An indicator is not useful because it has two decimal places or appears on a chart. It is useful when someone else can calculate it the same way, understand the decision it supports, and recognize what it might hide.

A coffee shop may see a hundred numbers in its sales system and need only a few to decide whether it is improving. A service provider might record dozens of tasks without knowing whether commitments are met. A digital product can collect endless analytics events without measuring when users obtain value. UK government guidance recommends defining a service's purpose, intended benefits, and hypotheses before choosing measures and data sources [R01]. This avoids starting with a generic list of “top KPIs.”

This article explains what to measure, exactly how to calculate it, and which decisions it informs. Cleaning and joining the original records belong to the business data integration guide. Charts, alerts, and regular review meetings will be addressed in the next Pillar 7 article.

1. Distinguish data, metrics, indicators, KPIs, objectives, targets, and thresholds

Term Practical meaning Hypothetical example
Data Elementary record of an event An order was delivered at 1:12 p.m.
Metric Measurement derived under a defined rule 396 on-time orders out of 440 delivered
Indicator Metric interpreted as evidence of performance or change On-time delivery rate of 90% over four weeks
KPI Key indicator because it directly supports an objective and decision Punctuality when reliable delivery is a central promise
Objective Desired result Fulfill delivery commitments reliably
Target Value to reach within a period At least 93% in the next comparable period
Action threshold Condition triggering investigation or action Investigate the same day if performance falls below 88%

The difference between an ordinary indicator and a KPI is contextual: a measure critical to an emergency service may not be central to a gift shop. The 93% target and 88% threshold are illustrative management choices, not universal industry benchmarks. Not everything measured should become a KPI.

UK service performance guidance distinguishes measurements from selected KPIs used to understand whether a service meets its purpose, and recommends combining service analytics with user research [R01][R02].

2. Build a small measurement tree from results back to evidence

Start with a question: “How do we know we are delivering reliable lunches without losing money?” One practical chain is:

  • Business outcome: a fulfilled service with quality and sustainable contribution.
  • Customer outcome: the complete order arrives by the time promised.
  • Controllable conditions: capacity checked before confirmation, appropriate preparation, packing, and dispatch verification.
  • Outcome indicators: punctuality, unmet requests, incomplete orders, and contribution per delivered order.
  • Leading signals: the share of bookings with capacity checked before acceptance; orders scheduled into an overloaded time window.
  • Guardrails: complaints, safety, rejected requests, and unusual costs.

Do not confuse an indicator with a cause. Two lines moving together may suggest a hypothesis, not prove that changing one caused the other. Causal investigation and experiments are covered separately in this pillar.

3. Write a complete technical definition for every KPI

A reproducible KPI needs a name, purpose, formula, numerator, denominator, population, segmentation, time period, source, owner, refresh frequency, unit, target, threshold, assumptions, exclusions, and definition-change record. If a denominator does not apply—for example, total sales—write not applicable, never zero.

Example: on-time delivery rate.

  • Question: What proportion of delivered orders with a promised time actually met that commitment?
  • Numerator: delivered orders arriving by their recorded promised delivery time, using a documented rule.
  • Denominator: all orders delivered in the period with a verifiable promised delivery time.
  • Formula: on-time rate (%) = 100 × on-time deliveries / delivered orders with a promised time.
  • Exclusions: orders canceled before delivery. Orders lacking promised times are categorized as incomplete data, not silently discarded. Business-initiated cancellations or rejected eligible requests remain visible as guardrails.
  • Segments: channel, day of week, service window, location; compare only groups with compatible definitions.
  • Source and owner: order history and dispatch log; operations lead.
  • Cadence: daily for alerts and four comparable weeks for trend review.
  • Interpretation: 90% means nine in ten orders in the defined denominator met their promise. It does not mean 90% of all incoming requests ended successfully.

The formula's precision is an editorial requirement. If two people disagree about “delivered,” “on time,” or “eligible order,” they have a definition problem, not a charting problem.

4. Combine leading signals, outcomes, and safeguards

Lagging indicators describe outcomes that have already occurred: contribution, on-time deliveries, or defects. Leading indicators describe conditions that may predict them: checking capacity before accepting an order or having enough ingredients for tomorrow. A leading signal is useful only when there is a plausible link to the objective and an available action, not merely because it is called leading.

Also distinguish controllable inputs from observed external conditions. A team can change the way it allocates delivery slots; it cannot control every storm or surprise in customer demand. Record contextual factors for interpretation, but do not treat them as blanket excuses for underperformance.

Volume metrics (440 deliveries) do not automatically replace rates (396/440 on time). A rate without its volume is misleading too: one complaint out of two deliveries is 50%; one out of five hundred is 0.2%. The two percentages carry different evidentiary weight and may require different responses.

5. Worked example: six KPIs for “Almuerzo Cerca”

Hypothetical data from the same four-week baseline used in the improvement-definition example: 500 meals prepared, 60 discarded, 440 delivered, and 20 requests unfulfilled because the product was unavailable. Also assume 396 on-time deliveries, 11 incomplete deliveries, 422 confirmed orders with capacity verified before acceptance, and 18 recorded complaints associated with delivered orders. Selling price S/30, variable cost per delivered lunch S/18, and lost ingredient/preparation cost for a discarded lunch S/10, without double counting. These numbers are educational assumptions, not observations of real customers.

KPI in this example Exact formula and population Result Decision supported
Production waste rate 60 discarded / 500 prepared × 100 12.00% Adjust batch size without driving stockouts
Unmet requests due to product shortage 20 unmet / (440 delivered + 20 unmet) × 100 4.35% Review availability and service windows
On-time delivery rate 396 on time / 440 delivered with a promise × 100 90.00% Investigate windows and late commitments
Incomplete order rate 11 incomplete / 440 delivered × 100 2.50% Review packing verification before dispatch
Confirmations with capacity verified 422 verified / 440 confirmed orders in the period × 100 95.91% Require capacity checks before acceptance
Contribution per delivered order after waste [440×(30−18)−60×10] / 440 S/10.64 Check that reducing waste does not destroy economics

Population clarification: for this teaching example, all 440 confirmed orders were delivered and had a promised time; the 20 unmet requests were never confirmed. A real operation may have cancellations, changed promises, refunds, or duplicate inquiries. These need distinct statuses before calculation. The 18 complaints would be an additional diagnostic metric, not necessarily a seventh KPI: 18/440 = 4.09%, under the definition “delivered orders with at least one complaint.” If one order generates three messages, count one affected order rather than three orders.

These six KPIs are selected for this case, not a requirement to track six in every company. With limited resources, an owner might start with waste, unmet requests, punctuality, and contribution, while retaining other numbers for diagnosis. Targets and action thresholds should be set using relevant local evidence, not inferred automatically from this hypothetical example.

6. Defend the denominator against perverse incentives

Suppose the business tries to raise its 90% on-time rate by rejecting 20 difficult orders it previously would have accepted. If it delivers 420 orders and the same 396 still arrive on time, the displayed result is 396/420 = 94.29%. The headline KPI improves, while more customer requests may go unmet. Detect this by also observing eligible requests accepted, business-initiated cancellations, defects, and customer experience. Do not allow late deliveries to be relabeled as cancellations without reason and an auditable history.

This is a general warning about measures that can encourage the appearance of progress. It is not evidence that targets always damage performance. If an indicator can improve merely by hiding difficult cases, reconsider its definition and guardrails before using it to reward a team.

7. Check sample size, segments, comparability, and version history

Before acting on a percentage, show n, time horizon, and context. Three late deliveries out of five mean 60% were late; that may signal a serious problem, but it is too small a sample to confidently predict performance across all locations. Aggregated averages may hide lunch-window delays while afternoon service runs smoothly. UK service guidance recommends comparing trends and segments rather than reading isolated values [R01].

If you change the meaning of “on time”—for example, from “by 1 p.m.” to “within fifteen minutes of the offered slot”—record the version, effective date, owner, and reason. A time series calculated under incompatible rules should not be compared without recalculation. When timestamps are missing, report data completeness; do not assume those orders were on time.

Data quality is part of KPI credibility: the UK framework treats accuracy, completeness, timeliness, uniqueness, consistency, and validity as distinct dimensions [R03]. A mathematically correct formula over duplicated records still yields a wrong answer.

8. Adapt KPIs to the business model; avoid universal lists

Business model Relevant outcome Possible indicator Caveat
Retail or coffee shop Complete deliveries and contribution Stockouts, waste, contribution per order A ticket and a customer are different units
Service provider Meet appointments and resolve issues on first visit On-time service, repeat visits Segment by job complexity
Production business Conforming units and safe operations Conforming yield, rework Do not trade safety for speed
Online store Completed, fulfilled purchases Net orders, returns, fulfillment A session is not a purchase [R04]
Subscription app Valuable use and retention Activation based on a defined event, cohorts Registration is not adoption; define the population

Acquisition indicators such as CAC, CPA, and ROAS are already covered in the marketing and sales performance guide. The focus here is choosing and documenting broader business indicators. The incremental economic impact of a specific initiative, including investment returns, is treated in a separate later Pillar 7 guide.

9. Reusable template: KPI dictionary

Someone who did not help define the indicator should be able to calculate it using this sheet:

Required field Question to complete
Name, ID, and purpose What result does it represent, and what decision changes?
Unit and formula Count, rate, currency per service? Exact calculation?
Numerator and denominator Which records count, which are excluded, and why?
Population and segments Which customers, orders, windows, or products count?
Source and quality Which systems or files supply the inputs? What is missing?
Period and refresh When did events happen, when were they exported, and when reviewed?
Owner Who resolves definition disputes and anomalies?
Baseline, target, and threshold What was observed, and what triggers action?
Guardrails What may not worsen even if the KPI improves?
Version and changes How does it remain comparable with its own history?

For an initial exercise, one sheet with these fields plus a second tab for period numerators and denominators is sufficient. You do not need a CRM, data warehouse, or BI subscription to define four indicators correctly. There is also no universal requirement to track exactly five or seven KPIs.

10. Turn a measure into a responsible decision

An indicator should lead to reviewing, investigating, retaining, or changing something specific. If punctuality falls below an agreed threshold, the next step may be to investigate orders by service window and separate kitchen, delivery, and overpromising problems. It does not automatically mean hiring more staff. If the rate rises, confirm that eligible requests, quality, and unit economics were protected.

What you should have by the end: a short set of KPIs that others can reproduce, each connected to a decision and supported by guardrails that prevent simplistic interpretations. The next step is visualization and a review routine, not drawing charts before the calculation rules are stable.

References