> For the complete documentation index, see [llms.txt](https://en.help.firstline.cc/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://en.help.firstline.cc/feature/dashboard/customer-service-efficiency-metrics.md). # Customer service efficiency metrics FIRST LINE’s “Customer Service Efficiency Metrics” are used to observe the overall efficiency from when customers enter a service, wait for a human reply, are picked up by an agent, to service completion. Managers can quickly determine whether issues mainly occur in the queue, first reply, follow-up replies, or overall handling time through metric cards and trend charts. ## Where to view Go to the “Dashboard” in FIRST LINE and find “Customer Service Efficiency Metrics” on the Basic Information page. The system shows recent data by default, and you can further filter by date, agent, department, group, channel, and business hours. The trend chart can be switched between “day/week/month,” making it suitable for observing whether customer service efficiency improves or worsens over time. ## What do the five metrics represent To make it easier to understand, the same service below is used as an example: * 10:00 Customer enters the service * 10:01 Customer sends the first message * 10:02:30 Agent picks up and starts the service * 10:04 Agent replies in person for the first time * 10:06 Customer sends another message * 10:07 Customer adds a second message * 10:08 The system or bot automatically replies * 10:10 Agent replies in person * 10:12 Customer sends another message * 10:13 Agent replies in person * 10:20 Service completed The example below is used to explain how a single service would be counted. In the actual dashboard, valid samples within the selected period are organized into statistics to help you observe the overall situation. ### First human reply after entering the service The time from when the customer enters the service until the agent replies in person for the first time. This number includes the time the customer waits for assignment, the agent’s pickup time, and the time between pickup and the first human reply, so it is suitable for observing the overall waiting experience from when the customer enters the service until someone actually responds. Bots, system messages, and automatic replies are not considered human replies. If some services do not have a clear entry time, the system uses the time the service record was created as the starting point, so the actual starting point may differ slightly depending on the service scenario. > **Example**: Customer enters the service at 10:00, and the agent replies in person for the first time at 10:04; this service will be counted as **4 minutes**into “First human reply after entering the service.” Even though an agent has already picked up at 10:02:30, the time is still counted until the actual human reply. ### First human reply time The time from when the customer sends the first message until the agent replies in person for the first time. This metric is closer to what is generally called “first response time,” and is used to observe how long customers wait to receive a human answer after they actually raise an issue. If a service has no customer message, it will not generate a valid sample for this metric. > **Example**: Customer sends the first message at 10:01, and the agent replies in person for the first time at 10:04; this service will be counted as **3 minutes**into “First human reply time.” For the same service, “First human reply after entering the service” is 4 minutes because the starting point of the two metrics is different. ### Pickup time The waiting time from when the customer enters the service until the agent actually picks it up and starts servicing. This metric mainly reflects queueing and assignment efficiency. An agent having picked it up does not mean the customer has already received the first human message, so “pickup time” and “first human reply time” may differ. If pickup time is high, it usually means customers spent more time waiting to be assigned or taken over. > **Example**: Customer enters the service at 10:00, and the agent picks it up at 10:02:30; this service will be counted as **2 minutes 30 seconds**into “Pickup time.” The agent’s actual reply at 10:04 later does not change this pickup time. ### Reply time After the agent completes the first human reply, the customer sends another message, until the time of the next human reply by the agent. If the customer sends multiple messages continuously while waiting, the system calculates the waiting time starting from the first message; automated bot or system replies also do not stop this waiting period. The sample count for this metric represents valid back-and-forth interactions between customers and agents, so it may usually be higher than the sample counts of other metrics. If first human reply time is normal but reply time is high, it usually means the initial response speed is fine, but the follow-up exchanges during the service are slower. > **Example**: The agent has already completed the first human reply at 10:04. The customer sends another message at 10:06, then adds one more at 10:07. Although there is a system automatic reply at 10:08, the waiting period does not end until the agent replies in person at 10:10, so this round will be counted as **4 minutes**4 minutes **1 minute**valid sample **2 reply samples**and not just as 1 service. ### Resolution time The time from the customer’s first message until service completion; if the service has no customer message, it is calculated from the time the customer enters the service. This metric reflects the overall handling cycle, so it is usually longer than the first reply or pickup time. A longer resolution time does not necessarily mean customer service is slow; it may also be related to issue complexity, waiting for customer information, cross-department confirmation, or other handling processes. > **Example**: Customer sends the first message at 10:01, and the service is completed at 10:20; this service will be counted as **19 minutes**into “Resolution time.” If a service has no customer message from start to finish, it will instead be calculated from the time the customer enters the service until completion. ## How to determine where the problem may be occurring Different metrics are suitable for answering different questions. It is recommended not to look at only one number, but to interpret several metrics together. * **High pickup time**: Customers may have spent more time waiting to be assigned or waiting for an agent to take over; you can further review staffing, shift scheduling, or assignment methods. * **Pickup time is normal, but first human reply time is high**: This means the agent has already picked it up, but the customer still waited a while before receiving the first human reply. * **First human reply time is normal, but reply time is high**: This means the initial reply was fast, but follow-up exchanges during the service were slower. * **The first few items are normal, but resolution time is high**: The issue may not be the reply speed, but rather the overall case needing a longer handling time, such as waiting for confirmation, additional information, or cross-department collaboration. These metrics are suitable for helping identify trends and possible bottlenecks, and are not recommended to be used alone to judge an individual agent’s service quality. ## How to read the mean, median, and P95 The same customer service efficiency metric can be viewed from different angles in terms of waiting time. The three numbers each answer different questions: * **Mean**: Average all valid samples together, suitable for viewing overall time spent. A few particularly long cases may push the mean upward. * **Median**: Arrange the waiting times from shortest to longest and look at the middle value. It is closer to “how long most customers usually wait,” and is less easily affected by a few extreme cases. * **P95**: Used to observe the slower group of customers. You can think of it as “about 95% of valid samples can be completed within this time.” {% hint style="info" %} **Quick interpretation** If the mean and median are very close, it usually means the waiting times of most cases do not differ much. If the mean is noticeably higher than the median, it usually means a few particularly long cases are pulling the overall average up. At this point, you can look at P95 to determine how long the slower group of customers actually waited. {% endhint %} ### Understand with 20 conversations Assume a certain metric has 20 valid samples. Arranged from fastest to slowest, they are as follows: `1, 1, 2, 2, 3, 3, 4, 5, 5, 6, 7, 8, 10, 12, 15, 20, 30, 60, 600, 86400 seconds` In this data set, most cases are completed in a very short time, but a few at the end waited for a very long time. * **Median**will fall near the middle of the data, so it can reflect the typical waiting situation for most customers. * **P95**will fall toward the slower end, so it can help you see whether “most customers have already been handled, but are the slowest small portion of customers still waiting too long?” * **Mean**will be clearly affected by the last 86,400-second case, so if you look only at the mean, you may feel the overall situation is slower than what most customers actually experience. This is why it is recommended to look at the mean, median, and P95 together: putting the three numbers side by side makes it easier to distinguish between “generally slow” and “only a few cases are especially slow” than looking at a single mean alone. ## How to read the sample count Each metric card shows “N in total,” which represents the number of valid samples actually included in the statistics. The sample counts for different metrics are not necessarily the same. For example, “reply time” is calculated based on valid back-and-forth interactions, while most other metrics are based on service records that meet the conditions. Therefore, do not directly compare the sample counts of different cards as service volume. > **Example**: Suppose there are only 10 services today, but each service has an average of 3 valid back-and-forth interactions of “customer sends another message → agent replies in person.” Then the sample count for “reply time” may be close to 30; this does not mean there were 30 services today, but rather about 30 valid reply back-and-forth interactions. The fewer the samples, the more a single especially long or short case can affect the statistical result. When you see a number suddenly rise or fall, it is recommended to first confirm whether the sample count is sufficient, then interpret it together with the mean, median, and P95. ## How to read the trend chart The trend chart is mainly used to observe how the same metric changes over time, such as whether the first human reply in a certain week is gradually getting slower, or whether the resolution time in a certain month is continuously decreasing. It is recommended to first compare changes in the same metric across different dates, rather than directly comparing the heights of different metric lines. Because different metrics represent different service stages, they may naturally fall within completely different time ranges. If a certain day suddenly becomes higher, you can first look at the sample count, then compare the mean, median, and P95: * All three numbers rise together: it may indicate that the overall service on that day has slowed down. * The mean rises, but the median changes little: it may be caused by a few particularly long cases. * P95 rises significantly: it means the slower group of customers waited longer, and it is worth further review. ## Why there is no value on a certain day, but the line is still connected “No data” and “0 seconds” do not mean the same thing. * **No data**: There are no valid samples that meet the conditions for that period. * **0 seconds**: There are indeed samples, and the calculation result is 0 seconds. FIRST LINE does not fill dates with no data as 0, to avoid making users think the customer service efficiency that day was completed in “0 seconds.” To make the overall trend easier to read, the chart may connect data points before and after dates with no samples by skipping over those dates; this line is only to help read the trend and does not mean the dates with missing data have actual values. ## Available filter conditions You can narrow the analysis scope as needed, for example by looking only at a certain department, a certain agent, a certain service channel, or by comparing efficiency only during business hours versus outside business hours. Common ways include: * Check whether there are obvious differences in first human reply time across departments. * Compare resolution time across different channels. * Look only at after-hours incoming services to confirm whether longer waits exist. * Switch between day/week/month to observe short-term fluctuations and long-term trends. If you want to compare different agents, departments, or channels, it is recommended to use the same date range and filter conditions as much as possible to avoid drawing conclusions directly because of differences in case volume or service scenarios. ## Notes for use These metrics reflect “time efficiency” and are not the same as service quality, customer satisfaction, or whether the issue has truly been properly resolved. It is recommended to interpret them together with satisfaction, service results, case volume, and other information. Mean, median, and P95 each have different uses; there is no need to pursue any one number being lower as necessarily better. If the service content is more complex and requires cross-department handling, the resolution time may naturally be longer. When numbers suddenly change, it is recommended to first confirm the sample count, then observe whether the mean, median, and P95 change together. This better reflects the actual service situation than evaluating an individual agent based on a single day or a single number. If you want to compare team performance, it is recommended to first look at the overall trend over a longer period, then interpret it with similar date ranges, filter conditions, and sample counts; this will be more informative.