A dashboard says completion is up. Someone asks what counts as complete. The meeting moves from the work to the meaning of the number.
Write that meaning down before the next meeting.
A business metric dictionary gives each number a definition someone else can reproduce. Start with one disputed metric. Name its business question, counting unit, eligible records, calculation, and date rules. Then identify who can change those rules.
The worksheet below is a proposed operating method. The service-request records are fictional. No software purchase or particular data platform is required.
Start With the Decision the Number Supports
Write a question with an operational consequence: “What share of requests due yesterday had status Closed at the day's end?”
That question can support a backlog review. It cannot tell you whether customers liked the work or whether the team has enough capacity. Keep those questions separate.
An operator dashboard addresses which indicators deserve attention. A dictionary defines the indicators selected for that dashboard. Copying a chart into another report should carry its definition with it.
Give the metric a stable identifier, such as requests_due_completion_rate. Keep the display label readable. Two charts can use different labels while referring to the same identifier and definition version.
Record the decision owner beside the question. A report builder can implement a calculation without having authority to decide which requests belong in it. Make that distinction explicit before work starts.
Ask the owner to explain what a changed number would prompt them to inspect. If the answer is unknown, leave the intended use unresolved. A precise calculation does not create a useful decision by itself.
Copy This Business Metric Dictionary Template
Use one definition sheet per metric. The entries below describe the proposed fields, not requirements imposed by a vendor or standard.
| Field | What to write |
|---|---|
| Identifier and version | Stable metric ID, definition version, and effective date |
| Business question | The question this calculation answers and the decision it informs |
| Unit | Requests, dollars, hours, or a ratio displayed as a percentage |
| Grain | What one input row represents and which key identifies it |
| Formula | Calculation, numerator, denominator, and rounding rule |
| Population | Exact conditions that make a record eligible |
| Exclusions | Test records, cancellations, or other agreed removals |
| Time basis | Date field, interval boundaries, time zone, and status cutoff |
| Source | System, dataset, fields, and reproducible query or export reference |
| Refresh time | Expected refresh, actual successful refresh, and source coverage watermark |
| Owner | Person approving meaning and person maintaining the implementation |
| Result state | Rules for zero, blank, incomplete periods, and unavailable sources |
| Comparison rules | Which versions, populations, and periods can be compared |
| Change record | Approved change, reason, affected history, and restatement decision |
Keep the source reference specific enough to find the input again. “Service system” is a starting point. “Daily request snapshot, request ID, due timestamp, status at cutoff, cancellation flag, test flag” gives the builder a field map.
Store the definition where the team can read it. A spreadsheet tab or a short document can hold the first version. Access should match the records it references. The dictionary does not need copies of customer data.
State What One Row Represents
Grain is the meaning of one row. A request table might contain one row per request. An activity table might contain several rows per request. Counting rows in those tables answers different questions.
Microsoft's Power BI star-schema guidance distinguishes fact tables containing observations from dimensions used for filtering and grouping. It recommends a consistent grain for fact tables. That guidance concerns Power BI models; the worksheet applies the same question to a small operational report.
For this example, write: “One row represents one request, identified by request ID, in the retained cutoff snapshot.” A reopened request keeps its identifier. The proposed metric counts eligible requests once. An eligible request reopened at cutoff stays in the denominator and contributes zero to the numerator.
If you need completion events, define another metric. One request can have more than one completion event after reopening. Do not change the counting unit because an event table is easier to query.
Also name the output grain. A daily team result is different from a daily result for each service type. Specify whether team membership follows the cutoff snapshot or the current assignment. A later reassignment should not change history without a documented rule.
Work Through a Fictional Completion Rate
The fictional team wants the share of eligible requests due on October 26 that had status Closed at the end of that day.
Use UTC for this example. The due interval starts at 2026.10.26 00:00 UTC and ends before 2026.10.27 00:00 UTC. Evaluate status immediately before the ending boundary. The retained extract was produced at 2026.10.27 08:00 UTC.
Exclude cancellations and test requests from both parts of the calculation. A request marked Resolved remains eligible, but it does not satisfy this definition's Closed condition.
All records in the table have due timestamps inside the interval. Every identifier and state is invented.
| Request | Status at cutoff | Other condition | Denominator contribution | Numerator contribution |
|---|---|---|---|---|
| R01 | Closed | None | 1 | 1 |
| R02 | Closed | None | 1 | 1 |
| R03 | Closed | None | 1 | 1 |
| R04 | Closed | None | 1 | 1 |
| R05 | Open | None | 1 | 0 |
| R06 | Resolved | None | 1 | 0 |
| R07 | Cancelled | Cancellation | 0 | 0 |
| R08 | Closed | Test request | 0 | 0 |
The denominator is six distinct eligible requests due in the interval. The numerator is four of those requests with status Closed at the cutoff. Calculate four divided by six, then multiply by one hundred. Display 66.7 percent after rounding to one decimal place.
Keep four and six available beside the percentage. A percentage alone hides the size of the population. Preserve the unrounded ratio for later calculations.
This definition measures completion of work due in the interval. It does not count every request closed during the interval. A request due earlier can close today without entering today's due population. Label those two measures separately.
Give Empty Results a Meaning
Write explicit display rules. Do not let a blank cell carry four possible explanations.
For the proposed completion rate, use these states:
| Observed condition | Display rule | Meaning |
|---|---|---|
| Six eligible requests, none Closed | 0.0 percent, with 0 of 6 | Known population with no qualifying completions |
| No eligible requests | Blank ratio with “No eligible requests,” with 0 of 0 | The denominator is zero; this ratio is undefined |
| Current day remains open | Result with “Incomplete period” and cutoff time | Observed work covers only part of the intended interval |
| Source extract failed | “Unavailable” with last successful refresh | The current result cannot be established |
A genuine zero requires evidence that the population was examined. A failed source cannot establish that no work occurred. If you retain the previous result during an outage, show its original period and refresh time.
An incomplete period can also have a zero numerator. Keep the value and state separate. “0 of 3, incomplete through 10:00 UTC” preserves information that an isolated zero cannot carry.
Do not infer completeness from the scheduled refresh time. Record the successful extraction time and the source coverage it confirmed. If coverage is unknown, say so.
Separate Event Time From Knowledge Time
The request's due time determines eligibility. Its status at cutoff determines numerator membership. The extract time states when the report obtained those records. These fields have different jobs.
Suppose an October 26 completion arrives in the reporting dataset on October 27 after the retained extract. Choose whether the historical result stays frozen or receives a labeled restatement. Record that choice in the definition.
The worksheet does not assume your source preserves historical status. If it holds only current status, a later export cannot establish yesterday's state by itself. Mark the missing evidence before presenting the metric as a historical measure.
Include the time zone in the definition and the report. For this example, the business selects UTC. Another business can choose a named local zone. It must then specify how its date boundaries become timestamp boundaries.
Use the dictionary when designing reporting automation. A scheduled refresh should carry the intended interval, actual source coverage, and result state into the report. A successful schedule run does not settle those business definitions.
Use dbt as an Example of Explicit Configuration
The dbt metrics documentation, in its section for dbt v1.12 and later, describes YAML metric definitions. It lists metric names, types, descriptions, and optional filters. Ratio metrics identify a numerator and denominator.
That is a useful example of making assumptions visible in configuration. It is not a requirement to install dbt. This article supplies no executable dbt configuration and makes no claim about an existing installation.
In a small worksheet, the equivalent discipline is concrete: name the numerator, name the denominator, and state the filter applying to each. Keep the readable definition aligned with the implemented calculation.
Ask a second person to reproduce the fictional four-of-six result from the sheet. If they count the test request or treat Resolved as Closed, revise the definition before debating software.
The calculation can be explicit while the business rule remains undecided. Mark an unapproved status mapping as proposed. Configuration records a choice; it does not supply the authority to make it.
Version a Change Before Comparing History
Suppose the owner decides that Resolved should count as complete. Under version two, R06 adds one to the numerator. The same six-request denominator now produces five of six, or 83.3 percent.
That increase comes from the definition change in this example. The underlying records did not change. Comparing version one's 66.7 percent with version two's 83.3 percent as a performance gain would misdescribe the evidence.
Record the old rule, new rule, approver, effective date, and affected reports. Identify which historical comparisons stop being valid. A new population exclusion, date field, or counting unit can also break comparability.
Choose a visible transition. You can retain the old series and begin a new versioned series. You can restate prior periods when the required evidence exists. You can display both versions during a defined review window.
If history cannot be recalculated, keep the break visible. Use a label such as “Definition changed: Resolved included from version two.” Do not draw an uninterrupted comparison that implies unchanged meaning.
The same discipline belongs in executive reporting. A concise chart still needs a traceable definition and a visible comparison boundary. Shortening a presentation should preserve those facts.
Assign Maintenance to a Person
Before accepting the sheet, have the owner check one ordinary record and one exclusion. Have the maintainer locate the source fields. Have a report reader explain the blank-result rule.
Record any disagreement as an open question with an owner. Keep an unapproved metric labeled proposed until that question is resolved. A polished card should not hide an unfinished definition.
Review the sheet when the source changes, the decision changes, or a reader finds a reproducibility problem. Keep previous versions accessible to the people interpreting old reports.
Every displayed number needs a meaning someone owns.
Sources
- dbt Developer Hub: Creating metrics. Metric configuration and ratio components; section marked dbt v1.12 and later. Checked 2026.10.28.
- Microsoft Learn: Understand star schema and the importance for Power BI. Fact-table grain and dimension roles in Power BI. Checked 2026.10.28.