1. Define the decision before selecting the chart
Start with a specific question from the process owner. Replace “How productive is the warehouse?” with “Are confirmed orders handed over by the agreed deadline, and where do delays develop?” The second question identifies relevant events, boundaries and possible actions. If a change in a measure affects no decision, it may remain useful reference information, but it is not automatically a key performance indicator.
The GOV.UK Service Manual connects metrics with service objectives. That is a useful principle for an operations team, not a ready-made set of targets. Beside each metric, record who reviews it, when they review it, what they investigate and what they can change. Queue growth may call for examining blocking reasons rather than immediately hiring more people. A metric signals a situation; explaining its cause requires additional evidence and observation. Do not let one chart become an automatic diagnosis. State what the measure cannot tell you as clearly as what it can.
2. Give every metric a definition card
Document its name, business question, formula, unit, period, event source, inclusions and exclusions. Define the start and finish precisely. Order processing time might begin at draft creation, payment confirmation or stock availability. Each can answer a useful question, but they are different measures and must not silently alternate beneath the same heading. Make the definition accessible from the report rather than keeping it in a private analyst notebook.
Specify the time zone, business or elapsed time, cancellations and reopening. For split shipments, decide whether completion means the first parcel or the entire agreed quantity. Assign an owner and record formula changes. Do not compare a new series with the old one without explaining the change. Readers need to know whether performance improved or counting changed. This matters especially when the result influences departmental assessment or customer commitments. Preserve sufficient event history to reproduce the calculation, and document any periods for which that history is unavailable.
3. Separate touch time from elapsed lead time
Lead time describes how long an object remains between defined process events. Touch time describes active work on it. An order may need ten minutes of activity and wait two days for clarification. Automating two minutes of entry will not remove that waiting. Understanding the delay requires waiting states or sampled observation, not only a timer spanning the entire workflow. Identify whether the waiting is necessary, avoidable or outside the current team’s control.
Do not sum parallel activities and call the result customer elapsed time. Two simultaneous five-minute checks do not necessarily add ten minutes to the journey. Explain whose time each measure represents. If active work is not recorded, do not infer it from two status timestamps without verification. A clearly labelled estimate is better than false precision. Initially, observing ordinary and problematic operations may be enough to identify which part of the delay is realistically changeable. Choose an intervention based on that evidence rather than on whichever duration is easiest to display.
4. Inspect open work, not only completed work
Completed-order performance can improve while difficult cases accumulate unfinished. Pair completion time with the number and age of open objects. Define which statuses belong in the queue and whether waiting for a customer or supplier is included. Quietly excluding those states can produce an optimistic report that does not resemble the buyer’s experience. If you need an internal controllable-time measure, show it separately from the end-to-end measure.
Segment open work by blocking reason and expose the oldest cases. A list of specific overdue orders is often more useful in an operational meeting than average age. Do not mix open-item age with completed-item duration: they describe different populations. If the system stores no historical queue snapshots or reconstructable events, today’s state cannot establish last Monday’s backlog. Agree on event retention or regular snapshots. Otherwise the historical chart may be an assumption presented as a fact. Also identify abandoned records so they are resolved through a defined process rather than silently removed to improve the number.
5. Use the correct denominator for rates
Consider a fictional example. One group completes 90 of 100 orders on time; another completes 5 of 10. Their rates are 90% and 50%. Averaging those percentages gives 70%, but the overall order-level result is 95 out of 110, approximately 86.4%. The groups have different sizes. Combine numerators and denominators for an overall rate when the definitions are equivalent, rather than taking an unweighted average of group percentages.
Show volume beside a percentage. One defect in ten and one hundred in a thousand both produce 10%, but create different recovery workloads. Define whether you count defects, affected orders or customers: one order may have several defects. Avoid mixing cohorts, such as today’s returns divided by today’s sales when returned items were purchased earlier. That ratio may describe current operational load, but it is not automatically the return rate of today’s purchases. State the intended interpretation so readers do not make a conclusion the calculation cannot support.
6. Balance speed, quality and unfinished work
A single metric can encourage unwanted behaviour. Counting closed tasks alone creates an incentive to close them before the intended outcome. Measuring picking speed alone may transfer verification work to another team or the customer. Add a measure that reveals the cost of acceleration: rework, post-handoff defects or returned tasks. Define how those outcomes link back to the originating work, including failures discovered after the reporting period.
For one workflow, begin with a compact set: completed volume, lead time, open work and completion without rework. Do not hide them inside an opaque overall score. Each must remain understandable and traceable to individual records. The worksheet below illustrates three signal-to-investigation relationships, not universal productivity standards. Set thresholds using customer commitments, actual workflow and available capacity. A change in demand or product mix may require context rather than a conclusion that employees are performing worse. Review whether the measures together describe the customer outcome, not merely the activity of the easiest department to measure.
From a signal to an investigation
Operational-review examples. A signal alone does not establish its cause.
Scroll horizontally to read the table.
| Signal | Next check | Do not automatically conclude |
|---|---|---|
| Open orders are getting older | Oldest records, blocking reasons and stock availability. | That the team is working more slowly. |
| Tasks complete faster | Reopening, post-handoff defects and changes in complexity. | That the entire customer journey is faster. |
| On-time completion rate falls | Volume, consistent definitions, channels and promised deadlines. | That an average of departmental percentages is the overall rate. |
Takeaway. After investigating, record a hypothesis, change owner and effect-review date.
7. Look beyond the average
Average duration can remain acceptable while some customers wait much longer. Inspect the distribution, typical cases and long delays. A median describes the middle of an ordered dataset but does not reveal every problematic case. With enough observations, an agreed percentile can add context; with a small sample, show the cases themselves and avoid overstating stability. Explain the selected measure in language its users understand.
Compare similar operations. A single-line order and a large order requiring inspection have different workloads. Segmenting by complexity or workflow type may explain a change in overall duration without any deterioration in execution. Do not create dozens of segments containing only a handful of records for the appearance of precision. Choose dimensions that change decisions: channel, warehouse, delivery type or blocking reason. Investigate order-mix changes before judging individuals. A process metric is not automatic evidence of personal effectiveness. Where case assignment differs, a direct employee ranking can conceal the difference instead of explaining it.
8. Make report quality visible
Show refresh time, source coverage and known limitations beside the results. If one store has not supplied data, a sales decline must not silently appear as a business outcome. Mark incomplete periods or missing sources. Agree when figures are provisional, when they become final and how late corrections are handled. A historical number that changes should have an explainable reason rather than leaving users to guess whether the report is broken.
Test the report on a small reference set: a normal operation, cancellation, reopening, split shipment and delayed event. Calculate expected results manually and compare them with the dashboard. Preserve those checks for future changes. Permissions can also affect numbers: a user may see only part of the population. Explain whether a result is local to their access or company-wide. Do not solve completeness by exposing unnecessary personal information. Aggregation and permissions should fit the management decision. Confirm that filters, exports and drill-downs preserve the same interpretation rather than presenting subtly different populations under an identical metric name.
9. End the review with an action and an effect check
During a regular review, select an exception, inspect a few records and formulate one hypothesis. Perhaps the queue is growing because item confirmation is missing, not because picking is slow. Assign an owner to a small change, set a review date and identify the measure that will indicate progress. Avoid changing everything simultaneously if you want to understand the contribution of a particular intervention. Record what was actually changed, not only the intention.
Annotate process changes, promotions, outages and other events affecting comparisons. Improvement after an intervention does not prove causation: demand or case mix may also have changed. State the limitations and keep observing. A useful dashboard does not replace management; it sharpens questions and makes consequences easier to examine. If nobody uses a measure for months, remove it from the main view or reconsider its purpose. The goal is a reliable decision loop, not an ever-growing collection of charts.
- Every key metric has a decision and an owner.
- Formula, period and exclusions are documented.
- Rates are displayed with their underlying volumes.
- Speed is considered alongside quality and open work.
- Reference scenarios verify the report’s calculations.
Official sources
Primary sources for definitions and further reading. Routio’s practical examples are educational, not customer case studies. Verify system configuration against the documentation for your version.