Customer Service KPIs for Small Shops in 2026: 8 Metrics, Their Formulas and a One-⁠Page Weekly Review

First, the honest part: there are no universal benchmarks

You have probably seen claims such as "first response time should be under so many minutes" or "resolution rate should reach such-⁠and-⁠such percent". We quote none of them, for three reasons.

• Industries differ. A cake shop's questions can mostly be answered in one line; a renovation firm has to estimate every job. Comparing the two means nothing.
• Definitions differ. Under the same name "resolution rate", some people divide by all conversations and others only by the conversations the AI handled.
• Sources are unclear. Many figures cannot be traced to an original survey, and you cannot tell whether shops like yours were in the sample.

So how do you measure customer service fairly? Establish your own baseline first, then watch the direction. Calculate with the same method for four weeks in a row and from week five you have a line of your own, which is more accurate than anyone else's number.

One more principle: watch two or three metrics at a time. Chase all eight at once and the usual result is that nobody looks at any of them.

The 8 customer service metrics in one table

The "what bad looks like" column deliberately contains no numbers. It describes what you will see in the conversations.

Which metrics first? Pick by what hurts most right now

You do not have to start all eight together. Ask one question: what do customers complain about most this month? Pick from there.

Once you have chosen, watch for four weeks. Do not switch metrics because one week looks bad. A metric is a bathroom scale, not a diet pill: weighing yourself does not make you lighter, but without it you cannot tell whether you are heading the right way. When those two numbers are stable and the weekly review has become a habit, add the next pair.

No helpdesk? Sample 30 conversations a week

The back ends of LINE (the dominant messaging app in Taiwan), Instagram and Facebook will not hand you these eight numbers in one report. That is fine. A manual sample does the job and takes about as long as a cup of coffee each week.

Three sampling rules

First, fix the method. For example ten conversations per platform, or a set number per day. Once decided, do not change it. Sampling LINE this week and Instagram next week gives you two numbers that cannot be compared.

Second, do not cherry-⁠pick. Take every few conversations from the top of the inbox downward. Do not choose only the smooth ones, and do not choose only the arguments.

Third, one conversation is one issue. The same customer asking about a booking and then parking on the same day is one conversation; coming back three days later about something else is a new one. If last week had fewer than 30 conversations, log them all.

Log sheet template

Paste this into a spreadsheet and fill one row per conversation.

[Customer service conversation sample log]
No. | Platform | Customer's first message (time) | First reply (time) | Who replied first (AI / person) | Handed to a person? | Resolved? (yes / no / waiting) | Closed (time) | Repeat contact? | Order placed? | Notes
1 | LINE | | | | | | | | |
2 | Instagram | | | | | | | | |
3 | Facebook | | | | | | | | |

The eight metrics, one by one

1. First response time

How long it takes from the customer's first message until they receive the first reply with real content.

First response time = time of first substantive reply − time of customer's first message
Weekly figure = median of the 30 conversations (sort from shortest to longest, take the middle one)

How to collect it: read the timestamps in the conversation, enter them in the log sheet and subtract. Calculate business hours and out-⁠of-⁠hours separately.

How it misleads: an average gets dragged around by one or two overnight conversations, so use the median. Also, "Hello, we will reply shortly" is not a substantive reply. An instant empty line makes the number look good while the customer is still waiting.

2. Resolution time

From the moment the customer speaks up until the issue is truly finished.

Resolution time = time closed − time of customer's first message
Closed = the customer has the answer or has completed the next step, and has not asked about the same issue again

How to collect it: use the time of the last effective reply as the closing time. Leave unclosed conversations out and count them separately.

How it misleads: a short resolution time does not prove good work. It may just mean the hard cases were left open. Always read it next to the number of conversations still "waiting".

3. Resolution rate

Of the conversations sampled, how many actually dealt with the customer's issue. If you use AI, also calculate a second figure: resolved without handover.

Resolution rate = resolved conversations ÷ conversations sampled
Resolved without handover = conversations resolved with no handover to a person ÷ conversations sampled

How to collect it: have the same person judge each conversation as yes, no or waiting. Write the criteria down so anyone else applies the same ones.

How it misleads: a customer who stops replying is not necessarily resolved. They may have given up. This is the box where resolution rates get inflated most often; when in doubt, count it as "no".

4. Handover rate

How many conversations had to pass from auto-⁠replies or AI to a person.

Handover rate = conversations handed to a person ÷ conversations sampled

How to collect it: the "Handed to a person?" column in the log sheet, with the reason in the notes.

How it misleads: neither high nor low is good or bad in itself. Too high may mean the knowledge has too many gaps; too low may mean complaints that should have been handed over were absorbed by the AI. What matters is whether the conversations that were handed over really needed a person.

5. Repeat-⁠contact rate

The share of customers who come back a second time about the same issue.

Repeat-⁠contact rate = conversations where the customer returned about the same issue within your chosen number of days ÷ closed conversations

How to collect it: choose the number of days yourself and then keep it fixed. When sampling, scroll back to that customer's previous conversation and check whether it is the same issue.

How it misleads: a customer returning with a new question is a good thing and does not count. Only "you said you would get back to me and did not" and "I could not understand the last answer" count.

6. Missed conversations

Conversations that nobody has answered past the time limit you set yourself.

Missed conversations = conversations with no reply past your own time limit
Missed rate = missed conversations ÷ all incoming conversations

How to collect it: do not sample this one. Open each platform's inbox every day and count unread and unanswered conversations. The time limit is yours to set, for example how long during business hours, and by what time overnight messages must be cleared.

How it misleads: counting only unread messages misses the ones that were read and never answered. On Meta's platforms a miss also has a practical consequence: the Messenger Platform policy states that businesses have up to 24 hours to respond to a user, and after that, sending a message has to meet other rules such as message tags.

7. Customer satisfaction (CSAT)

Ask the customer directly whether they were satisfied with this interaction.

CSAT = number of satisfied responses ÷ number of people who gave a rating
Rating response rate = number of people who gave a rating ÷ number of people asked

How to collect it: send one short message when you close the issue, for example "Did this reply help? Send 1 for yes, 2 for no." Define in advance what counts as satisfied and do not change it later.

How it misleads: when only a few people respond, the score describes those few people. Always read CSAT together with the rating response rate. If you only ask customers who seem happy, what you get is comfort, not a metric.

8. Conversations-⁠to-⁠order

Only for shops that can track orders.

Conversations-⁠to-⁠order = orders ÷ conversations opened

How to collect it: define what counts as an order first, then make each entry point distinguishable. We covered how in the conversational commerce guide and will not repeat it here.

How it misleads: not every conversation is a purchase enquiry. In a week with many after-⁠sales questions this ratio drops although the service has not got worse.

After you add AI, read your KPIs differently

The moment AI goes live, first response time looks excellent automatically, because the AI does not sleep. Keep staring at speed and you are staring at a number that will never tell you anything again. Measuring AI customer service means shifting attention to two things: whether the answers are right, and whether the handover is good.

Handover quality deserves one more number:

Handover wait time = time of the person's first reply − time the AI handed over

Below is an AI spot-⁠check sheet. Every week, sample a fixed number of conversations the AI answered and have a person judge each one.

[AI reply spot-⁠check sheet]
No. | What the customer asked | Was the AI's answer correct? (right / partly / wrong) | Did it come from approved knowledge? | Did it promise anything it should not? | Did it hand over when it should have? | Did it tell the customer about the handover? | How long until a person took over? | Knowledge or rule to fix
1 | | | | | | | |
2 | | | | | | | |
3 | | | | | | | |

Every "wrong" must map to something in the last column. How to test before and after launch, and when not to adopt AI at all, is laid out with a full test matrix in AI customer service implementation: 9 questions to ask first. For how to change the rules, see how to write AI customer service rules.

The weekly review fits on one page

Numbers without a review are numbers you did not need to calculate. No slides required; one page is enough.

[Weekly customer service review]
Week:        Reviewer:        Conversations sampled:

1. This week's numbers (last week in brackets)
Median first response time:      (      )
Resolution rate:      (      )
Handover rate:      (      )
Repeat-⁠contact rate:      (      )
Missed conversations:      (      )
CSAT and rating response rate:      (      )
Conversations-⁠to-⁠order:      (      )

2. The one metric that changed most:
3. Which conversations we read, and the line where the customer stopped:
4. The one thing we change this week:
5. Who owns it, and done by which day:
6. Did last week's change work:

Please write only one item under point 4. Change five things in a week and when the numbers move next week you will not know which change earned it.

Common pitfalls and how to avoid them

• Tying KPIs straight to bonuses. People start managing the numbers, for example racing to send an empty first line or marking hard conversations as resolved.
• Watching only speed. An instant wrong answer is just a more efficient way of making a mess.
• Changing the formula every week. Once the denominator changes, before and after cannot be compared; if you must change it, start a new baseline.
• Calculating without reading. Numbers tell you where things got worse. The reason is in the conversations.
• Looking at each platform separately and never adding up. A customer who gets no answer on Instagram and asks again on LINE looks normal on both sides. For how to organise across platforms, see the omnichannel customer service guide.
• Treating complaint volume as a bad metric. A customer who is willing to tell you is giving you a chance; how to handle it is in the complaint handling guide.

FAQ