Metric dictionary
An AI metric dictionary is a controlled record of the definitions used to measure AI investment, adoption, workflow performance, cost, quality, risk, and business outcomes. It tells every team exactly what a metric means, how it is calculated, where its data comes from, and how it should be interpreted.
The dictionary makes results reproducible. It prevents two dashboards from using the same metric name for different populations, formulas, or completion rules.
What should an AI metric dictionary contain?
Each entry should include the metric name, business question, formula, unit, eligible population, numerator, denominator, exclusions, source systems, refresh cadence, owner, and effective date.
Add the workflow and unit of work when they affect interpretation. A metric called completion rate is ambiguous unless the entry identifies what starts the workflow, what counts as completed, and which rejected or reopened cases remain in scope.
Include the AI measurement window and baseline period so reviewers know which observations the metric covers. Record known limitations and allowed breakdowns. These may include team, use case, work type, complexity, model, or region. State whether the measure is an input, activity, output, outcome, financial result, or guardrail.
NIST's AI Metrology Center asks metric contributors for a precise mathematical or textual definition and the resources needed to implement the measure (NIST metric submission guidance). An enterprise dictionary should apply the same discipline to internal metrics.
How is a metric dictionary different from a data dictionary?
A data dictionary describes fields, tables, types, relationships, and technical constraints in a dataset. A metric dictionary describes a business measure built from those fields.
The two records should connect. The metric entry might define accepted cases per labor hour, a measure covered under AI productivity measurement, while the data dictionary identifies the case-status field, employee-time field, join key, and source tables used in the calculation.
Data readiness for AI measurement depends on both. A well-written formula cannot compensate for missing source records, and technically complete data cannot compensate for an undefined business measure.
How should outcome and attribution fields be documented?
For every AI outcome metric, identify the decision it supports and the direction that represents improvement. Add quality or risk guardrails so a faster result cannot hide worse work.
For metrics used in impact attribution, document the intervention, comparison method, baseline period, analysis population, confounders, and uncertainty treatment. Keep observed change separate from attributed impact.
The US Government Accountability Office describes reliable data as accurate, complete, and applicable for the intended purpose (GAO data reliability guide). The metric dictionary should state that intended purpose so reviewers can judge whether the underlying data is suitable.
Who should own the dictionary?
A central measurement or data owner can maintain the format and approval process. Individual metric owners should remain accountable for business meaning, calculation, and source quality.
Changes should involve the workflow owner and the teams that use the result. Finance should review financial measures. Risk or security owners should review relevant guardrails. Data owners should confirm lineage and implementation.
Common metric-dictionary mistakes
Common failures include recording only a name and formula, omitting populations and exclusions, mixing raw fields with business metrics, allowing dashboards to redefine measures locally, and deleting previous definitions.
Another mistake is documenting every available metric without connecting it to a decision. Keep the dictionary broad enough for governance but prioritize the measures used to fund, expand, renew, improve, or stop AI initiatives. AI pilot success criteria show where those measures become decision thresholds. The AI initiative inventory records the baseline and target metrics for each initiative.