gives up and leaves is contained, not resolved. Pair this number with the reopen rate or it will flatter everyone.

AI CSAT. Satisfaction measured specifically on AI-handled conversations, tracked separately from human-handled ones. Blend them and you hide the truth in both directions: a great human team can mask a weak bot, and a strong bot can carry a struggling queue. If you want this at scale instead of from a thin survey sample, evaluate every AI conversation against your own quality criteria automatically rather than waiting for the small fraction of customers who fill out a survey.

Escalation and handoff rate. How often the AI passes a conversation to a person, and, more importantly, whether it hands off well: with the full context attached so the customer doesn't start over. A low escalation rate looks efficient and can be quietly terrible if those conversations should have reached a human sooner. Watch the escalation rate and the post-escalation CSAT together.

Self-service and knowledge effectiveness. Whether the answer existed and got found. When your AI can't resolve something, the first question is usually whether the knowledge to resolve it was even there. Track the questions that returned no good source, because that list is your content roadmap and your fastest route to a higher resolution rate.

How to Build a Scorecard That Doesn't Lie to You

Pick one metric from each group and read them as a set: a speed metric (resolution time), an outcome metric (first contact or resolution rate), a cost metric (cost per resolution), and, once AI is live, automated resolution rate paired with AI CSAT. Four or five numbers, reviewed together, weekly for the operational ones and monthly for the trend.

You read them together because every metric optimized alone corrodes another. Chase first response time and resolution suffers. Chase deflection and CSAT tanks. Chase handle time and next month's reopen rate climbs. A single-number dashboard is a gaming machine, and the team always works out the game, usually faster than you'd like.

Weight resolution above avoidance. "Did we solve it" is the job. "Did we dodge a human" is a cost lever, not a goal, and confusing the two is how support orgs win their metrics and lose their customers. Underneath all of it sits one practical requirement: you can only manage what you can actually see. That means conversation-level visibility into which decision the agent made, which source it used, and where it resolved versus where it should have escalated. Without that, every number above is a guess with a decimal point on it.

Where to Go Next

If you want to measure resolution instead of just deflection, that starts with seeing what happens inside each conversation. Take a look at how we approach customer support, then pick one metric from each group above and build the smallest scorecard that still tells you the truth.

Customer Service Metrics and KPIs: What to Track in 2026