EDUKATE SINGAPORE · GOLD STANDARD PRACTICAL CASE STUDY 223
An average response time can conceal a family waiting far longer
A fictional school announces a ten-minute average helpdesk response, but one of the ten recorded inquiries waited 91 minutes. The mean is correct—(nine one-minute waits plus one 91-minute wait) divided by ten equals ten minutes. It does not establish that everybody was answered promptly.
All responses, families and service targets in this case are imaginary. The service-desk image is an editorial illustration, not a depiction of the invented helpdesk.
The ten-case service record
- First meaningful response times (minutes): 1, 1, 1, 1, 1, 1, 1, 1, 1, 91.
- Total waiting time across cases: 100 minutes.
- Arithmetic mean: 100/10 = 10 minutes.
- Median: 1 minute.
- Fraction meeting the fictional 15-minute target: 9/10 = 90%.
- One 91-minute exception accounts for 91% of total observed waiting minutes, not 91% of families.
Chapter navigation
- 1. A friendly average conceals one long wait
- 2. Define exactly what a ‘response’ means
- 3. Work through the ten actual numbers
- 4. Why a mean can be dominated by a tail event
- 5. Google SRE explains why percentiles matter
- 6. A service target is not an SLA merely because it appears on a poster
- 7. One exception deserves investigation without automatic blame
- 8. Segment cases before making global promises
- 9. Missing and abandoned requests must not vanish from the denominator
- 10. A corrected statement can celebrate success without erasing the delay
- 11. Teaching exercise: make three accurate statements from one dataset
- 12. The management decision should follow the slowest important service need
1. A friendly average conceals one long wait
A fictional school begins tracking how quickly its family support desk answers incoming messages. In a small illustrative ten-case sample, nine families receive a first meaningful reply after one minute each. The tenth waits 91 minutes. The school reports an average first-response time of ten minutes and announces that everyone is answered promptly.
The average is arithmetically correct: nine times one minute, plus 91 minutes, gives 100 total minutes of waiting across ten cases; divided by ten, that is ten minutes. But a mean describes the distribution in only one way. It cannot establish that every family waited ten minutes or less, and it obscures how different one family’s experience was.
The school may have a fast and caring support team for most cases. The existence of one serious delay does not erase the team’s positive work. It does require a more honest performance claim and an investigation into the delayed case.
Every parent, school, case, service promise and response measurement in this story is invented. The featured service-desk illustration is conceptual, not an image of an actual parent’s inquiry.
2. Define exactly what a ‘response’ means
Before calculating anything, the helpdesk should agree on the event being timed. Does the clock begin when the message reaches the supported queue, when an agent opens it, or when a parent fills out the form? Does it stop at an automated acknowledgement, the first human message, or a useful answer? Those are distinct measures.
In the fictional exercise, the start is the timestamp at which a complete incoming request reaches the school’s designated helpdesk queue. The end is the timestamp of the first meaningful human response that addresses or advances the inquiry. A robot’s immediate ‘we received your message’ acknowledgement does not meet the chosen definition.
A response time is not resolution time. One family may receive a fast reply saying the issue is being investigated and wait much longer for closure. Another may have the whole question answered in the first message. A credible performance report can disclose both indicators without pretending they measure the same experience.
The Gold Standard requires the measurement boundary to be explicit before a dashboard displays any statistic. Otherwise a change in software wording can look like an improvement in service even when the experience of a family has not changed.
3. Work through the ten actual numbers
The ordered fictional response times, in minutes, are 1, 1, 1, 1, 1, 1, 1, 1, 1 and 91. The total is 100 minutes. The arithmetic mean is 100/10 = 10 minutes. The median, calculated from the fifth and sixth values, is one minute.
Nine out of ten cases had a first meaningful response in one minute, so the proportion receiving a response within a hypothetical fifteen-minute service window is 9/10, or 90%. One case did not meet the fictional fifteen-minute criterion. These figures can all be true simultaneously.
It would be wrong to report ‘ten out of ten families answered within ten minutes’ merely because the mean is ten minutes. Only nine cases meet that threshold in the sample. It would also be wrong to announce that 91 minutes is the typical experience of every parent. Most responses in the sample were short.
The numbers should be accompanied by sample size, period, case definition and data completeness. With only ten cases, the sample is useful for explaining the indicator but not sufficient to certify the school’s long-term performance across busy terms, weekends and different service categories.
4. Why a mean can be dominated by a tail event
An arithmetic mean is affected by every observation, including large outliers. In the toy sample, nine cases contribute a total of nine minutes and the one delayed case contributes 91 minutes. The single delayed case therefore accounts for 91% of total recorded waiting minutes, despite representing only ten percent of the inquiries.
That 91% figure is a share of observed waiting minutes, not a percentage of harmed families, failed inquiries or staff work. The distinction matters. A large wait may represent a serious communication problem, but its practical impact must be reviewed separately from its contribution to the arithmetic.
The median of one minute describes the middle of the distribution and is less affected by the one extreme value. Reporting only the median would hide the tail in a different way. A responsible summary therefore considers the spread and the proportion of cases meeting a specified response target, along with the raw extreme waiting times.
The purpose of measuring service is not to pick whichever statistic makes the school look fastest. It is to help it identify where families experience avoidable delay.
5. Google SRE explains why percentiles matter
Google’s Site Reliability Engineering guidance on service-level objectives explains that averages can hide the long tail of response times. It recommends distributional measures and percentiles for systems where a small subset of users experience severe delays.
Google’s discussion concerns online service reliability, not a binding standard for a school reception team. The transferable statistical principle is that the mean and median of a skewed response-time distribution can diverge and that high-end delays matter to actual users.
A percentile answers a different question from an average. For example, a stated 90th-percentile response time needs a precise calculation convention and sufficiently meaningful sample size. In this toy sample, nine of ten responses are at one minute and one is at 91; jumping from that to a polished ’99th percentile’ claim would be unstable and method-dependent.
The more dependable management practice is to report the fraction meeting the selected service window, meaningful distribution summaries and individually material exceptions. Each serves a different purpose.
6. A service target is not an SLA merely because it appears on a poster
The fictional desk’s internal goal is to send a first meaningful response within fifteen minutes during defined service hours. A sample where nine cases meet that window has 90% attainment for those ten completed observations. Whether that is acceptable depends on the target the school actually adopted, its scope and the needs of the families.
A service-level objective is an operational aim; a contractual service-level agreement may have additional terms, exclusions and consequences. The school should not invoke the terminology loosely to suggest compliance with obligations that were never defined.
The reporting team must state whether the clock runs outside staffed hours, how holidays and urgent incidents are handled and whether unresponded inquiries count as misses rather than disappearing from the dataset. Omitting still-open long waits can make an apparent success rate artificially high.
A clear target should be measurable, relevant to the user and worth acting on. A convenient ten-minute average is not automatically a good substitute for a service promise expressed as ‘reply within fifteen minutes’.
7. One exception deserves investigation without automatic blame
The fictional 91-minute wait may have arisen from a routing error, an unusually complex inquiry, a temporary absence, a broken notification or another cause. The reported time alone does not reveal which. A manager should not assume that a particular employee was negligent because their case appears at the end of a list.
The first review checks timestamps, service hours, case classification, queue assignment and whether urgent requests had a documented escalation route. It should identify whether the parent was eventually given a useful answer and whether the delay created a material consequence. Contact with the family should be respectful and protective of privacy.
If a genuine processing bottleneck is found, the team can correct it. If the measured time included a period outside declared service hours, the reporting definition may need clarification. Neither outcome permits silently deleting the slow case merely to improve the average.
The Gold Standard keeps operational evidence and personal judgement separate. The purpose is to repair a service, not manufacture a favourable metric or an accusation.
8. Segment cases before making global promises
A single aggregate can combine distinct request types. Routine timetable questions may be simple to answer, while urgent safety-related inquiries require different escalation. Translation or accessibility needs can also affect the way families receive help. A useful report identifies whether the same target is sensible for all categories.
The school should avoid treating slower replies to complex or accessible communications as proof that those families are less important. Rather, it should design channels, staffing and escalation paths appropriate to the service promised. Comparisons should be fair and should not exclude difficult inquiries.
Reporting by channel—phone, web form, email and in-person desk—may expose handoffs that a single average conceals. But subgroup samples can be small, and public reports should avoid disclosing personal information about identifiable families in tiny categories.
A good dashboard gives operators enough detail to improve the system while giving families a concise and truthful description of what performance they can expect.
9. Missing and abandoned requests must not vanish from the denominator
Imagine one additional fictional request arrived but never received a meaningful response during the observation window. If the analyst calculates response times for only the ten completed cases, the published mean applies to completed cases, not all requests received. That may be a useful calculation if clearly labelled, but it would not demonstrate universal timely response.
The reporting process should define how open cases, cancelled inquiries, duplicate tickets and automated acknowledgements are counted. Waiting-time measurements for unresolved cases are censored or ongoing rather than equal to zero. Treating them as zero would make the system look faster precisely when an answer is missing.
The correct denominator for a fifteen-minute service-attainment indicator needs to include all eligible cases whose outcomes can be determined for that target, including late or unanswered cases under the chosen rule. The operational data pipeline should document exclusions rather than letting them accumulate invisibly.
The Gold Standard asks whether every case enters the metric that corresponds to its outcome. A neat average over an incomplete population can be both mathematically correct and managerially misleading.
10. A corrected statement can celebrate success without erasing the delay
The school could say: ‘In this ten-case illustrative sample, nine families received a meaningful first reply within one minute, while one waited 91 minutes. The mean was ten minutes and the median was one minute. Nine of ten met our fictional fifteen-minute first-response goal. We are reviewing the delayed case and the reliability of our queue handling.’
That statement is longer than ‘everyone answered promptly’, but it is honest and actionable. It gives parents a sense of typical service, exposes the significant exception, and explains what management will do. With a larger real dataset, the same structure can be presented clearly without listing personal cases.
A corrective notice should replace the unsupported ‘everyone’ claim if it was distributed publicly. The school should preserve the original records, a dated correction and an owner for reviewing service performance after any process change.
Trust grows when a service measures what its users experience instead of selecting the most flattering single number.
11. Teaching exercise: make three accurate statements from one dataset
Give students the ten wait times and ask them to calculate the mean, median and fraction receiving a reply within fifteen minutes. They should obtain ten minutes, one minute and ninety percent. Then ask whether each number alone justifies the claim that every parent was answered quickly.
A second group can draw a simple dot plot with nine dots at one minute and one at 91 minutes. The plot makes the long-tail experience visible without requiring complex statistics. Another group can define a better service measure and explain what to do with unanswered requests.
Advanced learners can discuss why sample quantiles require a stated convention and why ten observations cannot support stable promises about the most extreme percentiles of future service. They should also propose how to separate response time from final resolution and how to avoid releasing identifying details in public reports.
The exercise transfers to hospital queues, software latency, call centres, library services and administrative applications. Whenever people wait, a mean is only one part of the experience.
12. The management decision should follow the slowest important service need
The school may report that its ten measured first responses averaged ten minutes, with a one-minute median, and that nine met its hypothetical fifteen-minute target. It cannot say every case met the target, because one was delayed 91 minutes. That exception deserves a scoped investigation and follow-up evidence.
A response-time dashboard should include a defensible service-level indicator, complete eligible cases, clearly defined measurement windows and enough distributional information to reveal problems. A later improvement must be verified under the same definitions rather than announced because a summary number changed.
No single observed delay proves that a service is generally failing; no low mean proves every family receives timely help. The right conclusion is bounded by the observed data.
The Gold Standard is a family support process that measures users’ real experiences, reports both normal service and meaningful exceptions, and uses those findings to improve the next interaction.
Practical performance reporting checklist
- Define first meaningful response separately from acknowledgement and resolution.
- Use complete eligible case populations and document unanswered inquiries.
- Report mean, median, target attainment and material long waits.
- Check service hours, urgent inquiry routing and accessibility needs.
- Investigate exceptions without assuming individual blame.
- Retest with consistent indicators after changing staffing or systems.
Frequently asked questions
If the average response is ten minutes, was everyone answered in ten minutes?
No. The ten-case example has nine one-minute waits and one 91-minute wait, producing a ten-minute mean.
What is the median in this dataset?
One minute, since the fifth and sixth sorted values are both one minute.
Does 90% attainment mean the school’s service is satisfactory?
Not by itself. That depends on the defined goal, sample coverage, service context and consequences of the delayed cases.
Can automated acknowledgements count as a response?
Only if the indicator explicitly defines that event. It should not be silently equated with a useful human answer.
Why report incomplete cases?
Excluding unanswered inquiries can bias a performance report toward fast completed cases and hide the very waits a service needs to address.
Authoritative performance references and connected cases
- Google SRE · Service Level Objectives
- Google Cloud · Building Good SLOs
- Google SRE Workbook · Implementing Service-Level Objectives
- The Gold Standard of Customer Service
- Gold Standard Case 218 · Nine survey responses and parent satisfaction
- Gold Standard Case 215 · Successful backups and failed restoration
- Gold Standard · Core Requirements and Evidence
- Gold Standard · Casebook
Google SRE resources provide general lessons about distributions and service-level objectives, not school-helpdesk accreditation. All response times and service rules are fictional instructional records. This case is about careful operational measurement and communication.