Skip to content

Data Quality Standard

Ready

Word Markdown

Purpose. Defines what "good enough" data means in measurable terms — the dimensions, how they are measured, the thresholds that apply, and who acts when a threshold is breached.

When to use it. Alongside the data governance policy, and specifically before any dataset is used to train or evaluate a model. Review thresholds annually.

How to use it. Do not attempt all six dimensions across the whole estate at once. Pick your two or three highest-value domains, set thresholds, and measure those until the process is real — then extend. A standard measured nowhere is worse than no standard, because it implies assurance you do not have.

Closes assessment gaps

This template addresses Q12, Q28 in the readiness assessment.

The template

1. Quality dimensions

Dimension Question it answers Example measure
Completeness Is the expected data present? % of mandatory fields populated
Accuracy Does it reflect the real world? % matching a verified source
Timeliness Is it current enough to use? Age at point of use vs. requirement
Consistency Does it agree across systems? % of records matching across systems
Validity Does it conform to its rules? % passing format / range rules
Uniqueness Are there duplicates? % duplicate records

2. Thresholds

Thresholds are set per domain by the Data Owner, based on use. Data feeding a regulatory return or a high-risk AI system needs tighter thresholds than data feeding an internal dashboard.

Domain Dimension Threshold Measure Frequency Owner
Customer Completeness ≥ 98% [rule] Daily [role]
Customer Uniqueness ≥ 99% [rule] Weekly [role]
[domain] [dimension]

3. Quality for AI training data

Additional requirements where data trains, validates, or tests a model — this is the Art. 10 obligation in practice:

Requirement What it means Evidence
Relevance The data suits the stated purpose. Documented rationale in the Model Card / Model Risk Documentation
Representativeness It reflects the population the system will act on. Distribution analysis vs. target population
Bias examination Known skews are identified and addressed. Bias assessment; mitigation record
Provenance Source and licence are known and lawful. Entry in Data Asset Register
Split integrity Train/validation/test do not leak into each other. Split methodology documented

Representativeness is not a percentage

A dataset can be 99% complete and still unrepresentative of the people the system will affect. Completeness measures the data you collected; representativeness asks who is missing from it entirely.

4. Measurement & reporting

Scores are produced monthly (weekly for data feeding a high-risk system), published to [location], and summarised in KPI / KRI Dashboard. Trend matters more than absolute level.

5. Breach handling

Severity Definition Response Timeframe
Critical Regulatory or high-risk AI impact Escalate to committee; consider suspending use 24 hours
Major Material business impact Owner-led remediation plan 5 working days
Minor Limited impact Backlog for next cycle 30 days

Breaches are recorded in the Issue & Incident Log.

6. Review

Version Date Owner Approved by
0.1 [date] [role] [committee]

Adaptation notes

  • Small organisations: Measure completeness and validity only, on your two most important domains. That covers most real-world failures at a fraction of the cost.
  • Regulated financial services: Align with BCBS 239 and make lineage evidence explicit — supervisors ask to trace a reported figure back to source.
  • ML-heavy organisations: Section 3 is the section that matters. Automate representativeness and drift checks into the training pipeline so the evidence is a by-product rather than a project.

Not legal advice

These templates are a head start, not a substitute for professional judgement. Adapt them to your jurisdiction, sector, and risk appetite, and have qualified counsel review anything material before you rely on it.