Data Readiness Before AI: A Practical Decision Table
A practical test for whether data is fit for a specific AI use case across purpose, ownership, quality, access, and evaluation—and what to fix before a pilot.
Updated 29 August 2026. Data is never “AI-ready” in the abstract. Readiness belongs to a specific use case. A knowledge base may be suitable for internal search but unsuitable for an automated customer decision. A small sample may be enough to test retrieval but nowhere near enough to train a model.
The useful question is not “Do we have a lot of data?” It is “Do we have the data required for this purpose, the right to use it, an understanding of its quality and limits, and a way to evaluate the system's result?”
Define the use case's data contract
Before selecting technology, write a short contract: purpose, source, owner, fields used, access, retention, expected output, and verification method. This exposes whether the real issue is missing data, conflicting definitions, weak ownership, or an unstable workflow.
Where personal data is involved, consult the SDAIA Personal Data Protection knowledge center, including purpose and data-minimization considerations. This article is an operating tool, not legal advice; the organization and its qualified advisers remain responsible for regulatory assessment.
The readiness decision table
| Dimension | Ready to pilot | Preparation needed | Temporary stop |
|---|---|---|---|
| Purpose | Named decision, user, and output | Broad goal without a baseline | No clear legitimate use |
| Ownership | Approved owner and known source | Distributed ownership without resolution | Unknown or unauthorized source |
| Quality | Representative sample and known error rate | Measurable gaps | Conflicting definitions prevent interpretation |
| Access | Least privilege and usage logging | Permissions wider than needed | Uncontrolled export of sensitive data |
| Coverage | Normal cases and exceptions represented | Time or segment bias | Not representative of operations |
| Evaluation | Reference outcomes and acceptance threshold | Human review without a standard | No way to recognize an error |
If a temporary stop appears, do not compensate with a larger model. Fix the right to use, the source, the definition, or the measurement method. A “preparation needed” result can become a bounded task inside a short pilot when the consequence of error remains limited and reversible.
Structured data and knowledge systems need different tests
For analysis or prediction, you need consistent definitions, comparable history, and an explanation of missing values and bias. For knowledge systems and retrieval-augmented generation, the priority shifts to source authority, chunking, freshness, permissions, and the ability to show the correct citation.
Finding a similar paragraph is not enough. Test whether the system retrieved the authoritative source, preserved context, refused when evidence was absent, and prevented a user from retrieving a document they could not otherwise access. These are product and governance decisions, not merely search settings.
Build a small evaluation set before integration
- Collect 30 to 50 questions or cases from real work, including ambiguous and rare examples.
- Assign an approved reference answer or outcome and source to each case.
- Define critical failures that cause rejection regardless of the average score.
- Test retrieval, output, permissions, cost, and latency separately.
- Store results with the version so regression becomes visible after a change.
If the team cannot build this set, it probably does not understand the decision well enough to operate it. Start with the data governance service to resolve ownership and definitions, then revisit feasibility with the use-case scorecard.
The final decision: pilot, prepare, or stop
- Pilot: purpose, owner, data, and evaluation are clear, and errors have limited consequence.
- Prepare first: value is plausible but quality, permissions, or definitions need a named intervention.
- Stop: there is no clear right to use the data, errors cannot be measured, or harm is irreversible.
After a pilot decision, connect the data contract to the risk register in the AI governance operating model. Rafid 360 in selected work provides public context for multi-organization collaboration and data without attributing unpublished outcomes. Readiness is not an endless cleaning program; it is a documented sufficiency threshold for a known use case.