Customer support metrics for small teams: the eight worth tracking

Most lists of customer support metrics give you twenty-one of them, one line each, and no formula. That is a glossary, not a plan. If three people answer your customers, you need a handful of numbers you can compute in a spreadsheet on the first Monday of the month, and a clear idea of what to do when one of them moves.

This is that list: eight metrics, the formula for each, whether it counts every conversation or only a sample, and where the number comes from. Four of them already have their own deep guides, so those get a short card and a link. The depth here goes to the ones nobody else covers for small teams: reopen rate, backlog age and contacts per 100 customers, plus a one-page monthly scorecard.

The eight, on one table

"Census" means the metric counts every conversation, so a small team can trust it. "Sample" means it counts only the customers who bothered to answer, which is why it needs more care.

Eight is enough because a three-person team can act on eight. If a metric won't change a decision, it's decoration.

Three rules before you count anything

Use the median for times. A handful of overnight emails drags an average far above what a typical customer experienced, and Front's support KPI guide makes the same case for resolution time (Front). The worked example is in the first response time guide, so we won't rerun it.

Say which clock you used. Business hours and wall-clock time give different numbers for the same conversation. Write the choice at the top of your scorecard and never switch it halfway through a quarter.

Report AI conversations separately. An AI answer isn't a human first response, and satisfaction on conversations the AI closed tells you something different from satisfaction on the ones a person handled. Mix them and both numbers go soft. The AI resolution rate guide explains how to keep them apart.

Volume: conversations per 100 customers

A raw count of conversations rises when you grow, which is good news and tells you nothing. The rate tells you whether each customer needs less help from you than last month.

Define "active customers" once and keep it. Paying accounts that logged in during the month works for a SaaS company; orders in the last 90 days works for a shop. Then:

Conversations per 100 customers = conversations ÷ active customers × 100

Say you handled 340 conversations last month with 1,250 active customers. That's 340 ÷ 1,250 × 100 = 27.2. This month it's 360 conversations and 1,500 customers: 360 ÷ 1,500 × 100 = 24.0. The raw count went up 5.9%, and the rate went down by 3.2 contacts per 100. You didn't get busier per customer; you got better.

The rate also feeds your staffing plan. Multiply it by your expected customer count and you have a conversation forecast, which the support staffing calculator turns into hours. How to find the reasons behind the number, and bring it down, is the subject of how to reduce support tickets.

Speed and promise: first response and SLA attainment

Median first response is the typical wait for a first human reply. SLA attainment is the share of conversations where that wait met your target. Both are covered in full, with the benchmarks and staffing math, in the first response time SLA guide.

One thing the guide doesn't say loudly enough for a metrics hub: attainment is calculated from conversations that have received a reply. A conversation nobody has answered yet isn't in the denominator. A team can post a lovely 95% while one customer waits three days, which is why the backlog check below sits next to it.

Look at the slow tail too. If your tool only reports the median, take a month of waits and look at the slowest tenth by hand.

Finish: resolution time and reopen rate

Median resolution time is created to resolved, middle value. It tells you how long problems take. It doesn't tell you whether they were fixed, and that's what reopen rate is for.

Reopen rate = conversations resolved that came back within 7 days ÷ conversations resolved.

Take a month with 120 resolved conversations, nine of which the customer wrote back about inside a week: 9 ÷ 120 = 7.5%. Its mirror image, the share fixed first time, is 92.5%. Vendors define "first contact resolution" differently, so we use reopen rate and count it yourself the same way every month.

Here is why it matters. Two imaginary teams:

Team A looks faster on the dashboard. But nearly one in five of its "resolved" conversations comes back, so it's closing threads to clear the queue. Team B is slower and its customers get the answer once. Resolution time without reopen rate rewards the wrong behavior.

Three practical details:

• Pick the window (7 days is a sensible start) and keep it. Only count conversations resolved at least that long ago, or the last few days of the month will look better than they are.
• Decide whether a new thread about the same problem counts. We'd count it. In Talvo, a customer's email reply to a resolved thread reopens it. Check what your tool does with a returning chat visitor.
• You'll probably count this one by hand. In Talvo you can add a conversation tag such as reopened when you spot one, then filter the inbox by that tag at month end. Analytics can't report by tag, so the count is a quick tally from the filtered list.

Talvo shows the median resolution time on the analytics page. Reopen rate isn't a report, so the tag-and-tally above is the method.

Backlog: the oldest open conversation

Most small teams look at what's arriving. Few look at what's been sitting there.

Once a day, at the end of your working day, write down two numbers: how many conversations are open, and how old the oldest one is. Count Open and Pending. Leave Snoozed ones out while they're snoozed, but glance at them weekly so a snooze doesn't become a hiding place.

An imaginary week:

Every day the median first response could still look fine, because new chats get quick answers while the old ones age. The count climbing by three a day and the oldest passing your longest SLA target by Thursday are the signals. The rule that works: if the oldest conversation is older than your longest target, someone is named for it first thing tomorrow.

In Talvo you can read this in the inbox by filtering to open conversations, and the analytics page's "By status" breakdown shows how many sit in each status. The no-reply reminder, 12 hours by default under Settings → Inbox → Reply reminders and available on every plan, alerts the team when a customer has been waiting, which is a safety net rather than a substitute for the daily check.

Satisfaction: CSAT, two ways, split by who answered

Report the average rating and the share of 4s and 5s together, with the number of ratings. The two often tell different stories, and a score with no count is an anecdote.

Take ten ratings: 5, 5, 5, 4, 4, 4, 3, 2, 5, 1. They add to 38, so the average is 3.8. Seven of the ten are a 4 or a 5, so 70% are satisfied. Both are true, and different tools headline different ones. Help Scout's Happiness Score, for example, is the percentage of Great ratings minus the percentage of Not Good ratings (Help Scout docs). Never compare your number with a vendor's until you know which convention it uses.

Talvo reports the average out of 5, with the rating count; compute the share of 4s and 5s yourself if you want it. At small counts a single rating moves everything. The CSAT vs NPS guide has the margin-of-error table, including roughly ±16 points at 20 responses, and the survey templates.

Split it by who answered. Talvo's CSAT shows AI-resolved and human-resolved conversations separately on every plan, which is the check to run before you turn automation up.

Automation: AI resolution rate and hand-off rate

If you run an AI agent, two ratios belong on the scorecard: AI-resolved conversations as a share of all conversations, and handed-off conversations as a share of those the AI engaged. A rising resolution rate with falling AI-side CSAT means the AI is closing things badly, not well. Definitions, a way to audit a sample and the pricing implications are in AI resolution rate, and what a healthy hand-off looks like is in the AI hand-off guide.

What we left off, and why

Average handle time. For chat and email in a tiny team it measures how long you were typing, not how well you helped. NPS. It's a relationship measure for a quarterly survey, not a support number; see CSAT vs NPS. First contact resolution. It's the inverse of reopen rate and vendors disagree on what "one touch" means. Cost per conversation. It's worth knowing once a year, and the support cost calculator does the work.

The monthly scorecard

One page, the same rows every month. Fill "target" only after you have three months of history, because a target you invented in week one is a guess.

1. 1Pull the numbersRead the analytics tiles and write the window they cover beside each number, so next month compares like with like.
2. 2Count the three by handTally reopened conversations, read your backlog log, and divide conversations by active customers.
3. 3Read twenty conversationsPick the longest-waiting, the lowest-rated and a few at random. Numbers say where to look; threads say why.
4. 4Write one sentence per moverFor every metric that moved, one line in the Note column about what you think happened.
5. 5Choose one changeOne thing to try next month, with an owner. More than one and you won't know what worked.

When a number moves, here's where to look first:

What Talvo reports, and what you compute yourself

Talvo's analytics page shows, live today:

• total conversations, average CSAT (out of 5, with the rating count), median first response and median resolution
• SLA attainment as a percentage with "met of responded on time", an estimated deflection, AI resolutions with the number of AI-engaged conversations, and channel mix
• a conversation volume chart and a "By status" breakdown
• a quality-and-automation card with CSAT for AI-resolved and for human-resolved conversations, plus campaign impressions and chats opened

History depends on your plan: 7 days on Free, 90 days on Starter, 1 year on Growth and unlimited on Scale. Whatever window the page shows, write it next to the number on your scorecard. "Estimated deflection" is a dollar estimate based on an assumed cost per ticket, so treat it as an estimate and keep it off a scorecard you intend to trust.

The AI performance panel, the funnel from engaged to resolved to handed off, with ranked hand-off topics, is on the Scale plan ($299 a month, or $249 billed annually). SLA policies are on the Growth plan ($149 a month, or $119 annually); business hours and the post-resolution CSAT prompt are on every plan.

The analytics can't report by tag, so there's no per-topic breakdown in the app. That is one reason the contact-reason tally in the ticket-reduction guide lives in a spreadsheet.

Two more limits worth knowing. Talvo covers live chat and email, so conversations on SMS, WhatsApp or the phone won't appear in any of these numbers. And Talvo is new and in beta: we have no customer benchmarks for any of these metrics and don't publish any. The more useful benchmark for a small team is your own previous month, measured the same way.