Data Quality Decisions That Matter before the First Build
Data quality is not a score a platform bestows on a warehouse. It is evidence that a named dataset is fit for a named decision at a known point in time. For a CTO, a useful starting case is the weekly revenue and retention pack: can the leadership team use it to set priorities, or must it be qualified because records arrived late, definitions changed, or ownership is unresolved? Write that decision down before selecting a rule. It exposes the consumer, reporting cadence, material fields, tolerated error, and cost of a misleading answer. Completeness may matter more than precision for a staffing forecast, while an automated entitlement decision may require stricter identity checks. The practical goal is a result that a user can trust, challenge, and correct, not an ever-growing catalogue of checks.
Take a customer record has two active account owners after a CRM merge. In data quality, that is not a minor edge case; it is the point at which assumptions about identity, timing, and meaning become visible. The team should decide in advance whether the record is rejected, quarantined, corrected, or reported with a qualification. A failed quality rule should block or clearly qualify a decision-ready output rather than quietly disappear. Making the boundary explicit prevents the common pattern in which people discover an ambiguity only after an executive meeting, customer interaction, or operational escalation.
Start with the decision boundary for data quality
A decision statement gives data quality a testable purpose. Name the decision, the accountable actor, the cadence, and the cost of being wrong or late. Then capture the minimum evidence that must accompany the result: source system, business definition, acceptable threshold, and accountable steward. This is more precise than collecting a broad list of desirable fields. It tells delivery teams which conditions are material and gives business owners a way to review trade-offs. A metric may be accurate enough for weekly planning and unsuitable for customer-facing automation; the boundary should say so.
| Question before build | Practical choice | Evidence to retain |
|---|---|---|
| Who takes action? | Name the owner who decides whether the weekly revenue and retention pack is fit for the leadership meeting. | Decision log and operating cadence. |
| What can change the answer? | List the material inputs and exclusions. | Definition, schema, and sample cases. |
| How current must it be? | Set a freshness or event-time expectation. | Last successful run and delayed-data policy. |
| What happens when it fails? | Choose block, qualify, or route for repair. | Alert owner, incident note, and correction record. |
Architecture and controls for data quality
The architecture should separate evidence capture, controlled calculation, publication, and observation. In practice, profile critical fields first, then write tests around the failure modes that would change the decision. Keep raw or source-shaped evidence accessible to authorized investigators; make the published layer small enough that a user can understand its grain, timing, and exclusions; and record the version of the logic that produced a consequential result. This division makes correction possible without pretending that every anomaly can be resolved automatically.
Ownership matters as much as the data path. The business owner approves meaning and prioritizes remediation; the technical owner operates collection, transformation, access, and recovery; consumers report confusing or surprising results through a visible route. For data quality, a review should use recent exceptions rather than slideware: inspect a failed rule, an unexpected trend, a delayed input, and one corrected record. That routine exposes whether the stated control actually works in daily use.
| Layer | Responsibility in this design | Failure signal |
|---|---|---|
| Evidence | Capture the identifiers, time, and source context needed to verify a case. | Missing key, late input, or unexpected volume. |
| Controlled logic | Apply approved rules and preserve calculation version. | Test failure, reconciliation gap, or schema change. |
| Published result | Show the answer, freshness, scope, and exception state. | Stale output, unexplained shift, or blocked access. |
| Operations | Route alerts, repair data, and communicate material changes. | Unowned incident or repeated manual workaround. |
A phased rollout for data quality
Begin with one decision, one data product, and a supervised reporting cycle. Use historical examples plus a small live sample, including incomplete, late, and corrected cases. Compare the new result with the current method and investigate differences before declaring one system authoritative. A good pilot produces a named baseline, acceptance criteria, support contact, and recovery exercise. It also produces a decision: extend the scope, revise the definition, or stop. That is a much stronger outcome than a technically successful demonstration with no evidence that the workflow can be operated.
- Write a one-sentence decision statement for data quality and have the action owner approve it.
- Select the smallest source-to-decision path and document the material fields, definitions, and exclusions.
- Create checks for the failure modes that would change whether the weekly revenue and retention pack is fit for the leadership meeting, including the case where a customer record has two active account owners after a CRM merge.
- Make freshness, scope, and exceptions visible to users rather than keeping them in an engineering runbook.
- Run the pilot alongside the existing process and retain explanations for material differences.
- Expand only after the owner can explain detection, communication, correction, and recovery.
Measures that show whether data quality is working
Measure behavior and reliability together. For data quality, track completeness, uniqueness, freshness, rule failures, and time to explain a material discrepancy. Pair these operational signals with a direct question for users: which decision changed because this evidence was available, and could they explain why they trusted it? Raw usage, query volume, or job-success counts are useful context, but none demonstrates that the result improved work. A temporary increase in questions can be healthy when it reveals definitions that were assumed instead of agreed.
Sources used for this data quality guide
The W3C Data Quality Vocabulary is useful here because it treats quality as fitness for a stated use, expressed through measurements, policies, and feedback rather than a universal score. W3C PROV-DM informs the recommendation to retain the activity, inputs, and agent behind a consequential result. dbt data tests documentation supports turning material assumptions into executable checks, while OpenTelemetry semantic conventions is a practical reference for consistent operational telemetry. These sources describe patterns, not a single threshold: the owner must set the tolerance appropriate to the decision.
Review data quality before wider release
Before a wider release, review one changed input, one failed or delayed run, and one user decision that depended on the result. Ask whether source system, business definition, acceptable threshold, and accountable steward still describe the real workflow and whether a person outside the delivery team can trace the answer without informal help. For data quality, the release record should identify the logic version, effective date, owner, and any known limitations. This review is deliberately modest. Its purpose is to catch a change that would alter whether the weekly revenue and retention pack is fit for the leadership meeting before it becomes embedded in a recurring meeting, automation, or customer process. Include a steward sign-off on threshold changes and retain the before-and-after exception count, so the quality rule remains connected to the reporting decision it protects rather than becoming an unexamined technical check. Capture the approver, rationale, and effective date whenever a material quality threshold changes. Review it at each release.
Use exception samples, not only aggregate success rates, to judge readiness. Reconstruct the treatment of the case where a customer record has two active account owners after a CRM merge; then verify that the published result, alert, or report would make the uncertainty visible to the intended user. Compare that exercise with completeness, uniqueness, freshness, rule failures, and time to explain a material discrepancy. If the team cannot explain a discrepancy, pause expansion and fix the definition, source contract, or recovery route. A narrow, explainable capability earns more trust than a broad data quality implementation whose assumptions are available only to its builders.
Key takeaways
- Data quality should begin with a consequential decision and named action owner.
- Treat definition, timing, provenance, and correction as visible parts of the product.
- Use a narrow pilot with real exceptions to test the operating model, not just the data path.
- Scale only when users can investigate a surprising answer and the team can recover a failed interval.
Frequently asked questions about data quality
What is the first useful milestone for data quality?
The first milestone is a supervised decision path, not a broad platform rollout. A named user should be able to obtain the result, see whether it is current and in scope, follow an exception to a responsible owner, and compare the answer with enough evidence to explain it. For data quality, keep this first path deliberately small. It should include the uncomfortable cases, because those reveal the controls and definitions that ordinary happy-path examples hide.
Do we need a new tool before implementing data quality?
Usually, no. First establish whether the existing stack can capture the necessary evidence, apply the agreed rules, restrict access where needed, expose timing and exceptions, and retain a correction path. A new tool is justified when it removes a demonstrated reliability, scale, security, or maintainability limit. Tool selection should follow the decision boundary for data quality; it cannot substitute for ownership, definitions, or a release and recovery practice.
Conclusion
The durable version of data quality is not a collection of reports, events, or jobs. It is an operating capability that helps finance and commercial leaders decide whether the weekly revenue and retention pack is fit for the leadership meeting with appropriate confidence. Start with the decision, state the evidence boundary, design for exceptions, and prove the workflow in a supervised pilot. That sequence keeps the build honest: it makes value visible early while preserving the controls needed to explain, correct, and improve the result over time.