First response time in chat: how to measure and genuinely reduce it

First response time is one of the few service metrics customers experience directly. They do not care about the number of messages handled or the team's workload—they care how long they waited before someone responded. That is precisely why this figure is so easy to improve superficially: simply change when the clock starts or count a welcome message as a response.

Define the measurement before trying to improve it

Two companies can report the same result while measuring entirely different things. Before comparing yourself with anyone, define four points and record them somewhere the whole team can access.

• Measurement startsUsually with the customer's first message in a new thread, not when they open the chat window.
• Measurement endsWith the first human response addressed to the customer. An automated greeting is a different metric.
• Business hoursA message at 11:40 p.m. when coverage ends at 5 p.m. should not damage the team's daily result.
• ReopeningA case reopened after a week starts a new measurement; it is not a continuation of the old record.

Separating an automated reply from a human response is crucial. A “thank you, we'll reply shortly” message has value because it confirms receipt, but it resolves nothing. Count it in the metric and you will see a result measured in seconds while losing the only warning signal you had. We show how to word such a message from the customer's side in our guide to what to tell customers waiting in live chat.

Median and p90 instead of a single mean

The mean is particularly vulnerable to distortion in customer service. A few conversations picked up after several hours can spoil the picture of a month in which most customers received a quick response. The reverse is also true: a flood of instant responses can hide a group of people who waited a very long time.

↔ Scroll the table sideways to see all columns.

A practical rule: the median tells you how you work from day to day, while p90 tells you who has every right to be dissatisfied. If the median improves while p90 rises, the team is usually handling the easiest conversations faster and putting off the difficult ones. That is natural behavior, but it is worth seeing.

Apply the same caution to resolution time. A fast first response followed by a week-long wait for resolution is something the response metric will not reveal. View both figures side by side, just as a website conversion audit examines every stage of the journey separately instead of judging only the final result.

What is a realistic target?

Copying someone else's claims rarely ends well because you do not know their measurement definition or team size. A more sensible starting point is your own median from the last quarter and the expectations associated with each channel. In chat, the customer is in the middle of a visit and waiting live—minutes, not hours, are the relevant unit. A much longer wait is acceptable for an offline form or email inquiry.

1. Measure the baseline.The median and p90 from the last three months, counting only the team's business hours.
2. Set a target for each channel.One for live chat and another for messages left outside business hours.
3. Check feasibility.Compare the number of peak-hour conversations with the number of available agents. A target without people is a wish.
4. Record exceptions.Cases requiring technical consultation may have a different target from pricing questions.
5. Advertise it cautiously.A promise displayed in the widget must still be kept on Friday afternoon.
6. Review it monthly.A target nobody checks stops working after the second month.

A target without people remains a wish, so the next step is usually to chart the hourly distribution of inquiries and plan coverage around it. We explain this in our article on customer service staffing.

To turn that target into an operational commitment, define when measurement starts, pauses and stops. See the policy example in the guide to customer support SLAs.

What genuinely reduces response time

Most gains come from organization, not faster typing. The changes below work regardless of team size, although the first two matter most for a one-person support function.

• Clear ownership: a named person for specified hours instead of “someone on the team will see it.”
• Route conversations to the right department immediately instead of transferring them after the first response.
• Use visible notifications outside the panel if the agent does not sit in it all day.
• Keep availability aligned with reality and switch to offline mode when no one is present.
• Use prepared responses to recurring questions as a starting point, not a finished message.
• Limit pre-chat questions to fields that genuinely speed up service.
• Disable proactive prompts during hours when no one can pick up conversations.

The last point often comes as a surprise. An invitation to chat sent while the team is unavailable generates messages with no recipient and damages the metric you want to improve. Our article on live chat automation explains when to reach out and when to hold back.

It is also worth considering language-model assistance when drafting responses. It can reduce preparation time, but it cannot replace human presence during peak hours. We discuss the limits in our article on AI in customer service.

Prepared answers to recurring questions shorten drafting time but can easily make you sound like a machine. Our article on customer service response templates explains how to build them while leaving room for a genuine conversation.

With a larger volume of cases, there is also the question of order—faster does not necessarily mean first come, first served. We explain the criteria in our article on prioritizing the ticket queue.

In a small team, the target must also survive holidays and peaks in incoming requests. Review staffing coverage for your SLA before making a tighter response-time promise.

Shortcuts that improve the number but harm the result

↔ Scroll the table sideways to see all columns.

The simplest way to check whether the improvement is real is to compare first response time with the service rating and the share of cases returning within 48 hours. If the metric falls but those two figures do not improve, the report was probably optimized rather than the service.

The most costly leak happens after contact: a message that reached no one. We have collected the full list of such points in our article on where you lose customer inquiries.

A monthly review routine

1. Check the median and p90.Separately for chat and for messages left outside business hours.
2. Find the hours when queues form.The hourly distribution shows whether the problem is headcount or how coverage is arranged.
3. Review the longest conversations.The five extreme cases usually share a cause, such as one department having no backup.
4. Compare it with the service rating.Faster but worse is not an improvement; it merely moves the problem.
5. Make one change.One change per month lets you identify what worked. Three at once do not.

This routine takes several dozen minutes a month and is usually enough to catch deterioration before customers notice it. More important than frequency is ensuring the review ends with a decision, not merely figures entered into a spreadsheet.

If these figures need to go beyond the team, limit them to a handful and add commentary. We explain how in our article on customer service reporting.