How to measure whether customer service resolves requests
A request is resolved when the outcome defined for that type of matter has been confirmed, not merely when someone has replied. Record requests, stages, reopenings and review work. To assess AI or manual service, use the same definitions and distinguish guidance, handover to the team and actual completion.
By the end, you can:
- Define what resolving each request type means
- Choose denominators and observation windows
- Review quality, reopenings and human work alongside automation
Write the expected outcome for each matter type
An address question may end when correct information is delivered; an appointment requires confirmation of availability; a delivery problem may need investigation and a decision. If all are counted as “conversation answered”, you lose the ability to understand what was genuinely completed. Define simple categories and observable outcomes before building the dashboard.
Separate resolution from administrative closure. A request may be closed because it was duplicated, cancelled or abandoned without its initial objective being achieved. Record the appropriate reason rather than improving the resolution rate by renaming states. Definitions should be understood by the people providing service and analysing it.
Choose comparable requests as the unit
Messages, contacts and requests are different counts. A person may send ten messages about the same problem or raise two matters in one conversation. Decide how each request is identified and how duplicates are avoided, retaining only data needed for operations. The reference should connect continuations without turning every new message into another chance to count success.
Define what enters the denominator: all requests received, only those eligible for a flow or a cohort from a specific period. Show exclusions and reasons. A rate calculated only over easy cases does not represent all customer service. If comparing AI and staff, also check whether they receive similarly complex matters or one mainly handles exceptions.
Combine outcome, waiting and quality
Choose a few indicators that inform decisions: unassigned requests, time to the first useful reply, confirmed resolution, reopenings and consequential errors. Record human work reviewing, correcting and handling exceptions. Automation can reply quickly while moving substantial work to later; looking only at the first message conceals that cost.
Define calendar time or covered hours and retain the distinction. A weekend request should not be compared without context to one received during a shift. Use distributions or ranges when an average hides very slow cases. Microsoft’s documented service dashboards illustrate the need for categories and case age; specific definitions should match your process.
| Indicator | Definition to approve | Interpretation caution |
|---|---|---|
| First useful reply | Reply that guides or advances the request | A greeting is insufficient |
| Confirmed resolution | Observed outcome by matter type | Do not confuse with closing the conversation |
| Reopening | Return of the same problem within the chosen window | Requires linking records |
| Total work | Service, review, correction and exceptions | Do not count only AI time |
Define when completion can be assessed
Some requests appear resolved at the time and return later. Choose an observation window suited to the service and explain how still-open cases are handled. A matter started yesterday may not have had time to reach an outcome; mixing it with older cases may lower a rate without indicating poorer quality. Analysis by entry cohort helps maintain comparability.
Do not choose the window after seeing the result simply to produce a favourable percentage. Record the definition, extraction date and treatment of reopenings. If operations change, show the change in the series. A dashboard silently changing its rules may appear to show progress while no longer measuring the same thing.
Calculate using requests and reopenings
Fictitious example: a pilot tracks twenty eligible requests from the same week. Twelve have the outcome initially confirmed without handover to another professional. During the chosen review window, three of those twelve reopen because the same problem continues. Under the stable first-contact resolution definition used in this exercise, nine out of twenty remain, or 45%.
The remaining requests need their own classification: pending, referred, cancelled or resolved later. The example does not set an industry target or prove that the tool performed well or poorly. It only shows why the initial message and final completion are not the same measure. The window and criteria should be chosen before the pilot.
Read cases to understand the number
Choose a sample including completions, reopenings, transfers and unanswered matters. Check whether information was correct, customers knew the next step and the state reflects what happened. In evaluations using real data, limit access and content to the purpose; for open training, use fictitious or appropriately prepared examples.
Classify errors by consequence rather than treating a comma and an incorrect commercial promise as equivalent. Prioritise false conditions, data exposure and important requests without a destination. Human review may find a problem that has not yet changed an aggregate metric. An indicator should guide an investigation, not replace it.
Compare changes with context and full cost
Keep a record of volume, request types, opening hours and staffing over the compared periods. An improvement may coincide with lower demand or removal of difficult cases from the flow. If you cannot separate these effects, present the conclusion as a limited observation. Do not automatically attribute increased sales to response speed.
Include implementation, training, source reviews and exception handling in the decision. The choice may be to keep automation for a simple group and refer other matters earlier. A lower percentage of automatic replies can accompany higher quality. The desired result is service that resolves requests safely and with proportionate effort within the business’s actual rules.
Keep the definitions before presenting the dashboard
The worksheet should explain each number’s meaning and allow the calculation to be reproduced. Include denominator, exclusions, window, source and person responsible. If a tool cannot measure a stage directly, identify the confirmation procedure or record the limitation. Do not fill a chart with a different indicator simply because it is available.
Review definitions with the team and keep examples for boundary cases. A request cancelled because of unavailable capacity, for instance, should not change classification depending on who closes the record. The dashboard becomes useful when several people reach the same interpretation and can choose a specific improvement from it.
Worksheet for your project
Request type and outcome defining resolution: Counting unit and how duplicates are avoided: Entry cohort, period and denominator: Exclusions and their reasons: First useful reply and hours considered: Completion evidence and reopening rule: Observation window and pending cases: Human work included: Reviewed sample and errors by consequence: Changes in volume, offer or staffing: Decision and next review:
Sources and criteria
Content prepared with AI assistance and research using technical sources, with checks of the relevant examples and calculations. The references support the guide’s decisions; examples identified as fictitious let you try out the method.
